sys/vfs/hammer2/hammer2.h

   1 /*
   2  * Copyright (c) 2011-2018 The DragonFly Project.  All rights reserved.
   3  *
   4  * This code is derived from software contributed to The DragonFly Project
   5  * by Matthew Dillon <dillon@dragonflybsd.org>
   6  * by Venkatesh Srinivas <vsrinivas@dragonflybsd.org>
   7  *
   8  * Redistribution and use in source and binary forms, with or without
   9  * modification, are permitted provided that the following conditions
  10  * are met:
  11  *
  12  * 1. Redistributions of source code must retain the above copyright
  13  *    notice, this list of conditions and the following disclaimer.
  14  * 2. Redistributions in binary form must reproduce the above copyright
  15  *    notice, this list of conditions and the following disclaimer in
  16  *    the documentation and/or other materials provided with the
  17  *    distribution.
  18  * 3. Neither the name of The DragonFly Project nor the names of its
  19  *    contributors may be used to endorse or promote products derived
  20  *    from this software without specific, prior written permission.
  21  *
  22  * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
  23  * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
  24  * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS
  25  * FOR A PARTICULAR PURPOSE ARE DISCLAIMED.  IN NO EVENT SHALL THE
  26  * COPYRIGHT HOLDERS OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT,
  27  * INCIDENTAL, SPECIAL, EXEMPLARY OR CONSEQUENTIAL DAMAGES (INCLUDING,
  28  * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
  29  * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED
  30  * AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
  31  * OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT
  32  * OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
  33  * SUCH DAMAGE.
  34  */
  35
  36 /*
  37  * HAMMER2 IN-MEMORY CACHE OF MEDIA STRUCTURES
  38  *
  39  * This header file contains structures used internally by the HAMMER2
  40  * implementation.  See hammer2_disk.h for on-disk structures.
  41  *
  42  * There is an in-memory representation of all on-media data structure.
  43  * Almost everything is represented by a hammer2_chain structure in-memory.
  44  * Other higher-level structures typically map to chains.
  45  *
  46  * A great deal of data is accessed simply via its buffer cache buffer,
  47  * which is mapped for the duration of the chain's lock.  Hammer2 must
  48  * implement its own buffer cache layer on top of the system layer to
  49  * allow for different threads to lock different sub-block-sized buffers.
  50  *
  51  * When modifications are made to a chain a new filesystem block must be
  52  * allocated.  Multiple modifications do not typically allocate new blocks
  53  * until the current block has been flushed.  Flushes do not block the
  54  * front-end unless the front-end operation crosses the current inode being
  55  * flushed.
  56  *
  57  * The in-memory representation may remain cached (for example in order to
  58  * placemark clustering locks) even after the related data has been
  59  * detached.
  60  */
  61
  62 #ifndef _VFS_HAMMER2_HAMMER2_H_
  63 #define _VFS_HAMMER2_HAMMER2_H_
  64
  65 #ifdef _KERNEL
  66 #include <sys/param.h>
  67 #endif
  68 #include <sys/types.h>
  69 #ifdef _KERNEL
  70 #include <sys/kernel.h>
  71 #endif
  72 #include <sys/conf.h>
  73 #ifdef _KERNEL
  74 #include <sys/systm.h>
  75 #endif
  76 #include <sys/tree.h>
  77 #include <sys/malloc.h>
  78 #include <sys/mount.h>
  79 #include <sys/vnode.h>
  80 #include <sys/proc.h>
  81 #include <sys/mountctl.h>
  82 #include <sys/priv.h>
  83 #include <sys/stat.h>
  84 #include <sys/thread.h>
  85 #include <sys/globaldata.h>
  86 #include <sys/lockf.h>
  87 #include <sys/buf.h>
  88 #include <sys/queue.h>
  89 #include <sys/limits.h>
  90 #include <sys/dmsg.h>
  91 #include <sys/mutex.h>
  92 #ifdef _KERNEL
  93 #include <sys/kern_syscall.h>
  94 #endif
  95
  96 #ifdef _KERNEL
  97 #include <sys/signal2.h>
  98 #include <sys/buf2.h>
  99 #include <sys/mutex2.h>
 100 #include <sys/thread2.h>
 101 #endif
 102
 103 #include "hammer2_xxhash.h"
 104 #include "hammer2_disk.h"
 105 #include "hammer2_mount.h"
 106 #include "hammer2_ioctl.h"
 107
 108 struct hammer2_io;
 109 struct hammer2_chain;
 110 struct hammer2_cluster;
 111 struct hammer2_inode;
 112 struct hammer2_dev;
 113 struct hammer2_pfs;
 114 struct hammer2_span;
 115 struct hammer2_msg;
 116 struct hammer2_thread;
 117 union hammer2_xop;
 118
 119 /*
 120  * Mutex and lock shims.  Hammer2 requires support for asynchronous and
 121  * abortable locks, and both exclusive and shared spinlocks.  Normal
 122  * synchronous non-abortable locks can be substituted for spinlocks.
