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_try(hammer2_mtx_t *mtx)
 151 {
 152         return mtx_upgrade_try(mtx);
 153 }
 154
 155 /*
 156  * The xid tracks internal transactional updates.
 157  *
 158  * XXX fix-me, really needs to be 64-bits
 159  */
 160 typedef uint32_t hammer2_xid_t;
 161
 162 #define HAMMER2_XID_MIN 0x00000000U
 163 #define HAMMER2_XID_MAX 0x7FFFFFFFU
 164
 165 /*
 166  * The chain structure tracks a portion of the media topology from the
 167  * root (volume) down.  Chains represent volumes, inodes, indirect blocks,
 168  * data blocks, and freemap nodes and leafs.
 169  *
 170  * The chain structure utilizes a simple singly-homed topology and the
 171  * chain's in-memory topology will move around as the chains do, due mainly
 172  * to renames and indirect block creation.
 173  *
 174  * Block Table Updates
 175  *
 176  *      Block table updates for insertions and updates are delayed until the
 177  *      flush.  This allows us to avoid having to modify the parent chain
 178  *      all the way to the root.
 179  *
 180  *      Block table deletions are performed immediately (modifying the parent
 181  *      in the process) because the flush code uses the chain structure to
 182  *      track delayed updates and the chain will be (likely) gone or moved to
 183  *      another location in the topology after a deletion.
 184  *
 185  *      A prior iteration of the code tried to keep the relationship intact
 186  *      on deletes by doing a delete-duplicate operation on the chain, but
 187  *      it added way too much complexity to the codebase.
 188  *
 189  * Flush Synchronization
 190  *
 191  *      The flush code must flush modified chains bottom-up.  Because chain
 192  *      structures can shift around and are NOT topologically stable,
 193  *      modified chains are independently indexed for the flush.  As the flush
 194  *      runs it modifies (or further modifies) and updates the parents,
 195  *      propagating the flush all the way to the volume root.
 196  *
 197  *      Modifying front-end operations can occur during a flush but will block
 198  *      in two cases: (1) when the front-end tries to operate on the inode
 199  *      currently in the midst of being flushed and (2) if the front-end
 200  *      crosses an inode currently being flushed (such as during a rename).
 201  *      So, for example, if you rename directory "x" to "a/b/c/d/e/f/g/x" and
 202  *      the flusher is currently working on "a/b/c", the rename will block
 203  *      temporarily in order to ensure that "x" exists in one place or the
 204  *      other.
 205  *
 206  *      Meta-data statistics are updated by the flusher.  The front-end will
 207  *      make estimates but meta-data must be fully synchronized only during a
 208  *      flush in order to ensure that it remains correct across a crash.
 209  *
 210  *      Multiple flush synchronizations can theoretically be in-flight at the
 211  *      same time but the implementation is not coded to handle the case and
 212  *      currently serializes them.
 213  *
 214  * Snapshots:
 215  *
 216  *      Snapshots currently require the subdirectory tree being snapshotted
 217  *      to be flushed.  The snapshot then creates a new super-root inode which
 218  *      copies the flushed blockdata of the directory or file that was
 219  *      snapshotted.
 220  *
 221  * RBTREE NOTES:
 222  *
 223  *      - Note that the radix tree runs in powers of 2 only so sub-trees
 224  *        cannot straddle edges.
 225  */
 226 RB_HEAD(hammer2_chain_tree, hammer2_chain);
 227 TAILQ_HEAD(h2_flush_list, hammer2_chain);
 228 TAILQ_HEAD(h2_core_list, hammer2_chain);
 229 TAILQ_HEAD(h2_iocb_list, hammer2_iocb);
 230
 231 #define CHAIN_CORE_DELETE_BMAP_ENTRIES  \
 232         (HAMMER2_PBUFSIZE / sizeof(hammer2_blockref_t) / sizeof(uint32_t))
 233
 234 /*
 235  * Core topology for chain (embedded in chain).  Protected by a spinlock.
 236  */
 237 struct hammer2_chain_core {
 238         hammer2_spin_t  spin;
 239         struct hammer2_chain_tree rbtree; /* sub-chains */
 240         int             live_zero;      /* blockref array opt */
 241         u_int           live_count;     /* live (not deleted) chains in tree */
 242         u_int           chain_count;    /* live + deleted chains under core */
 243         int             generation;     /* generation number (inserts only) */
 244 };
 245
 246 typedef struct hammer2_chain_core hammer2_chain_core_t;
 247
 248 RB_HEAD(hammer2_io_tree, hammer2_io);
 249
 250 /*
 251  * IOCB - IO callback (into chain, cluster, or manual request)
 252  */
 253 struct hammer2_iocb {
 254         TAILQ_ENTRY(hammer2_iocb) entry;
 255         void (*callback)(struct hammer2_iocb *iocb);
 256         struct hammer2_io       *dio;
 257         struct hammer2_cluster  *cluster;
 258         struct hammer2_chain    *chain;
 259         void                    *ptr;
 260         off_t                   lbase;
 261         int                     lsize;
 262         uint32_t                flags;
 263         int                     error;
 264         int                     btype;
 265 };
 266
 267 typedef struct hammer2_iocb hammer2_iocb_t;
 268
 269 #define HAMMER2_IOCB_INTERLOCK  0x00000001
 270 #define HAMMER2_IOCB_ONQ        0x00000002
 271 #define HAMMER2_IOCB_DONE       0x00000004
 272 #define HAMMER2_IOCB_INPROG     0x00000008
 273 #define HAMMER2_IOCB_UNUSED10   0x00000010
 274 #define HAMMER2_IOCB_QUICK      0x00010000
 275 #define HAMMER2_IOCB_ZERO       0x00020000
 276 #define HAMMER2_IOCB_READ       0x00040000
 277 #define HAMMER2_IOCB_WAKEUP     0x00080000
 278
 279 /*
 280  * DIO - Management structure wrapping system buffer cache.
 281  *
 282  *       Used for multiple purposes including concurrent management
 283  *       if small requests by chains into larger DIOs.
 284  */
 285 struct hammer2_io {
 286         RB_ENTRY(hammer2_io) rbnode;    /* indexed by device offset */
 287         struct h2_iocb_list iocbq;
 288         struct spinlock spin;
 289         struct hammer2_dev *hmp;
 290         struct buf      *bp;
 291         uint64_t        crc_good_mask;
 292         off_t           pbase;
 293         int             psize;
 294         int             refs;
 295         int             act;            /* activity */
 296         int             btype;          /* approximate BREF_TYPE_* */
 297 };
 298
 299 typedef struct hammer2_io hammer2_io_t;
 300
 301 #define HAMMER2_DIO_INPROG      0x80000000      /* bio in progress */
 302 #define HAMMER2_DIO_GOOD        0x40000000      /* dio->bp is stable */
 303 #define HAMMER2_DIO_WAITING     0x20000000      /* wait for inprog clr */
 304 #define HAMMER2_DIO_DIRTY       0x10000000      /* flush on last drop */
 305
 306 #define HAMMER2_DIO_MASK        0x0FFFFFFF
 307
 308 /*
 309  * Primary chain structure keeps track of the topology in-memory.
 310  */
 311 struct hammer2_chain {
 312         hammer2_mtx_t           lock;
 313         hammer2_chain_core_t    core;
 314         RB_ENTRY(hammer2_chain) rbnode;         /* live chain(s) */
 315         hammer2_blockref_t      bref;
 316         struct hammer2_chain    *parent;
 317         struct hammer2_state    *state;         /* if active cache msg */
 318         struct hammer2_dev      *hmp;
 319         struct hammer2_pfs      *pmp;           /* A PFS or super-root (spmp) */
 320
 321         hammer2_io_t    *dio;                   /* physical data buffer */
 322         u_int           bytes;                  /* physical data size */
 323         u_int           flags;
 324         u_int           refs;
 325         u_int           lockcnt;
 326         int             error;                  /* on-lock data error state */
 327         int             persist_refs;           /* (aka ip->cluster) */
 328
 329         hammer2_media_data_t *data;             /* data pointer shortcut */
 330         TAILQ_ENTRY(hammer2_chain) flush_node;  /* flush list */
 331         TAILQ_ENTRY(hammer2_chain) lru_node;    /* 0-refs LRU */
 332 };
 333
 334 typedef struct hammer2_chain hammer2_chain_t;
 335
 336 int hammer2_chain_cmp(hammer2_chain_t *chain1, hammer2_chain_t *chain2);
 337 RB_PROTOTYPE(hammer2_chain_tree, hammer2_chain, rbnode, hammer2_chain_cmp);
 338
 339 /*
 340  * Special notes on flags:
 341  *
 342  * INITIAL      - This flag allows a chain to be created and for storage to
 343  *                be allocated without having to immediately instantiate the
 344  *                related buffer.  The data is assumed to be all-zeros.  It
 345  *                is primarily used for indirect blocks.
 346  *
 347  * MODIFIED     - The chain's media data has been modified.  Prevents chain
 348  *                free on lastdrop if still in the topology.
 349  *
 350  * UPDATE       - Chain might not be modified but parent blocktable needs
 351  *                an update.  Prevents chain free on lastdrop if still in
 352  *                the topology.
 353  *
 354  * FICTITIOUS   - Faked chain as a placeholder for an error condition.  This
 355  *                chain is unsuitable for I/O.
 356  *
 357  * BMAPPED      - Indicates that the chain is present in the parent blockmap.
 358  *
 359  * BMAPUPD      - Indicates that the chain is present but needs to be updated
 360  *                in the parent blockmap.