 123  */
 124 typedef mtx_t                           hammer2_mtx_t;
 125 typedef mtx_link_t                      hammer2_mtx_link_t;
 126 typedef mtx_state_t                     hammer2_mtx_state_t;
 127
 128 typedef struct spinlock                 hammer2_spin_t;
 129
 130 #define hammer2_mtx_ex                  mtx_lock_ex_quick
 131 #define hammer2_mtx_ex_try              mtx_lock_ex_try
 132 #define hammer2_mtx_sh                  mtx_lock_sh_quick
 133 #define hammer2_mtx_sh_again            mtx_lock_sh_again
 134 #define hammer2_mtx_sh_try              mtx_lock_sh_try
 135 #define hammer2_mtx_unlock              mtx_unlock
 136 #define hammer2_mtx_downgrade           mtx_downgrade
 137 #define hammer2_mtx_owned               mtx_owned
 138 #define hammer2_mtx_init                mtx_init
 139 #define hammer2_mtx_temp_release        mtx_lock_temp_release
 140 #define hammer2_mtx_temp_restore        mtx_lock_temp_restore
 141 #define hammer2_mtx_refs                mtx_lockrefs
 142
 143 #define hammer2_spin_init               spin_init
 144 #define hammer2_spin_sh                 spin_lock_shared
 145 #define hammer2_spin_ex                 spin_lock
 146 #define hammer2_spin_unsh               spin_unlock_shared
 147 #define hammer2_spin_unex               spin_unlock
 148
 149 TAILQ_HEAD(hammer2_xop_list, hammer2_xop_head);
 150 TAILQ_HEAD(hammer2_chain_list, hammer2_chain);
 151
 152 typedef struct hammer2_xop_list hammer2_xop_list_t;
 153
 154 #ifdef _KERNEL
 155 /*
 156  * General lock support
 157  */
 158 static __inline
 159 int
 160 hammer2_mtx_upgrade_try(hammer2_mtx_t *mtx)
 161 {
 162         return mtx_upgrade_try(mtx);
 163 }
 164
 165 #endif
 166
 167 /*
 168  * The xid tracks internal transactional updates.
 169  *
 170  * XXX fix-me, really needs to be 64-bits
 171  */
 172 typedef uint32_t hammer2_xid_t;
 173
 174 #define HAMMER2_XID_MIN                 0x00000000U
 175 #define HAMMER2_XID_MAX                 0x7FFFFFFFU
 176
 177 #define HAMMER2_LIMIT_DIRTY_CHAINS      (65536)
 178 #define HAMMER2_LIMIT_DIRTY_INODES      (16384)
 179
 180 /*
 181  * The chain structure tracks a portion of the media topology from the
 182  * root (volume) down.  Chains represent volumes, inodes, indirect blocks,
 183  * data blocks, and freemap nodes and leafs.
 184  *
 185  * The chain structure utilizes a simple singly-homed topology and the
 186  * chain's in-memory topology will move around as the chains do, due mainly
 187  * to renames and indirect block creation.
 188  *
 189  * Block Table Updates
 190  *
 191  *      Block table updates for insertions and updates are delayed until the
 192  *      flush.  This allows us to avoid having to modify the parent chain
 193  *      all the way to the root.
 194  *
 195  *      Block table deletions are performed immediately (modifying the parent
 196  *      in the process) because the flush code uses the chain structure to
 197  *      track delayed updates and the chain will be (likely) gone or moved to
 198  *      another location in the topology after a deletion.
 199  *
 200  *      A prior iteration of the code tried to keep the relationship intact
 201  *      on deletes by doing a delete-duplicate operation on the chain, but
 202  *      it added way too much complexity to the codebase.
 203  *
 204  * Flush Synchronization
 205  *
 206  *      The flush code must flush modified chains bottom-up.  Because chain
 207  *      structures can shift around and are NOT topologically stable,
 208  *      modified chains are independently indexed for the flush.  As the flush
 209  *      runs it modifies (or further modifies) and updates the parents,
 210  *      propagating the flush all the way to the volume root.
 211  *
 212  *      Modifying front-end operations can occur during a flush but will block
 213  *      in two cases: (1) when the front-end tries to operate on the inode
 214  *      currently in the midst of being flushed and (2) if the front-end
 215  *      crosses an inode currently being flushed (such as during a rename).
 216  *      So, for example, if you rename directory "x" to "a/b/c/d/e/f/g/x" and
 217  *      the flusher is currently working on "a/b/c", the rename will block
 218  *      temporarily in order to ensure that "x" exists in one place or the
 219  *      other.
 220  *
 221  *      Meta-data statistics are updated by the flusher.  The front-end will
 222  *      make estimates but meta-data must be fully synchronized only during a
 223  *      flush in order to ensure that it remains correct across a crash.
 224  *
 225  *      Multiple flush synchronizations can theoretically be in-flight at the
 226  *      same time but the implementation is not coded to handle the case and
 227  *      currently serializes them.
 228  *
 229  * Snapshots:
 230  *
 231  *      Snapshots currently require the subdirectory tree being snapshotted
 232  *      to be flushed.  The snapshot then creates a new super-root inode which
 233  *      copies the flushed blockdata of the directory or file that was
 234  *      snapshotted.
 235  *
 236  * RBTREE NOTES:
 237  *
 238  *      - Note that the radix tree runs in powers of 2 only so sub-trees
 239  *        cannot straddle edges.
 240  */
 241 RB_HEAD(hammer2_chain_tree, hammer2_chain);
 242 TAILQ_HEAD(h2_flush_list, hammer2_chain);
 243 TAILQ_HEAD(h2_core_list, hammer2_chain);
 244
 245 #define CHAIN_CORE_DELETE_BMAP_ENTRIES  \
 246         (HAMMER2_PBUFSIZE / sizeof(hammer2_blockref_t) / sizeof(uint32_t))
 247
 248 struct hammer2_reptrack {
 249         hammer2_spin_t  spin;
 250         struct hammer2_reptrack *next;
 251         struct hammer2_chain    *chain;
 252 };
 253
 254 /*
 255  * Core topology for chain (embedded in chain).  Protected by a spinlock.
 256  */
 257 struct hammer2_chain_core {
 258         hammer2_spin_t  spin;
 259         struct hammer2_reptrack *reptrack;
 260         struct hammer2_chain_tree rbtree; /* sub-chains */
 261         int             live_zero;      /* blockref array opt */
 262         u_int           live_count;     /* live (not deleted) chains in tree */
 263         u_int           chain_count;    /* live + deleted chains under core */
 264         int             generation;     /* generation number (inserts only) */
 265 };
 266
 267 typedef struct hammer2_chain_core hammer2_chain_core_t;
 268
 269 RB_HEAD(hammer2_io_tree, hammer2_io);
 270
 271 /*
 272  * DIO - Management structure wrapping system buffer cache.
 273  *
 274  * HAMMER2 uses an I/O abstraction that allows it to cache and manipulate
 275  * fixed-sized filesystem buffers frontend by variable-sized hammer2_chain
 276  * structures.
 277  */
 278 struct hammer2_io {
 279         RB_ENTRY(hammer2_io) rbnode;    /* indexed by device offset */
 280         struct hammer2_dev *hmp;
 281         struct buf      *bp;
 282         off_t           pbase;
 283         uint64_t        refs;
 284         int             psize;
 285         int             act;            /* activity */
 286         int             btype;          /* approximate BREF_TYPE_* */
 287         int             ticks;
 288         int             error;
 289         int             unused01;
 290         uint64_t        dedup_valid;    /* valid for dedup operation */
 291         uint64_t        dedup_alloc;    /* allocated / de-dupable */
 292 };
 293
 294 typedef struct hammer2_io hammer2_io_t;
 295
 296 #define HAMMER2_DIO_INPROG      0x8000000000000000LLU   /* bio in progress */
 297 #define HAMMER2_DIO_GOOD        0x4000000000000000LLU   /* dio->bp is stable */
 298 #define HAMMER2_DIO_WAITING     0x2000000000000000LLU   /* wait on INPROG */
 299 #define HAMMER2_DIO_DIRTY       0x1000000000000000LLU   /* flush last drop */
 300
 301 #define HAMMER2_DIO_MASK        0x00FFFFFFFFFFFFFFLLU
 302
 303 /*
 304  * Primary chain structure keeps track of the topology in-memory.