 361  */
 362 #define HAMMER2_CHAIN_MODIFIED          0x00000001      /* dirty chain data */
 363 #define HAMMER2_CHAIN_ALLOCATED         0x00000002      /* kmalloc'd chain */
 364 #define HAMMER2_CHAIN_DESTROY           0x00000004
 365 #define HAMMER2_CHAIN_DEDUP             0x00000008      /* used as dedup src */
 366 #define HAMMER2_CHAIN_DELETED           0x00000010      /* deleted chain */
 367 #define HAMMER2_CHAIN_INITIAL           0x00000020      /* initial create */
 368 #define HAMMER2_CHAIN_UPDATE            0x00000040      /* need parent update */
 369 #define HAMMER2_CHAIN_DEFERRED          0x00000080      /* flush depth defer */
 370 #define HAMMER2_CHAIN_TESTEDGOOD        0x00000100      /* crc tested good */
 371 #define HAMMER2_CHAIN_ONFLUSH           0x00000200      /* on a flush list */
 372 #define HAMMER2_CHAIN_FICTITIOUS        0x00000400      /* unsuitable for I/O */
 373 #define HAMMER2_CHAIN_VOLUMESYNC        0x00000800      /* needs volume sync */
 374 #define HAMMER2_CHAIN_DELAYED           0x00001000      /* delayed flush */
 375 #define HAMMER2_CHAIN_COUNTEDBREFS      0x00002000      /* block table stats */
 376 #define HAMMER2_CHAIN_ONRBTREE          0x00004000      /* on parent RB tree */
 377 #define HAMMER2_CHAIN_ONLRU             0x00008000      /* on LRU list */
 378 #define HAMMER2_CHAIN_EMBEDDED          0x00010000      /* embedded data */
 379 #define HAMMER2_CHAIN_RELEASE           0x00020000      /* don't keep around */
 380 #define HAMMER2_CHAIN_BMAPPED           0x00040000      /* present in blkmap */
 381 #define HAMMER2_CHAIN_BMAPUPD           0x00080000      /* +needs updating */
 382 #define HAMMER2_CHAIN_IOINPROG          0x00100000      /* I/O interlock */
 383 #define HAMMER2_CHAIN_IOSIGNAL          0x00200000      /* I/O interlock */
 384 #define HAMMER2_CHAIN_PFSBOUNDARY       0x00400000      /* super->pfs inode */
 385
 386 #define HAMMER2_CHAIN_FLUSH_MASK        (HAMMER2_CHAIN_MODIFIED |       \
 387                                          HAMMER2_CHAIN_UPDATE |         \
 388                                          HAMMER2_CHAIN_ONFLUSH |        \
 389                                          HAMMER2_CHAIN_DESTROY)
 390
 391 /*
 392  * Hammer2 error codes, used by chain->error and cluster->error.  The error
 393  * code is typically set on-lock unless no I/O was requested, and set on
 394  * I/O otherwise.  If set for a cluster it generally means that the cluster
 395  * code could not find a valid copy to present.
 396  *
 397  * IO           - An I/O error occurred
 398  * CHECK        - I/O succeeded but did not match the check code
 399  * INCOMPLETE   - A cluster is not complete enough to use, or
 400  *                a chain cannot be loaded because its parent has an error.
 401  *
 402  * NOTE: API allows callers to check zero/non-zero to determine if an error
 403  *       condition exists.
 404  *
 405  * NOTE: Chain's data field is usually NULL on an IO error but not necessarily
 406  *       NULL on other errors.  Check chain->error, not chain->data.
 407  */
 408 #define HAMMER2_ERROR_NONE              0
 409 #define HAMMER2_ERROR_IO                1       /* device I/O error */
 410 #define HAMMER2_ERROR_CHECK             2       /* check code mismatch */
 411 #define HAMMER2_ERROR_INCOMPLETE        3       /* incomplete cluster */
 412 #define HAMMER2_ERROR_DEPTH             4       /* temporary depth limit */
 413
 414 /*
 415  * Flags passed to hammer2_chain_lookup() and hammer2_chain_next()
 416  *
 417  * NOTES:
 418  *      NOLOCK      - Input and output chains are referenced only and not
 419  *                    locked.  Output chain might be temporarily locked
 420  *                    internally.
 421  *
 422  *      NODATA      - Asks that the chain->data not be resolved in order
 423  *                    to avoid I/O.
 424  *
 425  *      NODIRECT    - Prevents a lookup of offset 0 in an inode from returning
 426  *                    the inode itself if the inode is in DIRECTDATA mode
 427  *                    (i.e. file is <= 512 bytes).  Used by the synchronization
 428  *                    code to prevent confusion.
 429  *
 430  *      SHARED      - The input chain is expected to be locked shared,
 431  *                    and the output chain is locked shared.
 432  *
 433  *      MATCHIND    - Allows an indirect block / freemap node to be returned
 434  *                    when the passed key range matches the radix.  Remember
 435  *                    that key_end is inclusive (e.g. {0x000,0xFFF},
 436  *                    not {0x000,0x1000}).
 437  *
 438  *                    (Cannot be used for remote or cluster ops).
 439  *
 440  *      ALLNODES    - Allows NULL focus.
 441  *
 442  *      ALWAYS      - Always resolve the data.  If ALWAYS and NODATA are both
 443  *                    missing, bulk file data is not resolved but inodes and
 444  *                    other meta-data will.
 445  *
 446  *      NOUNLOCK    - Used by hammer2_chain_next() to leave the lock on
 447  *                    the input chain intact.  The chain is still dropped.
 448  *                    This allows the caller to add a reference to the chain
 449  *                    and retain it in a locked state (used by the
 450  *                    XOP/feed/collect code).
 451  */
 452 #define HAMMER2_LOOKUP_NOLOCK           0x00000001      /* ref only */
 453 #define HAMMER2_LOOKUP_NODATA           0x00000002      /* data left NULL */
 454 #define HAMMER2_LOOKUP_NODIRECT         0x00000004      /* no offset=0 DD */
 455 #define HAMMER2_LOOKUP_SHARED           0x00000100
 456 #define HAMMER2_LOOKUP_MATCHIND         0x00000200      /* return all chains */
 457 #define HAMMER2_LOOKUP_ALLNODES         0x00000400      /* allow NULL focus */
 458 #define HAMMER2_LOOKUP_ALWAYS           0x00000800      /* resolve data */
 459 #define HAMMER2_LOOKUP_NOUNLOCK         0x00001000      /* leave lock intact */
 460
 461 /*
 462  * Flags passed to hammer2_chain_modify() and hammer2_chain_resize()
 463  *
 464  * NOTE: OPTDATA allows us to avoid instantiating buffers for INDIRECT
 465  *       blocks in the INITIAL-create state.
 466  */
 467 #define HAMMER2_MODIFY_OPTDATA          0x00000002      /* data can be NULL */
 468 #define HAMMER2_MODIFY_NO_MODIFY_TID    0x00000004
 469 #define HAMMER2_MODIFY_UNUSED0008       0x00000008
 470
 471 /*
 472  * Flags passed to hammer2_chain_lock()
 473  *
 474  * NOTE: RDONLY is set to optimize cluster operations when *no* modifications
 475  *       will be made to either the cluster being locked or any underlying
 476  *       cluster.  It allows the cluster to lock and access data for a subset
 477  *       of available nodes instead of all available nodes.
 478  */
 479 #define HAMMER2_RESOLVE_NEVER           1
 480 #define HAMMER2_RESOLVE_MAYBE           2
 481 #define HAMMER2_RESOLVE_ALWAYS          3
 482 #define HAMMER2_RESOLVE_MASK            0x0F
 483
 484 #define HAMMER2_RESOLVE_SHARED          0x10    /* request shared lock */
 485 #define HAMMER2_RESOLVE_UNUSED20        0x20
 486 #define HAMMER2_RESOLVE_RDONLY          0x40    /* higher level op flag */
 487
 488 /*
 489  * Flags passed to hammer2_chain_delete()
 490  */
 491 #define HAMMER2_DELETE_PERMANENT        0x0001
 492
 493 /*
 494  * Flags passed to hammer2_chain_insert() or hammer2_chain_rename()
 495  */
 496 #define HAMMER2_INSERT_PFSROOT          0x0004
 497
 498 /*
 499  * Flags passed to hammer2_chain_delete_duplicate()
 500  */
 501 #define HAMMER2_DELDUP_RECORE           0x0001
 502
 503 /*
 504  * Cluster different types of storage together for allocations
 505  */
 506 #define HAMMER2_FREECACHE_INODE         0
 507 #define HAMMER2_FREECACHE_INDIR         1
 508 #define HAMMER2_FREECACHE_DATA          2
 509 #define HAMMER2_FREECACHE_UNUSED3       3
 510 #define HAMMER2_FREECACHE_TYPES         4
 511
 512 /*
 513  * hammer2_freemap_alloc() block preference
 514  */
 515 #define HAMMER2_OFF_NOPREF              ((hammer2_off_t)-1)
 516
 517 /*
 518  * BMAP read-ahead maximum parameters
 519  */
 520 #define HAMMER2_BMAP_COUNT              16      /* max bmap read-ahead */
 521 #define HAMMER2_BMAP_BYTES              (HAMMER2_PBUFSIZE * HAMMER2_BMAP_COUNT)
 522
 523 /*
 524  * hammer2_freemap_adjust()
 525  */
 526 #define HAMMER2_FREEMAP_DORECOVER       1
 527 #define HAMMER2_FREEMAP_DOMAYFREE       2
 528 #define HAMMER2_FREEMAP_DOREALFREE      3
 529
 530 /*
 531  * HAMMER2 cluster - A set of chains representing the same entity.
 532  *
 533  * hammer2_cluster typically represents a temporary set of representitive
 534  * chains.  The one exception is that a hammer2_cluster is embedded in
 535  * hammer2_inode.  This embedded cluster is ONLY used to track the
 536  * representitive chains and cannot be directly locked.
 537  *
 538  * A cluster is usually temporary (and thus per-thread) for locking purposes,
 539  * allowing us to embed the asynchronous storage required for cluster
 540  * operations in the cluster itself and adjust the state and status without
 541  * having to worry too much about SMP issues.