 305  */
 306 struct hammer2_chain {
 307         hammer2_mtx_t           lock;
 308         hammer2_chain_core_t    core;
 309         RB_ENTRY(hammer2_chain) rbnode;         /* live chain(s) */
 310         hammer2_blockref_t      bref;
 311         struct hammer2_chain    *parent;
 312         struct hammer2_dev      *hmp;
 313         struct hammer2_pfs      *pmp;           /* A PFS or super-root (spmp) */
 314
 315         hammer2_io_t    *dio;                   /* physical data buffer */
 316         u_int           bytes;                  /* physical data size */
 317         u_int           flags;
 318         u_int           refs;
 319         u_int           lockcnt;
 320         int             error;                  /* on-lock data error state */
 321         int             cache_index;            /* heur speeds up lookup */
 322
 323         hammer2_media_data_t *data;             /* data pointer shortcut */
 324         TAILQ_ENTRY(hammer2_chain) flush_node;  /* flush list */
 325         TAILQ_ENTRY(hammer2_chain) lru_node;    /* 0-refs LRU */
 326 };
 327
 328 typedef struct hammer2_chain hammer2_chain_t;
 329
 330 int hammer2_chain_cmp(hammer2_chain_t *chain1, hammer2_chain_t *chain2);
 331 RB_PROTOTYPE(hammer2_chain_tree, hammer2_chain, rbnode, hammer2_chain_cmp);
 332
 333 /*
 334  * Special notes on flags:
 335  *
 336  * INITIAL      - This flag allows a chain to be created and for storage to
 337  *                be allocated without having to immediately instantiate the
 338  *                related buffer.  The data is assumed to be all-zeros.  It
 339  *                is primarily used for indirect blocks.
 340  *
 341  * MODIFIED     - The chain's media data has been modified.  Prevents chain
 342  *                free on lastdrop if still in the topology.
 343  *
 344  * UPDATE       - Chain might not be modified but parent blocktable needs
 345  *                an update.  Prevents chain free on lastdrop if still in
 346  *                the topology.
 347  *
 348  * FICTITIOUS   - Faked chain as a placeholder for an error condition.  This
 349  *                chain is unsuitable for I/O.
 350  *
 351  * BMAPPED      - Indicates that the chain is present in the parent blockmap.
 352  *
 353  * BMAPUPD      - Indicates that the chain is present but needs to be updated
 354  *                in the parent blockmap.
 355  */
 356 #define HAMMER2_CHAIN_MODIFIED          0x00000001      /* dirty chain data */
 357 #define HAMMER2_CHAIN_ALLOCATED         0x00000002      /* kmalloc'd chain */
 358 #define HAMMER2_CHAIN_DESTROY           0x00000004
 359 #define HAMMER2_CHAIN_DEDUPABLE         0x00000008      /* registered w/dedup */
 360 #define HAMMER2_CHAIN_DELETED           0x00000010      /* deleted chain */
 361 #define HAMMER2_CHAIN_INITIAL           0x00000020      /* initial create */
 362 #define HAMMER2_CHAIN_UPDATE            0x00000040      /* need parent update */
 363 #define HAMMER2_CHAIN_DEFERRED          0x00000080      /* flush depth defer */
 364 #define HAMMER2_CHAIN_TESTEDGOOD        0x00000100      /* crc tested good */
 365 #define HAMMER2_CHAIN_ONFLUSH           0x00000200      /* on a flush list */
 366 #define HAMMER2_CHAIN_FICTITIOUS        0x00000400      /* unsuitable for I/O */
 367 #define HAMMER2_CHAIN_VOLUMESYNC        0x00000800      /* needs volume sync */
 368 #define HAMMER2_CHAIN_DELAYED           0x00001000      /* delayed flush */
 369 #define HAMMER2_CHAIN_COUNTEDBREFS      0x00002000      /* block table stats */
 370 #define HAMMER2_CHAIN_ONRBTREE          0x00004000      /* on parent RB tree */
 371 #define HAMMER2_CHAIN_ONLRU             0x00008000      /* on LRU list */
 372 #define HAMMER2_CHAIN_EMBEDDED          0x00010000      /* embedded data */
 373 #define HAMMER2_CHAIN_RELEASE           0x00020000      /* don't keep around */
 374 #define HAMMER2_CHAIN_BMAPPED           0x00040000      /* present in blkmap */
 375 #define HAMMER2_CHAIN_BMAPUPD           0x00080000      /* +needs updating */
 376 #define HAMMER2_CHAIN_IOINPROG          0x00100000      /* I/O interlock */
 377 #define HAMMER2_CHAIN_IOSIGNAL          0x00200000      /* I/O interlock */
 378 #define HAMMER2_CHAIN_PFSBOUNDARY       0x00400000      /* super->pfs inode */
 379 #define HAMMER2_CHAIN_HINT_LEAF_COUNT   0x00800000      /* redo leaf count */
 380 #define HAMMER2_CHAIN_LRUHINT           0x01000000      /* was reused */
 381
 382 #define HAMMER2_CHAIN_FLUSH_MASK        (HAMMER2_CHAIN_MODIFIED |       \
 383                                          HAMMER2_CHAIN_UPDATE |         \
 384                                          HAMMER2_CHAIN_ONFLUSH |        \
 385                                          HAMMER2_CHAIN_DESTROY)
 386
 387 /*
 388  * Hammer2 error codes, used by chain->error and cluster->error.  The error
 389  * code is typically set on-lock unless no I/O was requested, and set on
 390  * I/O otherwise.  If set for a cluster it generally means that the cluster
 391  * code could not find a valid copy to present.