 542  *
 543  * The exception is the cluster embedded in the hammer2_inode structure.
 544  * This is used to cache the cluster state on an inode-by-inode basis.
 545  * Individual hammer2_chain structures not incorporated into clusters might
 546  * also stick around to cache miscellanious elements.
 547  *
 548  * Because the cluster is a 'working copy' and is usually subject to cluster
 549  * quorum rules, it is quite possible for us to end up with an insufficient
 550  * number of live chains to execute an operation.  If an insufficient number
 551  * of chains remain in a working copy, the operation may have to be
 552  * downgraded, retried, stall until the requisit number of chains are
 553  * available, or possibly even error out depending on the mount type.
 554  *
 555  * A cluster's focus is set when it is locked.  The focus can only be set
 556  * to a chain still part of the synchronized set.
 557  */
 558 #define HAMMER2_MAXCLUSTER      8
 559 #define HAMMER2_XOPMASK_CLUSTER ((1U << HAMMER2_MAXCLUSTER) - 1)
 560 #define HAMMER2_XOPFIFO         16
 561 #define HAMMER2_XOPFIFO_MASK    (HAMMER2_XOPFIFO - 1)
 562 #define HAMMER2_XOPGROUPS       16
 563 #define HAMMER2_XOPGROUPS_MASK  (HAMMER2_XOPGROUPS - 1)
 564 #define HAMMER2_XOPMASK_VOP     0x80000000U
 565 #define HAMMER2_XOPMASK_FIFOW   0x40000000U
 566
 567 #define HAMMER2_XOPMASK_ALLDONE (HAMMER2_XOPMASK_VOP | HAMMER2_XOPMASK_CLUSTER)
 568
 569 #define HAMMER2_SPECTHREADS     1       /* sync */
 570
 571 struct hammer2_cluster_item {
 572         hammer2_chain_t         *chain;
 573         int                     cache_index;
 574         int                     error;
 575         uint32_t                flags;
 576 };
 577
 578 typedef struct hammer2_cluster_item hammer2_cluster_item_t;
 579
 580 /*
 581  * INVALID      - Invalid for focus, i.e. not part of synchronized set.
 582  *                Once set, this bit is sticky across operations.
 583  *
 584  * FEMOD        - Indicates that front-end modifying operations can
 585  *                mess with this entry and MODSYNC will copy also
 586  *                effect it.
 587  */
 588 #define HAMMER2_CITEM_INVALID   0x00000001
 589 #define HAMMER2_CITEM_FEMOD     0x00000002
 590 #define HAMMER2_CITEM_NULL      0x00000004
 591
 592 struct hammer2_cluster {
 593         int                     refs;           /* track for deallocation */
 594         int                     ddflag;
 595         struct hammer2_pfs      *pmp;
 596         uint32_t                flags;
 597         int                     nchains;
 598         int                     error;          /* error code valid on lock */
 599         int                     focus_index;
 600         hammer2_iocb_t          iocb;
 601         hammer2_chain_t         *focus;         /* current focus (or mod) */
 602         hammer2_cluster_item_t  array[HAMMER2_MAXCLUSTER];
 603 };
 604
 605 typedef struct hammer2_cluster  hammer2_cluster_t;
 606
 607 /*
 608  * WRHARD       - Hard mounts can write fully synchronized
 609  * RDHARD       - Hard mounts can read fully synchronized
 610  * UNHARD       - Unsynchronized masters present
 611  * NOHARD       - No masters visible
 612  * WRSOFT       - Soft mounts can write to at least the SOFT_MASTER
 613  * RDSOFT       - Soft mounts can read from at least a SOFT_SLAVE
 614  * UNSOFT       - Unsynchronized slaves present
 615  * NOSOFT       - No slaves visible
 616  * RDSLAVE      - slaves are accessible (possibly unsynchronized or remote).
 617  * MSYNCED      - All masters are fully synchronized
 618  * SSYNCED      - All known local slaves are fully synchronized to masters
 619  *
 620  * All available masters are always incorporated.  All PFSs belonging to a
 621  * cluster (master, slave, copy, whatever) always try to synchronize the
 622  * total number of known masters in the PFSs root inode.
 623  *
 624  * A cluster might have access to many slaves, copies, or caches, but we
 625  * have a limited number of cluster slots.  Any such elements which are
 626  * directly mounted from block device(s) will always be incorporated.   Note
 627  * that SSYNCED only applies to such elements which are directly mounted,
 628  * not to any remote slaves, copies, or caches that could be available.  These
 629  * bits are used to monitor and drive our synchronization threads.
 630  *
 631  * When asking the question 'is any data accessible at all', then a simple
 632  * test against (RDHARD|RDSOFT|RDSLAVE) gives you the answer.  If any of
 633  * these bits are set the object can be read with certain caveats:
 634  * RDHARD - no caveats.  RDSOFT - authoritative but might not be synchronized.
 635  * and RDSLAVE - not authoritative, has some data but it could be old or
 636  * incomplete.
 637  *
 638  * When both soft and hard mounts are available, data will be read and written
 639  * via the soft mount only.  But all might be in the cluster because
 640  * background synchronization threads still need to do their work.
 641  */
 642 #define HAMMER2_CLUSTER_INODE   0x00000001      /* embedded in inode struct */
 643 #define HAMMER2_CLUSTER_UNUSED2 0x00000002
 644 #define HAMMER2_CLUSTER_LOCKED  0x00000004      /* cluster lks not recursive */
 645 #define HAMMER2_CLUSTER_WRHARD  0x00000100      /* hard-mount can write */
 646 #define HAMMER2_CLUSTER_RDHARD  0x00000200      /* hard-mount can read */
 647 #define HAMMER2_CLUSTER_UNHARD  0x00000400      /* unsynchronized masters */
 648 #define HAMMER2_CLUSTER_NOHARD  0x00000800      /* no masters visible */
 649 #define HAMMER2_CLUSTER_WRSOFT  0x00001000      /* soft-mount can write */
 650 #define HAMMER2_CLUSTER_RDSOFT  0x00002000      /* soft-mount can read */
 651 #define HAMMER2_CLUSTER_UNSOFT  0x00004000      /* unsynchronized slaves */
 652 #define HAMMER2_CLUSTER_NOSOFT  0x00008000      /* no slaves visible */
 653 #define HAMMER2_CLUSTER_MSYNCED 0x00010000      /* all masters synchronized */
 654 #define HAMMER2_CLUSTER_SSYNCED 0x00020000      /* known slaves synchronized */
 655
 656 #define HAMMER2_CLUSTER_ANYDATA ( HAMMER2_CLUSTER_RDHARD |      \
 657                                   HAMMER2_CLUSTER_RDSOFT |      \
 658                                   HAMMER2_CLUSTER_RDSLAVE)
 659
 660 #define HAMMER2_CLUSTER_RDOK    ( HAMMER2_CLUSTER_RDHARD |      \
 661                                   HAMMER2_CLUSTER_RDSOFT)
 662
 663 #define HAMMER2_CLUSTER_WROK    ( HAMMER2_CLUSTER_WRHARD |      \
 664                                   HAMMER2_CLUSTER_WRSOFT)
 665
 666 #define HAMMER2_CLUSTER_ZFLAGS  ( HAMMER2_CLUSTER_WRHARD |      \
 667                                   HAMMER2_CLUSTER_RDHARD |      \
 668                                   HAMMER2_CLUSTER_WRSOFT |      \
 669                                   HAMMER2_CLUSTER_RDSOFT |      \
 670                                   HAMMER2_CLUSTER_MSYNCED |     \
 671                                   HAMMER2_CLUSTER_SSYNCED)
 672
 673 /*
 674  * Helper functions (cluster must be locked for flags to be valid).
 675  */
 676 static __inline
 677 int
 678 hammer2_cluster_rdok(hammer2_cluster_t *cluster)
 679 {
 680         return (cluster->flags & HAMMER2_CLUSTER_RDOK);
 681 }
 682
 683 static __inline
 684 int
 685 hammer2_cluster_wrok(hammer2_cluster_t *cluster)
 686 {
 687         return (cluster->flags & HAMMER2_CLUSTER_WROK);
 688 }
 689
 690 RB_HEAD(hammer2_inode_tree, hammer2_inode);
 691
 692 /*
 693  * A hammer2 inode.
 694  *
 695  * NOTE: The inode-embedded cluster is never used directly for I/O (since
 696  *       it may be shared).  Instead it will be replicated-in and synchronized
 697  *       back out if changed.
 698  */
 699 struct hammer2_inode {
 700         RB_ENTRY(hammer2_inode) rbnode;         /* inumber lookup (HL) */
 701         hammer2_mtx_t           lock;           /* inode lock */
 702         hammer2_mtx_t           truncate_lock;  /* prevent truncates */
 703         struct hammer2_pfs      *pmp;           /* PFS mount */
 704         struct hammer2_inode    *pip;           /* parent inode */
 705         struct vnode            *vp;
 706         struct spinlock         cluster_spin;   /* update cluster */
 707         hammer2_cluster_t       cluster;
 708         struct lockf            advlock;
 709         u_int                   flags;
 710         u_int                   refs;           /* +vpref, +flushref */
 711         uint8_t                 comp_heuristic;
 712         hammer2_inode_meta_t    meta;           /* copy of meta-data */
 713         hammer2_off_t           osize;
 714         hammer2_cluster_t       *cluster_cache;
 715 };
 716
 717 typedef struct hammer2_inode hammer2_inode_t;
 718
 719 /*
 720  * MODIFIED     - Inode is in a modified state, ip->meta may have changes.
 721  * RESIZED      - Inode truncated (any) or inode extended beyond
 722  *                EMBEDDED_BYTES.