 392  *
 393  * All H2 error codes are flags and can be accumulated by ORing them
 394  * together.
 395  *
 396  * IO           - An I/O error occurred
 397  * CHECK        - I/O succeeded but did not match the check code
 398  * INCOMPLETE   - A cluster is not complete enough to use, or
 399  *                a chain cannot be loaded because its parent has an error.
 400  *
 401  * NOTE: API allows callers to check zero/non-zero to determine if an error
 402  *       condition exists.
 403  *
 404  * NOTE: Chain's data field is usually NULL on an IO error but not necessarily
 405  *       NULL on other errors.  Check chain->error, not chain->data.
 406  */
 407 #define HAMMER2_ERROR_NONE              0       /* no error (must be 0) */
 408 #define HAMMER2_ERROR_EIO               0x00000001      /* device I/O error */
 409 #define HAMMER2_ERROR_CHECK             0x00000002      /* check code error */
 410 #define HAMMER2_ERROR_INCOMPLETE        0x00000004      /* incomplete cluster */
 411 #define HAMMER2_ERROR_DEPTH             0x00000008      /* tmp depth limit */
 412 #define HAMMER2_ERROR_BADBREF           0x00000010      /* illegal bref */
 413 #define HAMMER2_ERROR_ENOSPC            0x00000020      /* allocation failure */
 414 #define HAMMER2_ERROR_ENOENT            0x00000040      /* entry not found */
 415 #define HAMMER2_ERROR_ENOTEMPTY         0x00000080      /* dir not empty */
 416 #define HAMMER2_ERROR_EAGAIN            0x00000100      /* retry */
 417 #define HAMMER2_ERROR_ENOTDIR           0x00000200      /* not directory */
 418 #define HAMMER2_ERROR_EISDIR            0x00000400      /* is directory */
 419 #define HAMMER2_ERROR_EINPROGRESS       0x00000800      /* already running */
 420 #define HAMMER2_ERROR_ABORTED           0x00001000      /* aborted operation */
 421 #define HAMMER2_ERROR_EOF               0x00002000      /* end of scan */
 422 #define HAMMER2_ERROR_EINVAL            0x00004000      /* catch-all */
 423 #define HAMMER2_ERROR_EEXIST            0x00008000      /* entry exists */
 424 #define HAMMER2_ERROR_EDEADLK           0x00010000
 425 #define HAMMER2_ERROR_ESRCH             0x00020000
 426 #define HAMMER2_ERROR_ETIMEDOUT         0x00040000
 427
 428 /*
 429  * Flags passed to hammer2_chain_lookup() and hammer2_chain_next()
 430  *
 431  * NOTES:
 432  *      NODATA      - Asks that the chain->data not be resolved in order
 433  *                    to avoid I/O.
 434  *
 435  *      NODIRECT    - Prevents a lookup of offset 0 in an inode from returning
 436  *                    the inode itself if the inode is in DIRECTDATA mode
 437  *                    (i.e. file is <= 512 bytes).  Used by the synchronization
 438  *                    code to prevent confusion.
 439  *
 440  *      SHARED      - The input chain is expected to be locked shared,
 441  *                    and the output chain is locked shared.
 442  *
 443  *      MATCHIND    - Allows an indirect block / freemap node to be returned
 444  *                    when the passed key range matches the radix.  Remember
 445  *                    that key_end is inclusive (e.g. {0x000,0xFFF},
 446  *                    not {0x000,0x1000}).
 447  *
 448  *                    (Cannot be used for remote or cluster ops).
 449  *
 450  *      ALLNODES    - Allows NULL focus.
 451  *
 452  *      ALWAYS      - Always resolve the data.  If ALWAYS and NODATA are both
 453  *                    missing, bulk file data is not resolved but inodes and
 454  *                    other meta-data will.
 455  */
 456 #define HAMMER2_LOOKUP_UNUSED0001       0x00000001
 457 #define HAMMER2_LOOKUP_NODATA           0x00000002      /* data left NULL */
 458 #define HAMMER2_LOOKUP_NODIRECT         0x00000004      /* no offset=0 DD */
 459 #define HAMMER2_LOOKUP_SHARED           0x00000100
 460 #define HAMMER2_LOOKUP_MATCHIND         0x00000200      /* return all chains */
 461 #define HAMMER2_LOOKUP_ALLNODES         0x00000400      /* allow NULL focus */
 462 #define HAMMER2_LOOKUP_ALWAYS           0x00000800      /* resolve data */
 463 #define HAMMER2_LOOKUP_UNUSED1000       0x00001000
 464
 465 /*
 466  * Flags passed to hammer2_chain_modify() and hammer2_chain_resize()
 467  *
 468  * NOTE: OPTDATA allows us to avoid instantiating buffers for INDIRECT
 469  *       blocks in the INITIAL-create state.
 470  */
 471 #define HAMMER2_MODIFY_OPTDATA          0x00000002      /* data can be NULL */
 472 #define HAMMER2_MODIFY_NO_MODIFY_TID    0x00000004
 473 #define HAMMER2_MODIFY_UNUSED0008       0x00000008
 474
 475 /*
 476  * Flags passed to hammer2_chain_lock()
 477  *
 478  * NOTE: RDONLY is set to optimize cluster operations when *no* modifications
 479  *       will be made to either the cluster being locked or any underlying
 480  *       cluster.  It allows the cluster to lock and access data for a subset
 481  *       of available nodes instead of all available nodes.
 482  *
 483  * NOTE: NONBLOCK is only used for hammer2_chain_repparent() and getparent(),
 484  *       other functions (e.g. hammer2_chain_lookup(), etc) can't handle its
 485  *       operation.
 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_LOCKAGAIN       0x20    /* another shared lock */
 494 #define HAMMER2_RESOLVE_RDONLY          0x40    /* higher level op flag */
 495 #define HAMMER2_RESOLVE_NONBLOCK        0x80    /* non-blocking */
 496
 497 /*
 498  * Flags passed to hammer2_chain_delete()
 499  */
 500 #define HAMMER2_DELETE_PERMANENT        0x0001
 501
 502 /*
 503  * Flags passed to hammer2_chain_insert() or hammer2_chain_rename()
 504  * or hammer2_chain_create().
 505  */
 506 #define HAMMER2_INSERT_PFSROOT          0x0004
 507 #define HAMMER2_INSERT_SAMEPARENT       0x0008
 508
 509 /*
 510  * Flags passed to hammer2_chain_delete_duplicate()
 511  */
 512 #define HAMMER2_DELDUP_RECORE           0x0001
 513
 514 /*
 515  * Cluster different types of storage together for allocations
 516  */
 517 #define HAMMER2_FREECACHE_INODE         0
 518 #define HAMMER2_FREECACHE_INDIR         1
 519 #define HAMMER2_FREECACHE_DATA          2
 520 #define HAMMER2_FREECACHE_UNUSED3       3
 521 #define HAMMER2_FREECACHE_TYPES         4
 522
 523 /*
 524  * hammer2_freemap_alloc() block preference
 525  */
 526 #define HAMMER2_OFF_NOPREF              ((hammer2_off_t)-1)
 527
 528 /*
 529  * BMAP read-ahead maximum parameters
 530  */
 531 #define HAMMER2_BMAP_COUNT              16      /* max bmap read-ahead */
 532 #define HAMMER2_BMAP_BYTES              (HAMMER2_PBUFSIZE * HAMMER2_BMAP_COUNT)
 533
 534 /*
 535  * hammer2_freemap_adjust()
 536  */
 537 #define HAMMER2_FREEMAP_DORECOVER       1
 538 #define HAMMER2_FREEMAP_DOMAYFREE       2
 539 #define HAMMER2_FREEMAP_DOREALFREE      3
 540
 541 /*
 542  * HAMMER2 cluster - A set of chains representing the same entity.
 543  *
 544  * hammer2_cluster typically represents a temporary set of representitive
 545  * chains.  The one exception is that a hammer2_cluster is embedded in
 546  * hammer2_inode.  This embedded cluster is ONLY used to track the
 547  * representitive chains and cannot be directly locked.
 548  *
 549  * A cluster is usually temporary (and thus per-thread) for locking purposes,
 550  * allowing us to embed the asynchronous storage required for cluster
 551  * operations in the cluster itself and adjust the state and status without
 552  * having to worry too much about SMP issues.
 553  *
 554  * The exception is the cluster embedded in the hammer2_inode structure.
 555  * This is used to cache the cluster state on an inode-by-inode basis.
 556  * Individual hammer2_chain structures not incorporated into clusters might
 557  * also stick around to cache miscellanious elements.