 723  */
 724 #define HAMMER2_INODE_MODIFIED          0x0001
 725 #define HAMMER2_INODE_SROOT             0x0002  /* kmalloc special case */
 726 #define HAMMER2_INODE_RENAME_INPROG     0x0004
 727 #define HAMMER2_INODE_ONRBTREE          0x0008
 728 #define HAMMER2_INODE_RESIZED           0x0010  /* requires inode_fsync */
 729 #define HAMMER2_INODE_ISDELETED         0x0020  /* deleted, not in ihidden */
 730 #define HAMMER2_INODE_ISUNLINKED        0x0040
 731 #define HAMMER2_INODE_METAGOOD          0x0080  /* inode meta-data good */
 732 #define HAMMER2_INODE_ONSIDEQ           0x0100  /* on side processing queue */
 733
 734 int hammer2_inode_cmp(hammer2_inode_t *ip1, hammer2_inode_t *ip2);
 735 RB_PROTOTYPE2(hammer2_inode_tree, hammer2_inode, rbnode, hammer2_inode_cmp,
 736                 hammer2_tid_t);
 737
 738 /*
 739  * inode-unlink side-structure
 740  */
 741 struct hammer2_inode_sideq {
 742         TAILQ_ENTRY(hammer2_inode_sideq) entry;
 743         hammer2_inode_t *ip;
 744 };
 745 TAILQ_HEAD(h2_sideq_list, hammer2_inode_sideq);
 746
 747 typedef struct hammer2_inode_sideq hammer2_inode_sideq_t;
 748
 749 /*
 750  * Transaction management sub-structure under hammer2_pfs
 751  */
 752 struct hammer2_trans {
 753         uint32_t                flags;
 754         uint32_t                sync_wait;
 755 };
 756
 757 typedef struct hammer2_trans hammer2_trans_t;
 758
 759 #define HAMMER2_TRANS_ISFLUSH           0x80000000      /* flush code */
 760 #define HAMMER2_TRANS_BUFCACHE          0x40000000      /* bio strategy */
 761 #define HAMMER2_TRANS_PREFLUSH          0x20000000      /* preflush state */
 762 #define HAMMER2_TRANS_FPENDING          0x10000000      /* flush pending */
 763 #define HAMMER2_TRANS_WAITING           0x08000000      /* someone waiting */
 764 #define HAMMER2_TRANS_MASK              0x00FFFFFF      /* count mask */
 765
 766 #define HAMMER2_FREEMAP_HEUR_NRADIX     4       /* pwr 2 PBUFRADIX-MINIORADIX */
 767 #define HAMMER2_FREEMAP_HEUR_TYPES      8
 768 #define HAMMER2_FREEMAP_HEUR_SIZE       (HAMMER2_FREEMAP_HEUR_NRADIX * \
 769                                          HAMMER2_FREEMAP_HEUR_TYPES)
 770
 771 #define HAMMER2_DEDUP_HEUR_SIZE         65536
 772 #define HAMMER2_DEDUP_HEUR_MASK         (HAMMER2_DEDUP_HEUR_SIZE - 1)
 773
 774 #define HAMMER2_FLUSH_TOP               0x0001
 775 #define HAMMER2_FLUSH_ALL               0x0002
 776
 777
 778 /*
 779  * Hammer2 support thread element.
 780  *
 781  * Potentially many support threads can hang off of hammer2, primarily
 782  * off the hammer2_pfs structure.  Typically:
 783  *
 784  * td x Nodes                   A synchronization thread for each node.
 785  * td x Nodes x workers         Worker threads for frontend operations.
 786  * td x 1                       Bioq thread for logical buffer writes.
 787  *
 788  * In addition, the synchronization thread(s) associated with the
 789  * super-root PFS (spmp) for a node is responsible for automatic bulkfree
 790  * and dedup scans.
 791  */
 792 struct hammer2_thread {
 793         struct hammer2_pfs *pmp;
 794         hammer2_xop_list_t *xopq;       /* points into pmp->xopq[] */
 795         thread_t        td;
 796         uint32_t        flags;
 797         int             depth;
 798         int             clindex;        /* cluster element index */
 799         int             repidx;
 800 };
 801
 802 typedef struct hammer2_thread hammer2_thread_t;
 803
 804 #define HAMMER2_THREAD_UNMOUNTING       0x0001  /* unmount request */
 805 #define HAMMER2_THREAD_DEV              0x0002  /* related to dev, not pfs */
 806 #define HAMMER2_THREAD_WAITING          0x0004  /* thread in idle tsleep */
 807 #define HAMMER2_THREAD_REMASTER         0x0008  /* remaster request */
 808 #define HAMMER2_THREAD_STOP             0x0010  /* exit request */
 809 #define HAMMER2_THREAD_FREEZE           0x0020  /* force idle */
 810 #define HAMMER2_THREAD_FROZEN           0x0040  /* thread is frozen */
 811 #define HAMMER2_THREAD_XOPQ             0x0080  /* work pending */
 812 #define HAMMER2_THREAD_STOPPED          0x0100  /* thread has stopped */
 813 #define HAMMER2_THREAD_UNFREEZE         0x0200
 814 #define HAMMER2_THREAD_CLIENTWAIT       0x0400
 815
 816 #define HAMMER2_THREAD_WAKEUP_MASK      (HAMMER2_THREAD_UNMOUNTING |    \
 817                                          HAMMER2_THREAD_REMASTER |      \
 818                                          HAMMER2_THREAD_STOP |          \
 819                                          HAMMER2_THREAD_FREEZE |        \
 820                                          HAMMER2_THREAD_XOPQ)
 821
 822 /*
 823  * Support structure for dedup heuristic.
 824  */
 825 struct hammer2_dedup {
 826         hammer2_off_t   data_off;
 827         uint64_t        data_crc;
 828         uint32_t        ticks;
 829         uint32_t        unused03;
 830 };
 831
 832 typedef struct hammer2_dedup hammer2_dedup_t;
 833
 834 /*
 835  * hammer2_xop - container for VOP/XOP operation (allocated, not on stack).
 836  *
 837  * This structure is used to distribute a VOP operation across multiple
 838  * nodes.  It provides a rendezvous for concurrent node execution and
 839  * can be detached from the frontend operation to allow the frontend to
 840  * return early.
 841  *
 842  * This structure also sequences operations on up to three inodes.
 843  */
 844 typedef void (*hammer2_xop_func_t)(union hammer2_xop *xop, int clidx);
 845
 846 struct hammer2_xop_fifo {
 847         TAILQ_ENTRY(hammer2_xop_head) entry;
 848         hammer2_chain_t         *array[HAMMER2_XOPFIFO];
 849         int                     errors[HAMMER2_XOPFIFO];
 850         int                     ri;
 851         int                     wi;
 852         int                     flags;
 853 };
 854
 855 typedef struct hammer2_xop_fifo hammer2_xop_fifo_t;
 856
 857 #define HAMMER2_XOP_FIFO_RUN    0x0001
 858 #define HAMMER2_XOP_FIFO_STALL  0x0002
 859
 860 struct hammer2_xop_head {
 861         hammer2_xop_func_t      func;
 862         hammer2_tid_t           mtid;
 863         struct hammer2_inode    *ip1;
 864         struct hammer2_inode    *ip2;
 865         struct hammer2_inode    *ip3;
 866         uint32_t                check_counter;
 867         uint32_t                run_mask;
 868         uint32_t                chk_mask;
 869         int                     flags;
 870         int                     state;
 871         int                     error;
 872         hammer2_key_t           collect_key;
 873         char                    *name1;
 874         size_t                  name1_len;
 875         char                    *name2;
 876         size_t                  name2_len;
 877         hammer2_xop_fifo_t      collect[HAMMER2_MAXCLUSTER];
 878         hammer2_cluster_t       cluster;        /* help collections */
 879 };
 880
 881 typedef struct hammer2_xop_head hammer2_xop_head_t;
 882
 883 struct hammer2_xop_ipcluster {
 884         hammer2_xop_head_t      head;
 885 };
 886
 887 struct hammer2_xop_strategy {
 888         hammer2_xop_head_t      head;
 889         hammer2_key_t           lbase;
 890         int                     finished;
 891         hammer2_mtx_t           lock;
 892         struct bio              *bio;
 893 };
 894
 895 struct hammer2_xop_readdir {
 896         hammer2_xop_head_t      head;
 897         hammer2_key_t           lkey;
 898 };
 899
 900 struct hammer2_xop_nresolve {
 901         hammer2_xop_head_t      head;
 902         hammer2_key_t           lhc;    /* if name is NULL used lhc */
 903 };
 904
 905 struct hammer2_xop_nlink {
 906         hammer2_xop_head_t      head;
 907         int                     nlinks_delta;
 908 };
 909
 910 struct hammer2_xop_unlink {
 911         hammer2_xop_head_t      head;
 912         int                     isdir;
 913         int                     dopermanent;
 914 };
 915
 916 struct hammer2_xop_nrename {
 917         hammer2_xop_head_t      head;
 918         hammer2_tid_t           lhc;
 919         int                     ip_key;
 920 };
 921
 922 struct hammer2_xop_scanlhc {
 923         hammer2_xop_head_t      head;
 924         hammer2_key_t           lhc;
 925 };
 926
 927 struct hammer2_xop_scanall {
 928         hammer2_xop_head_t      head;
 929         hammer2_key_t           key_beg;        /* inclusive */
 930         hammer2_key_t           key_end;        /* inclusive */
 931         int                     resolve_flags;
 932         int                     lookup_flags;