 558  *
 559  * Because the cluster is a 'working copy' and is usually subject to cluster
 560  * quorum rules, it is quite possible for us to end up with an insufficient
 561  * number of live chains to execute an operation.  If an insufficient number
 562  * of chains remain in a working copy, the operation may have to be
 563  * downgraded, retried, stall until the requisit number of chains are
 564  * available, or possibly even error out depending on the mount type.
 565  *
 566  * A cluster's focus is set when it is locked.  The focus can only be set
 567  * to a chain still part of the synchronized set.
 568  */
 569 #define HAMMER2_XOPFIFO         16
 570 #define HAMMER2_XOPFIFO_MASK    (HAMMER2_XOPFIFO - 1)
 571 #define HAMMER2_XOPGROUPS       32
 572 #define HAMMER2_XOPGROUPS_MASK  (HAMMER2_XOPGROUPS - 1)
 573
 574 #define HAMMER2_MAXCLUSTER      8
 575 #define HAMMER2_XOPMASK_CLUSTER (uint64_t)((1LLU << HAMMER2_MAXCLUSTER) - 1)
 576 #define HAMMER2_XOPMASK_VOP     (uint64_t)0x0000000080000000LLU
 577 #define HAMMER2_XOPMASK_FIFOW   (uint64_t)0x0000000040000000LLU
 578 #define HAMMER2_XOPMASK_WAIT    (uint64_t)0x0000000020000000LLU
 579 #define HAMMER2_XOPMASK_FEED    (uint64_t)0x0000000100000000LLU
 580
 581 #define HAMMER2_XOPMASK_ALLDONE (HAMMER2_XOPMASK_VOP | HAMMER2_XOPMASK_CLUSTER)
 582
 583 #define HAMMER2_SPECTHREADS     1       /* sync */
 584
 585 struct hammer2_cluster_item {
 586         hammer2_chain_t         *chain;
 587         int                     error;
 588         uint32_t                flags;
 589 };
 590
 591 typedef struct hammer2_cluster_item hammer2_cluster_item_t;
 592
 593 /*
 594  * INVALID      - Invalid for focus, i.e. not part of synchronized set.
 595  *                Once set, this bit is sticky across operations.
 596  *
 597  * FEMOD        - Indicates that front-end modifying operations can
 598  *                mess with this entry and MODSYNC will copy also
 599  *                effect it.
 600  */
 601 #define HAMMER2_CITEM_INVALID   0x00000001
 602 #define HAMMER2_CITEM_FEMOD     0x00000002
 603 #define HAMMER2_CITEM_NULL      0x00000004
 604
 605 struct hammer2_cluster {
 606         int                     refs;           /* track for deallocation */
 607         int                     ddflag;
 608         struct hammer2_pfs      *pmp;
 609         uint32_t                flags;
 610         int                     nchains;
 611         int                     error;          /* error code valid on lock */
 612         int                     focus_index;
 613         hammer2_chain_t         *focus;         /* current focus (or mod) */
 614         hammer2_cluster_item_t  array[HAMMER2_MAXCLUSTER];
 615 };
 616
 617 typedef struct hammer2_cluster  hammer2_cluster_t;
 618
 619 /*
 620  * WRHARD       - Hard mounts can write fully synchronized
 621  * RDHARD       - Hard mounts can read fully synchronized
 622  * UNHARD       - Unsynchronized masters present
 623  * NOHARD       - No masters visible
 624  * WRSOFT       - Soft mounts can write to at least the SOFT_MASTER
 625  * RDSOFT       - Soft mounts can read from at least a SOFT_SLAVE
 626  * UNSOFT       - Unsynchronized slaves present
 627  * NOSOFT       - No slaves visible
 628  * RDSLAVE      - slaves are accessible (possibly unsynchronized or remote).
 629  * MSYNCED      - All masters are fully synchronized
 630  * SSYNCED      - All known local slaves are fully synchronized to masters
 631  *
 632  * All available masters are always incorporated.  All PFSs belonging to a
 633  * cluster (master, slave, copy, whatever) always try to synchronize the
 634  * total number of known masters in the PFSs root inode.
 635  *
 636  * A cluster might have access to many slaves, copies, or caches, but we
 637  * have a limited number of cluster slots.  Any such elements which are
 638  * directly mounted from block device(s) will always be incorporated.   Note
 639  * that SSYNCED only applies to such elements which are directly mounted,
 640  * not to any remote slaves, copies, or caches that could be available.  These
 641  * bits are used to monitor and drive our synchronization threads.
 642  *
 643  * When asking the question 'is any data accessible at all', then a simple
 644  * test against (RDHARD|RDSOFT|RDSLAVE) gives you the answer.  If any of
 645  * these bits are set the object can be read with certain caveats:
 646  * RDHARD - no caveats.  RDSOFT - authoritative but might not be synchronized.