 933 };
 934
 935 struct hammer2_xop_lookup {
 936         hammer2_xop_head_t      head;
 937         hammer2_key_t           lhc;
 938 };
 939
 940 struct hammer2_xop_create {
 941         hammer2_xop_head_t      head;
 942         hammer2_inode_meta_t    meta;           /* initial metadata */
 943         hammer2_key_t           lhc;
 944         int                     flags;
 945 };
 946
 947 struct hammer2_xop_destroy {
 948         hammer2_xop_head_t      head;
 949 };
 950
 951 struct hammer2_xop_fsync {
 952         hammer2_xop_head_t      head;
 953         hammer2_inode_meta_t    meta;
 954         hammer2_off_t           osize;
 955         u_int                   ipflags;
 956         int                     clear_directdata;
 957 };
 958
 959 struct hammer2_xop_unlinkall {
 960         hammer2_xop_head_t      head;
 961         hammer2_key_t           key_beg;
 962         hammer2_key_t           key_end;
 963 };
 964
 965 struct hammer2_xop_connect {
 966         hammer2_xop_head_t      head;
 967         hammer2_key_t           lhc;
 968 };
 969
 970 struct hammer2_xop_flush {
 971         hammer2_xop_head_t      head;
 972 };
 973
 974 typedef struct hammer2_xop_readdir hammer2_xop_readdir_t;
 975 typedef struct hammer2_xop_nresolve hammer2_xop_nresolve_t;
 976 typedef struct hammer2_xop_nlink hammer2_xop_nlink_t;
 977 typedef struct hammer2_xop_unlink hammer2_xop_unlink_t;
 978 typedef struct hammer2_xop_nrename hammer2_xop_nrename_t;
 979 typedef struct hammer2_xop_ipcluster hammer2_xop_ipcluster_t;
 980 typedef struct hammer2_xop_strategy hammer2_xop_strategy_t;
 981 typedef struct hammer2_xop_create hammer2_xop_create_t;
 982 typedef struct hammer2_xop_destroy hammer2_xop_destroy_t;
 983 typedef struct hammer2_xop_fsync hammer2_xop_fsync_t;
 984 typedef struct hammer2_xop_unlinkall hammer2_xop_unlinkall_t;
 985 typedef struct hammer2_xop_scanlhc hammer2_xop_scanlhc_t;
 986 typedef struct hammer2_xop_scanall hammer2_xop_scanall_t;
 987 typedef struct hammer2_xop_lookup hammer2_xop_lookup_t;
 988 typedef struct hammer2_xop_connect hammer2_xop_connect_t;
 989 typedef struct hammer2_xop_flush hammer2_xop_flush_t;
 990
 991 union hammer2_xop {
 992         hammer2_xop_head_t      head;
 993         hammer2_xop_ipcluster_t xop_ipcluster;
 994         hammer2_xop_readdir_t   xop_readdir;
 995         hammer2_xop_nresolve_t  xop_nresolve;
 996         hammer2_xop_nlink_t     xop_nlink;
 997         hammer2_xop_unlink_t    xop_unlink;
 998         hammer2_xop_nrename_t   xop_nrename;
 999         hammer2_xop_strategy_t  xop_strategy;
1000         hammer2_xop_create_t    xop_create;
1001         hammer2_xop_destroy_t   xop_destroy;
1002         hammer2_xop_fsync_t     xop_fsync;
1003         hammer2_xop_unlinkall_t xop_unlinkall;
1004         hammer2_xop_scanlhc_t   xop_scanlhc;
1005         hammer2_xop_scanall_t   xop_scanall;
1006         hammer2_xop_lookup_t    xop_lookup;
1007         hammer2_xop_flush_t     xop_flush;
1008         hammer2_xop_connect_t   xop_connect;
1009 };
1010
1011 typedef union hammer2_xop hammer2_xop_t;
1012
1013 /*
1014  * hammer2_xop_group - Manage XOP support threads.
1015  */
1016 struct hammer2_xop_group {
1017         hammer2_thread_t        thrs[HAMMER2_MAXCLUSTER];
1018 };
1019
1020 typedef struct hammer2_xop_group hammer2_xop_group_t;
1021
1022 /*
1023  * flags to hammer2_xop_collect()
1024  */
1025 #define HAMMER2_XOP_COLLECT_NOWAIT      0x00000001
1026 #define HAMMER2_XOP_COLLECT_WAITALL     0x00000002
1027
1028 /*
1029  * flags to hammer2_xop_alloc()
1030  *
1031  * MODIFYING    - This is a modifying transaction, allocate a mtid.
1032  */
1033 #define HAMMER2_XOP_MODIFYING           0x00000001
1034 #define HAMMER2_XOP_ITERATOR            0x00000002
1035
1036 /*
1037  * Global (per partition) management structure, represents a hard block
1038  * device.  Typically referenced by hammer2_chain structures when applicable.
1039  * Typically not used for network-managed elements.
1040  *
1041  * Note that a single hammer2_dev can be indirectly tied to multiple system
1042  * mount points.  There is no direct relationship.  System mounts are
1043  * per-cluster-id, not per-block-device, and a single hard mount might contain
1044  * many PFSs and those PFSs might combine together in various ways to form
1045  * the set of available clusters.
1046  */
1047 struct hammer2_dev {
1048         struct vnode    *devvp;         /* device vnode */
1049         int             ronly;          /* read-only mount */
1050         int             mount_count;    /* number of actively mounted PFSs */
1051         TAILQ_ENTRY(hammer2_dev) mntentry; /* hammer2_mntlist */
1052
1053         struct malloc_type *mchain;
1054         int             nipstacks;
1055         int             maxipstacks;
1056         kdmsg_iocom_t   iocom;          /* volume-level dmsg interface */
1057         struct spinlock io_spin;        /* iotree access */
1058         struct hammer2_io_tree iotree;
1059         int             iofree_count;
1060         hammer2_chain_t vchain;         /* anchor chain (topology) */
1061         hammer2_chain_t fchain;         /* anchor chain (freemap) */
1062         struct spinlock list_spin;
1063         struct h2_flush_list    flushq; /* flush seeds */
1064         struct hammer2_pfs *spmp;       /* super-root pmp for transactions */
1065         struct lock     bulklk;         /* bulkfree lock */
1066         struct lock     vollk;          /* lockmgr lock */
1067         hammer2_off_t   heur_freemap[HAMMER2_FREEMAP_HEUR_SIZE];
1068         hammer2_dedup_t heur_dedup[HAMMER2_DEDUP_HEUR_SIZE];
1069         int             volhdrno;       /* last volhdrno written */
1070         int             hflags;         /* HMNT2 flags applicable to device */
1071         char            devrepname[64]; /* for kprintf */
1072         hammer2_volume_data_t voldata;
1073         hammer2_volume_data_t volsync;  /* synchronized voldata */
1074 };
1075
1076 typedef struct hammer2_dev hammer2_dev_t;
1077
1078 /*
1079  * Helper functions (cluster must be locked for flags to be valid).
1080  */
1081 static __inline
1082 int
1083 hammer2_chain_rdok(hammer2_chain_t *chain)
1084 {
1085         return (chain->error == 0);
1086 }
1087
1088 static __inline
1089 int
1090 hammer2_chain_wrok(hammer2_chain_t *chain)
1091 {
1092         return (chain->error == 0 && chain->hmp->ronly == 0);
1093 }
1094
1095 /*
1096  * Per-cluster management structure.  This structure will be tied to a
1097  * system mount point if the system is mounting the PFS, but is also used
1098  * to manage clusters encountered during the super-root scan or received
1099  * via LNK_SPANs that might not be mounted.
1100  *
1101  * This structure is also used to represent the super-root that hangs off
1102  * of a hard mount point.  The super-root is not really a cluster element.
1103  * In this case the spmp_hmp field will be non-NULL.  It's just easier to do
1104  * this than to special case super-root manipulation in the hammer2_chain*
1105  * code as being only hammer2_dev-related.
1106  *
1107  * pfs_mode and pfs_nmasters are rollup fields which critically describes
1108  * how elements of the cluster act on the cluster.  pfs_mode is only applicable
1109  * when a PFS is mounted by the system.  pfs_nmasters is our best guess as to
1110  * how many masters have been configured for a cluster and is always
1111  * applicable.  pfs_types[] is an array with 1:1 correspondance to the
1112  * iroot cluster and describes the PFS types of the nodes making up the
1113  * cluster.
1114  *
1115  * WARNING! Portions of this structure have deferred initialization.  In
1116  *          particular, if not mounted there will be no ihidden or wthread.
1117  *          umounted network PFSs will also be missing iroot and numerous
1118  *          other fields will not be initialized prior to mount.
1119  *
1120  *          Synchronization threads are chain-specific and only applicable
1121  *          to local hard PFS entries.  A hammer2_pfs structure may contain
1122  *          more than one when multiple hard PFSs are present on the local
1123  *          machine which require synchronization monitoring.  Most PFSs
1124  *          (such as snapshots) are 1xMASTER PFSs which do not need a
1125  *          synchronization thread.
1126  *
1127  * WARNING! The chains making up pfs->iroot's cluster are accounted for in
1128  *          hammer2_dev->mount_count when the pfs is associated with a mount
1129  *          point.
1130  */
1131 struct hammer2_pfs {
1132         struct mount            *mp;
1133         TAILQ_ENTRY(hammer2_pfs) mntentry;      /* hammer2_pfslist */
1134         uuid_t                  pfs_clid;
1135         hammer2_dev_t           *spmp_hmp;      /* only if super-root pmp */
1136         hammer2_dev_t           *force_local;   /* only if 'local' mount */
1137         hammer2_inode_t         *iroot;         /* PFS root inode */
1138         hammer2_inode_t         *ihidden;       /* PFS hidden directory */
1139         uint8_t                 pfs_types[HAMMER2_MAXCLUSTER];
1140         char                    *pfs_names[HAMMER2_MAXCLUSTER];
1141         hammer2_dev_t           *pfs_hmps[HAMMER2_MAXCLUSTER];
1142         hammer2_trans_t         trans;
1143         struct lock             lock;           /* PFS lock for certain ops */
1144         struct lock             lock_nlink;     /* rename and nlink lock */
1145         struct netexport        export;         /* nfs export */
1146         int                     ronly;          /* read-only mount */
1147         int                     hflags;         /* pfs-specific mount flags */
1148         struct malloc_type      *minode;
1149         struct malloc_type      *mmsg;
1150         struct spinlock         inum_spin;      /* inumber lookup */
1151         struct hammer2_inode_tree inum_tree;    /* (not applicable to spmp) */
1152         struct spinlock         lru_spin;       /* inumber lookup */
1153         struct hammer2_chain_list lru_list;     /* chains on LRU */
1154         int                     lru_count;      /* #of chains on LRU */
1155         hammer2_tid_t           modify_tid;     /* modify transaction id */
1156         hammer2_tid_t           inode_tid;      /* inode allocator */
1157         uint8_t                 pfs_nmasters;   /* total masters */
1158         uint8_t                 pfs_mode;       /* operating mode PFSMODE */
1159         uint8_t                 unused01;
1160         uint8_t                 unused02;
1161         int                     xop_iterator;
1162         long                    inmem_inodes;
1163         uint32_t                inmem_dirty_chains;
1164         int                     count_lwinprog; /* logical write in prog */
1165         struct spinlock         list_spin;
1166         struct h2_sideq_list    sideq;          /* last-close dirty/unlink */
1167         hammer2_thread_t        sync_thrs[HAMMER2_MAXCLUSTER];
1168         uint32_t                cluster_flags;  /* cached cluster flags */
1169         int                     has_xop_threads;
1170         struct spinlock         xop_spin;       /* xop sequencer */
1171         hammer2_xop_group_t     xop_groups[HAMMER2_XOPGROUPS];
1172         hammer2_xop_list_t      xopq[HAMMER2_MAXCLUSTER][HAMMER2_XOPGROUPS+
1173                                                          HAMMER2_SPECTHREADS];
1174 };
1175
1176 typedef struct hammer2_pfs hammer2_pfs_t;
1177
1178 #define HAMMER2_LRU_LIMIT               1024    /* per pmp lru_list */
1179
1180 #define HAMMER2_DIRTYCHAIN_WAITING      0x80000000
1181 #define HAMMER2_DIRTYCHAIN_MASK         0x7FFFFFFF
1182
1183 #define HAMMER2_LWINPROG_WAITING        0x80000000
1184 #define HAMMER2_LWINPROG_WAITING0       0x40000000
1185 #define HAMMER2_LWINPROG_MASK           0x3FFFFFFF
1186
1187 /*
1188  * hammer2_cluster_check
1189  */
1190 #define HAMMER2_CHECK_NULL      0x00000001
1191
1192 /*
1193  * Bulkscan
1194  */
1195 #define HAMMER2_BULK_ABORT      0x00000001
1196
1197 /*
1198  * Misc
1199  */
1200 #if defined(_KERNEL)
1201
1202 MALLOC_DECLARE(M_HAMMER2);
1203
1204 #define VTOI(vp)        ((hammer2_inode_t *)(vp)->v_data)
1205 #define ITOV(ip)        ((ip)->vp)
1206
1207 /*
1208  * Currently locked chains retain the locked buffer cache buffer for
1209  * indirect blocks, and indirect blocks can be one of two sizes.  The
1210  * device buffer has to match the case to avoid deadlocking recursive
1211  * chains that might otherwise try to access different offsets within
1212  * the same device buffer.
1213  */
1214 static __inline
1215 int
1216 hammer2_devblkradix(int radix)
1217 {
1218 #if 0
1219         if (radix <= HAMMER2_LBUFRADIX) {
1220                 return (HAMMER2_LBUFRADIX);
1221         } else {
1222                 return (HAMMER2_PBUFRADIX);
1223         }
1224 #endif
1225         return (HAMMER2_PBUFRADIX);
1226 }
1227
1228 /*
1229  * XXX almost time to remove this.  DIO uses PBUFSIZE exclusively now.
1230  */
1231 static __inline
1232 size_t
1233 hammer2_devblksize(size_t bytes)
1234 {
1235 #if 0
1236         if (bytes <= HAMMER2_LBUFSIZE) {
1237                 return(HAMMER2_LBUFSIZE);
1238         } else {
1239                 KKASSERT(bytes <= HAMMER2_PBUFSIZE &&
1240                          (bytes ^ (bytes - 1)) == ((bytes << 1) - 1));
1241                 return (HAMMER2_PBUFSIZE);
1242         }
1243 #endif
1244         return (HAMMER2_PBUFSIZE);
1245 }
1246
1247
1248 static __inline
1249 hammer2_pfs_t *
1250 MPTOPMP(struct mount *mp)
1251 {
1252         return ((hammer2_pfs_t *)mp->mnt_data);
1253 }
1254
1255 #define LOCKSTART       int __nlocks = curthread->td_locks
1256 #define LOCKENTER       (++curthread->td_locks)
1257 #define LOCKEXIT        (--curthread->td_locks)
1258 #define LOCKSTOP        KKASSERT(curthread->td_locks == __nlocks)
1259
1260 extern struct vop_ops hammer2_vnode_vops;
1261 extern struct vop_ops hammer2_spec_vops;
1262 extern struct vop_ops hammer2_fifo_vops;
1263
1264 extern int hammer2_debug;
1265 extern int hammer2_cluster_enable;
1266 extern int hammer2_hardlink_enable;
1267 extern int hammer2_flush_pipe;
1268 extern int hammer2_synchronous_flush;
1269 extern int hammer2_dio_count;
1270 extern long hammer2_chain_allocs;
1271 extern long hammer2_chain_frees;
1272 extern long hammer2_limit_dirty_chains;
1273 extern long hammer2_count_modified_chains;
1274 extern long hammer2_iod_file_read;
1275 extern long hammer2_iod_meta_read;
1276 extern long hammer2_iod_indr_read;
1277 extern long hammer2_iod_fmap_read;
1278 extern long hammer2_iod_volu_read;
1279 extern long hammer2_iod_file_write;
1280 extern long hammer2_iod_file_wembed;
1281 extern long hammer2_iod_file_wzero;
1282 extern long hammer2_iod_file_wdedup;
1283 extern long hammer2_iod_meta_write;
1284 extern long hammer2_iod_indr_write;
1285 extern long hammer2_iod_fmap_write;
1286 extern long hammer2_iod_volu_write;
1287 extern long hammer2_ioa_file_read;
1288 extern long hammer2_ioa_meta_read;
1289 extern long hammer2_ioa_indr_read;
1290 extern long hammer2_ioa_fmap_read;
1291 extern long hammer2_ioa_volu_read;
1292 extern long hammer2_ioa_file_write;
1293 extern long hammer2_ioa_meta_write;
1294 extern long hammer2_ioa_indr_write;
1295 extern long hammer2_ioa_fmap_write;
1296 extern long hammer2_ioa_volu_write;
1297
1298 extern long hammer2_check_xxhash64;
1299 extern long hammer2_check_icrc32;
1300
1301 extern struct objcache *cache_buffer_read;
1302 extern struct objcache *cache_buffer_write;
1303 extern struct objcache *cache_xops;
1304
1305 /*
1306  * hammer2_subr.c
1307  */
1308 #define hammer2_icrc32(buf, size)       iscsi_crc32((buf), (size))
1309 #define hammer2_icrc32c(buf, size, crc) iscsi_crc32_ext((buf), (size), (crc))
1310
1311 int hammer2_signal_check(time_t *timep);
1312 const char *hammer2_error_str(int error);
1313
1314 void hammer2_inode_lock(hammer2_inode_t *ip, int how);
1315 void hammer2_inode_unlock(hammer2_inode_t *ip);
1316 hammer2_chain_t *hammer2_inode_chain(hammer2_inode_t *ip, int clindex, int how);
1317 hammer2_chain_t *hammer2_inode_chain_and_parent(hammer2_inode_t *ip,
1318                         int clindex, hammer2_chain_t **parentp, int how);
1319 hammer2_mtx_state_t hammer2_inode_lock_temp_release(hammer2_inode_t *ip);
1320 void hammer2_inode_lock_temp_restore(hammer2_inode_t *ip,
1321                         hammer2_mtx_state_t ostate);
1322 int hammer2_inode_lock_upgrade(hammer2_inode_t *ip);
1323 void hammer2_inode_lock_downgrade(hammer2_inode_t *ip, int);
1324
1325 void hammer2_dev_exlock(hammer2_dev_t *hmp);
1326 void hammer2_dev_shlock(hammer2_dev_t *hmp);
1327 void hammer2_dev_unlock(hammer2_dev_t *hmp);
1328
1329 int hammer2_get_dtype(const hammer2_inode_data_t *ipdata);
1330 int hammer2_get_vtype(uint8_t type);
1331 uint8_t hammer2_get_obj_type(enum vtype vtype);
1332 void hammer2_time_to_timespec(uint64_t xtime, struct timespec *ts);
1333 uint64_t hammer2_timespec_to_time(const struct timespec *ts);
1334 uint32_t hammer2_to_unix_xid(const uuid_t *uuid);
1335 void hammer2_guid_to_uuid(uuid_t *uuid, uint32_t guid);
1336 void hammer2_trans_manage_init(hammer2_pfs_t *pmp);
1337
1338 hammer2_key_t hammer2_dirhash(const unsigned char *name, size_t len);
1339 int hammer2_getradix(size_t bytes);
1340
1341 int hammer2_calc_logical(hammer2_inode_t *ip, hammer2_off_t uoff,
1342                         hammer2_key_t *lbasep, hammer2_key_t *leofp);
1343 int hammer2_calc_physical(hammer2_inode_t *ip, hammer2_key_t lbase);
1344 void hammer2_update_time(uint64_t *timep);
1345 void hammer2_adjreadcounter(hammer2_blockref_t *bref, size_t bytes);
1346
1347 /*
1348  * hammer2_inode.c
1349  */
1350 struct vnode *hammer2_igetv(hammer2_inode_t *ip, int *errorp);
1351 hammer2_inode_t *hammer2_inode_lookup(hammer2_pfs_t *pmp,
1352                         hammer2_tid_t inum);
1353 hammer2_inode_t *hammer2_inode_get(hammer2_pfs_t *pmp, hammer2_inode_t *dip,
1354                         hammer2_cluster_t *cluster, int idx);
1355 void hammer2_inode_free(hammer2_inode_t *ip);
1356 void hammer2_inode_ref(hammer2_inode_t *ip);
1357 void hammer2_inode_drop(hammer2_inode_t *ip);
1358 void hammer2_inode_repoint(hammer2_inode_t *ip, hammer2_inode_t *pip,
1359                         hammer2_cluster_t *cluster);
1360 void hammer2_inode_repoint_one(hammer2_inode_t *ip, hammer2_cluster_t *cluster,
1361                         int idx);
1362 void hammer2_inode_modify(hammer2_inode_t *ip);
1363 void hammer2_inode_run_sideq(hammer2_pfs_t *pmp);
1364
1365 hammer2_inode_t *hammer2_inode_create(hammer2_inode_t *dip,
1366                         struct vattr *vap, struct ucred *cred,
1367                         const uint8_t *name, size_t name_len, hammer2_key_t lhc,
1368                         hammer2_key_t inum, uint8_t type, uint8_t target_type,
1369                         int flags, int *errorp);
1370 int hammer2_inode_connect(hammer2_inode_t *dip, hammer2_inode_t *ip,
1371                         const char *name, size_t name_len,
1372                         hammer2_key_t lhc);
1373 hammer2_inode_t *hammer2_inode_common_parent(hammer2_inode_t *fdip,
1374                         hammer2_inode_t *tdip, int *errorp, int ishardlink);
1375 void hammer2_inode_chain_sync(hammer2_inode_t *ip);
1376 int hammer2_inode_unlink_finisher(hammer2_inode_t *ip, int isopen);
1377 void hammer2_inode_install_hidden(hammer2_pfs_t *pmp);
1378
1379 /*
1380  * hammer2_chain.c
1381  */
1382 void hammer2_voldata_lock(hammer2_dev_t *hmp);
1383 void hammer2_voldata_unlock(hammer2_dev_t *hmp);
1384 void hammer2_voldata_modify(hammer2_dev_t *hmp);
1385 hammer2_chain_t *hammer2_chain_alloc(hammer2_dev_t *hmp,
1386                                 hammer2_pfs_t *pmp,
1387                                 hammer2_blockref_t *bref);
1388 void hammer2_chain_core_init(hammer2_chain_t *chain);
1389 void hammer2_chain_ref(hammer2_chain_t *chain);
1390 void hammer2_chain_drop(hammer2_chain_t *chain);
1391 void hammer2_chain_lock(hammer2_chain_t *chain, int how);
1392 void hammer2_chain_lock_downgrade(hammer2_chain_t *chain);
1393 void hammer2_chain_push_shared_lock(hammer2_chain_t *chain);
1394 void hammer2_chain_pull_shared_lock(hammer2_chain_t *chain);
1395 void hammer2_chain_load_data(hammer2_chain_t *chain);
1396 const hammer2_media_data_t *hammer2_chain_rdata(hammer2_chain_t *chain);
1397 hammer2_media_data_t *hammer2_chain_wdata(hammer2_chain_t *chain);
1398 int hammer2_chain_snapshot(hammer2_chain_t *chain, hammer2_ioc_pfs_t *pmp,
1399                                 hammer2_tid_t mtid);
1400
1401 int hammer2_chain_hardlink_find(hammer2_inode_t *dip,
1402                                 hammer2_chain_t **parentp,
1403                                 hammer2_chain_t **chainp,
1404                                 int flags);
1405 void hammer2_chain_modify(hammer2_chain_t *chain, hammer2_tid_t mtid,
1406                                 hammer2_off_t dedup_off, int flags);
1407 void hammer2_chain_modify_ip(hammer2_inode_t *ip, hammer2_chain_t *chain,
1408                                 hammer2_tid_t mtid, int flags);
1409 void hammer2_chain_resize(hammer2_inode_t *ip, hammer2_chain_t *parent,
1410                                 hammer2_chain_t *chain,
1411                                 hammer2_tid_t mtid, hammer2_off_t dedup_off,
1412                                 int nradix, int flags);
1413 void hammer2_chain_unlock(hammer2_chain_t *chain);
1414 void hammer2_chain_wait(hammer2_chain_t *chain);
1415 hammer2_chain_t *hammer2_chain_get(hammer2_chain_t *parent, int generation,
1416                                 hammer2_blockref_t *bref);
1417 hammer2_chain_t *hammer2_chain_lookup_init(hammer2_chain_t *parent, int flags);
1418 void hammer2_chain_lookup_done(hammer2_chain_t *parent);
1419 hammer2_chain_t *hammer2_chain_getparent(hammer2_chain_t **parentp, int how);
1420 hammer2_chain_t *hammer2_chain_lookup(hammer2_chain_t **parentp,
1421                                 hammer2_key_t *key_nextp,
1422                                 hammer2_key_t key_beg, hammer2_key_t key_end,
1423                                 int *cache_indexp, int flags);
1424 hammer2_chain_t *hammer2_chain_next(hammer2_chain_t **parentp,
1425                                 hammer2_chain_t *chain,
1426                                 hammer2_key_t *key_nextp,
1427                                 hammer2_key_t key_beg, hammer2_key_t key_end,
1428                                 int *cache_indexp, int flags);
1429 hammer2_blockref_t *hammer2_chain_scan(hammer2_chain_t *parent,
1430                                 hammer2_chain_t **chainp,
1431                                 hammer2_blockref_t *bref,
1432                                 int *firstp, int *cache_indexp, int flags);
1433
1434 int hammer2_chain_create(hammer2_chain_t **parentp,
1435                                 hammer2_chain_t **chainp, hammer2_pfs_t *pmp,
1436                                 hammer2_key_t key, int keybits,
1437                                 int type, size_t bytes, hammer2_tid_t mtid,
1438                                 hammer2_off_t dedup_off, int flags);
1439 void hammer2_chain_rename(hammer2_blockref_t *bref,
1440                                 hammer2_chain_t **parentp,
1441                                 hammer2_chain_t *chain,
1442                                 hammer2_tid_t mtid, int flags);
1443 void hammer2_chain_delete(hammer2_chain_t *parent, hammer2_chain_t *chain,
1444                                 hammer2_tid_t mtid, int flags);
1445 void hammer2_chain_setflush(hammer2_chain_t *chain);
1446 void hammer2_chain_countbrefs(hammer2_chain_t *chain,
1447                                 hammer2_blockref_t *base, int count);
1448 hammer2_chain_t *hammer2_chain_bulksnap(hammer2_chain_t *chain);
1449 void hammer2_chain_bulkdrop(hammer2_chain_t *copy);
1450
1451 void hammer2_chain_setcheck(hammer2_chain_t *chain, void *bdata);
1452 int hammer2_chain_testcheck(hammer2_chain_t *chain, void *bdata);
1453
1454
1455 void hammer2_pfs_memory_wait(hammer2_pfs_t *pmp);
1456 void hammer2_pfs_memory_inc(hammer2_pfs_t *pmp);
1457 void hammer2_pfs_memory_wakeup(hammer2_pfs_t *pmp);
1458
1459 void hammer2_base_delete(hammer2_chain_t *chain,
1460                                 hammer2_blockref_t *base, int count,
1461                                 int *cache_indexp, hammer2_chain_t *child);
1462 void hammer2_base_insert(hammer2_chain_t *chain,
1463                                 hammer2_blockref_t *base, int count,
1464                                 int *cache_indexp, hammer2_chain_t *child);
1465
1466 /*
1467  * hammer2_flush.c
1468  */
1469 void hammer2_flush(hammer2_chain_t *chain, int istop);
1470 void hammer2_flush_quick(hammer2_dev_t *hmp);
1471 void hammer2_delayed_flush(hammer2_chain_t *chain);
1472
1473 /*
1474  * hammer2_trans.c
1475  */
1476 void hammer2_trans_init(hammer2_pfs_t *pmp, uint32_t flags);
1477 hammer2_tid_t hammer2_trans_sub(hammer2_pfs_t *pmp);
1478 void hammer2_trans_clear_preflush(hammer2_pfs_t *pmp);
1479 void hammer2_trans_done(hammer2_pfs_t *pmp);
1480 hammer2_tid_t hammer2_trans_newinum(hammer2_pfs_t *pmp);
1481 void hammer2_trans_assert_strategy(hammer2_pfs_t *pmp);
1482
1483 /*
1484  * hammer2_ioctl.c
1485  */
1486 int hammer2_ioctl(hammer2_inode_t *ip, u_long com, void *data,
1487                                 int fflag, struct ucred *cred);
1488
1489 /*
1490  * hammer2_io.c
1491  */
1492 void hammer2_io_putblk(hammer2_io_t **diop);
1493 void hammer2_io_cleanup(hammer2_dev_t *hmp, struct hammer2_io_tree *tree);
1494 char *hammer2_io_data(hammer2_io_t *dio, off_t lbase);
1495 hammer2_io_t *hammer2_io_getquick(hammer2_dev_t *hmp, off_t lbase, int lsize);
1496 void hammer2_io_getblk(hammer2_dev_t *hmp, off_t lbase, int lsize,
1497                                 hammer2_iocb_t *iocb);
1498 void hammer2_io_complete(hammer2_iocb_t *iocb);
1499 void hammer2_io_callback(struct bio *bio);
1500 void hammer2_iocb_wait(hammer2_iocb_t *iocb);
1501 int hammer2_io_new(hammer2_dev_t *hmp, int btype, off_t lbase, int lsize,
1502                                 hammer2_io_t **diop);
1503 int hammer2_io_newnz(hammer2_dev_t *hmp, int btype, off_t lbase, int lsize,
1504                                 hammer2_io_t **diop);
1505 int hammer2_io_newq(hammer2_dev_t *hmp, int btype, off_t lbase, int lsize,
1506                                 hammer2_io_t **diop);
1507 int hammer2_io_bread(hammer2_dev_t *hmp, int btype, off_t lbase, int lsize,
1508                                 hammer2_io_t **diop);
1509 void hammer2_io_bawrite(hammer2_io_t **diop);
1510 void hammer2_io_bdwrite(hammer2_io_t **diop);
1511 int hammer2_io_bwrite(hammer2_io_t **diop);
1512 int hammer2_io_isdirty(hammer2_io_t *dio);
1513 void hammer2_io_setdirty(hammer2_io_t *dio);
1514 void hammer2_io_setinval(hammer2_io_t *dio, hammer2_off_t off, u_int bytes);
1515 void hammer2_io_brelse(hammer2_io_t **diop);
1516 void hammer2_io_bqrelse(hammer2_io_t **diop);
1517 int hammer2_io_crc_good(hammer2_chain_t *chain, uint64_t *maskp);
1518 void hammer2_io_crc_setmask(hammer2_io_t *dio, uint64_t mask);
1519 void hammer2_io_crc_clrmask(hammer2_io_t *dio, uint64_t mask);
1520
1521 /*
1522  * hammer2_thread.c
1523  */
1524 void hammer2_thr_signal(hammer2_thread_t *thr, uint32_t flags);
1525 void hammer2_thr_return(hammer2_thread_t *thr, uint32_t flags);
1526 void hammer2_thr_wait(hammer2_thread_t *thr, uint32_t flags);
1527 void hammer2_thr_wait_neg(hammer2_thread_t *thr, uint32_t flags);
1528 void hammer2_thr_create(hammer2_thread_t *thr, hammer2_pfs_t *pmp,
1529                         const char *id, int clindex, int repidx,
1530                         void (*func)(void *arg));
1531 void hammer2_thr_delete(hammer2_thread_t *thr);
1532 void hammer2_thr_remaster(hammer2_thread_t *thr);
1533 void hammer2_thr_freeze_async(hammer2_thread_t *thr);
1534 void hammer2_thr_freeze(hammer2_thread_t *thr);
1535 void hammer2_thr_unfreeze(hammer2_thread_t *thr);
1536 int hammer2_thr_break(hammer2_thread_t *thr);
1537 void hammer2_primary_xops_thread(void *arg);
1538
1539 /*
1540  * hammer2_thread.c (XOP API)
1541  */
1542 void hammer2_xop_group_init(hammer2_pfs_t *pmp, hammer2_xop_group_t *xgrp);
1543 void *hammer2_xop_alloc(hammer2_inode_t *ip, int flags);
1544 void hammer2_xop_setname(hammer2_xop_head_t *xop,
1545                                 const char *name, size_t name_len);
1546 void hammer2_xop_setname2(hammer2_xop_head_t *xop,
1547                                 const char *name, size_t name_len);
1548 void hammer2_xop_setip2(hammer2_xop_head_t *xop, hammer2_inode_t *ip2);
1549 void hammer2_xop_setip3(hammer2_xop_head_t *xop, hammer2_inode_t *ip3);
1550 void hammer2_xop_reinit(hammer2_xop_head_t *xop);
1551 void hammer2_xop_helper_create(hammer2_pfs_t *pmp);
1552 void hammer2_xop_helper_cleanup(hammer2_pfs_t *pmp);
1553 void hammer2_xop_start(hammer2_xop_head_t *xop, hammer2_xop_func_t func);
1554 void hammer2_xop_start_except(hammer2_xop_head_t *xop, hammer2_xop_func_t func,
1555                                 int notidx);
1556 int hammer2_xop_collect(hammer2_xop_head_t *xop, int flags);
1557 void hammer2_xop_retire(hammer2_xop_head_t *xop, uint32_t mask);
1558 int hammer2_xop_active(hammer2_xop_head_t *xop);
1559 int hammer2_xop_feed(hammer2_xop_head_t *xop, hammer2_chain_t *chain,
1560                                 int clindex, int error);
1561
1562 /*
1563  * hammer2_synchro.c
1564  */
1565 void hammer2_primary_sync_thread(void *arg);
1566
1567 /*
1568  * XOP backends in hammer2_xops.c, primarily for VNOPS.  Other XOP backends
1569  * may be integrated into other source files.
1570  */
1571 void hammer2_xop_ipcluster(hammer2_xop_t *xop, int clidx);
1572 void hammer2_xop_readdir(hammer2_xop_t *xop, int clidx);
1573 void hammer2_xop_nresolve(hammer2_xop_t *xop, int clidx);
1574 void hammer2_xop_unlink(hammer2_xop_t *xop, int clidx);
1575 void hammer2_xop_nrename(hammer2_xop_t *xop, int clidx);
1576 void hammer2_xop_nlink(hammer2_xop_t *xop, int clidx);
1577 void hammer2_xop_scanlhc(hammer2_xop_t *xop, int clidx);
1578 void hammer2_xop_scanall(hammer2_xop_t *xop, int clidx);
1579 void hammer2_xop_lookup(hammer2_xop_t *xop, int clidx);
1580 void hammer2_inode_xop_create(hammer2_xop_t *xop, int clidx);
1581 void hammer2_inode_xop_destroy(hammer2_xop_t *xop, int clidx);
1582 void hammer2_inode_xop_chain_sync(hammer2_xop_t *xop, int clidx);
1583 void hammer2_inode_xop_unlinkall(hammer2_xop_t *xop, int clidx);
1584 void hammer2_inode_xop_connect(hammer2_xop_t *xop, int clidx);
1585 void hammer2_inode_xop_flush(hammer2_xop_t *xop, int clidx);
1586
1587 /*
1588  * hammer2_msgops.c
1589  */
1590 int hammer2_msg_dbg_rcvmsg(kdmsg_msg_t *msg);
1591 int hammer2_msg_adhoc_input(kdmsg_msg_t *msg);
1592
1593 /*
1594  * hammer2_vfsops.c
1595  */
1596 void hammer2_volconf_update(hammer2_dev_t *hmp, int index);
1597 void hammer2_dump_chain(hammer2_chain_t *chain, int tab, int *countp, char pfx);
1598 int hammer2_vfs_sync(struct mount *mp, int waitflags);
1599 hammer2_pfs_t *hammer2_pfsalloc(hammer2_chain_t *chain,
1600                                 const hammer2_inode_data_t *ripdata,
1601                                 hammer2_tid_t modify_tid,
1602                                 hammer2_dev_t *force_local);
1603
1604 void hammer2_lwinprog_ref(hammer2_pfs_t *pmp);
1605 void hammer2_lwinprog_drop(hammer2_pfs_t *pmp);
1606 void hammer2_lwinprog_wait(hammer2_pfs_t *pmp, int pipe);
1607
1608 /*
1609  * hammer2_freemap.c
1610  */
1611 int hammer2_freemap_alloc(hammer2_chain_t *chain, size_t bytes);
1612 void hammer2_freemap_adjust(hammer2_dev_t *hmp,
1613                                 hammer2_blockref_t *bref, int how);
1614
1615 /*
1616  * hammer2_cluster.c
1617  */
1618 uint8_t hammer2_cluster_type(hammer2_cluster_t *cluster);
1619 const hammer2_media_data_t *hammer2_cluster_rdata(hammer2_cluster_t *cluster);
1620 hammer2_media_data_t *hammer2_cluster_wdata(hammer2_cluster_t *cluster);
1621 hammer2_cluster_t *hammer2_cluster_from_chain(hammer2_chain_t *chain);
1622 void hammer2_cluster_bref(hammer2_cluster_t *cluster, hammer2_blockref_t *bref);
1623 hammer2_cluster_t *hammer2_cluster_alloc(hammer2_pfs_t *pmp,
1624                                 hammer2_blockref_t *bref);
1625 void hammer2_cluster_ref(hammer2_cluster_t *cluster);
1626 void hammer2_cluster_drop(hammer2_cluster_t *cluster);
1627 void hammer2_cluster_lock(hammer2_cluster_t *cluster, int how);
1628 int hammer2_cluster_check(hammer2_cluster_t *cluster, hammer2_key_t lokey,
1629                         int flags);
1630 void hammer2_cluster_resolve(hammer2_cluster_t *cluster);
1631 void hammer2_cluster_forcegood(hammer2_cluster_t *cluster);
1632 hammer2_cluster_t *hammer2_cluster_copy(hammer2_cluster_t *ocluster);
1633 void hammer2_cluster_unlock(hammer2_cluster_t *cluster);
1634
1635 int hammer2_bulkfree_pass(hammer2_dev_t *hmp,
1636                         struct hammer2_ioc_bulkfree *bfi);
1637
1638 /*
1639  * hammer2_iocom.c
1640  */
1641 void hammer2_iocom_init(hammer2_dev_t *hmp);
1642 void hammer2_iocom_uninit(hammer2_dev_t *hmp);
1643 void hammer2_cluster_reconnect(hammer2_dev_t *hmp, struct file *fp);
1644
1645 /*
1646  * hammer2_strategy.c
1647  */
1648 int hammer2_vop_strategy(struct vop_strategy_args *ap);
1649 int hammer2_vop_bmap(struct vop_bmap_args *ap);
1650 void hammer2_write_thread(void *arg);
1651 void hammer2_bioq_sync(hammer2_pfs_t *pmp);
1652 void hammer2_dedup_clear(hammer2_dev_t *hmp);
1653
1654 #endif /* !_KERNEL */
1655 #endif /* !_VFS_HAMMER2_HAMMER2_H_ */