 647  * and RDSLAVE - not authoritative, has some data but it could be old or
 648  * incomplete.
 649  *
 650  * When both soft and hard mounts are available, data will be read and written
 651  * via the soft mount only.  But all might be in the cluster because
 652  * background synchronization threads still need to do their work.
 653  */
 654 #define HAMMER2_CLUSTER_INODE   0x00000001      /* embedded in inode struct */
 655 #define HAMMER2_CLUSTER_UNUSED2 0x00000002
 656 #define HAMMER2_CLUSTER_LOCKED  0x00000004      /* cluster lks not recursive */
 657 #define HAMMER2_CLUSTER_WRHARD  0x00000100      /* hard-mount can write */
 658 #define HAMMER2_CLUSTER_RDHARD  0x00000200      /* hard-mount can read */
 659 #define HAMMER2_CLUSTER_UNHARD  0x00000400      /* unsynchronized masters */
 660 #define HAMMER2_CLUSTER_NOHARD  0x00000800      /* no masters visible */
 661 #define HAMMER2_CLUSTER_WRSOFT  0x00001000      /* soft-mount can write */
 662 #define HAMMER2_CLUSTER_RDSOFT  0x00002000      /* soft-mount can read */
 663 #define HAMMER2_CLUSTER_UNSOFT  0x00004000      /* unsynchronized slaves */
 664 #define HAMMER2_CLUSTER_NOSOFT  0x00008000      /* no slaves visible */
 665 #define HAMMER2_CLUSTER_MSYNCED 0x00010000      /* all masters synchronized */
 666 #define HAMMER2_CLUSTER_SSYNCED 0x00020000      /* known slaves synchronized */
 667
 668 #define HAMMER2_CLUSTER_ANYDATA ( HAMMER2_CLUSTER_RDHARD |      \
 669                                   HAMMER2_CLUSTER_RDSOFT |      \
 670                                   HAMMER2_CLUSTER_RDSLAVE)
 671
 672 #define HAMMER2_CLUSTER_RDOK    ( HAMMER2_CLUSTER_RDHARD |      \
 673                                   HAMMER2_CLUSTER_RDSOFT)
 674
 675 #define HAMMER2_CLUSTER_WROK    ( HAMMER2_CLUSTER_WRHARD |      \
 676                                   HAMMER2_CLUSTER_WRSOFT)
 677
 678 #define HAMMER2_CLUSTER_ZFLAGS  ( HAMMER2_CLUSTER_WRHARD |      \
 679                                   HAMMER2_CLUSTER_RDHARD |      \
 680                                   HAMMER2_CLUSTER_WRSOFT |      \
 681                                   HAMMER2_CLUSTER_RDSOFT |      \
 682                                   HAMMER2_CLUSTER_MSYNCED |     \
 683                                   HAMMER2_CLUSTER_SSYNCED)
 684
 685 /*
 686  * Helper functions (cluster must be locked for flags to be valid).
 687  */
 688 static __inline
 689 int
 690 hammer2_cluster_rdok(hammer2_cluster_t *cluster)
 691 {
 692         return (cluster->flags & HAMMER2_CLUSTER_RDOK);
 693 }
 694
 695 static __inline
 696 int
 697 hammer2_cluster_wrok(hammer2_cluster_t *cluster)
 698 {
 699         return (cluster->flags & HAMMER2_CLUSTER_WROK);
 700 }
 701
 702 RB_HEAD(hammer2_inode_tree, hammer2_inode);
 703
 704 /*
 705  * A hammer2 inode.
 706  *
 707  * NOTE: The inode-embedded cluster is never used directly for I/O (since
 708  *       it may be shared).  Instead it will be replicated-in and synchronized
 709  *       back out if changed.
 710  */
 711 struct hammer2_inode {
 712         RB_ENTRY(hammer2_inode) rbnode;         /* inumber lookup (HL) */
 713         hammer2_mtx_t           lock;           /* inode lock */
 714         hammer2_mtx_t           truncate_lock;  /* prevent truncates */
 715         struct hammer2_pfs      *pmp;           /* PFS mount */
 716         struct vnode            *vp;
 717         struct spinlock         cluster_spin;   /* update cluster */
 718         hammer2_cluster_t       cluster;
 719         struct lockf            advlock;
 720         u_int                   flags;
 721         u_int                   refs;           /* +vpref, +flushref */
 722         uint8_t                 comp_heuristic;
 723         hammer2_inode_meta_t    meta;           /* copy of meta-data */
 724         hammer2_off_t           osize;
 725 };
 726
 727 typedef struct hammer2_inode hammer2_inode_t;
 728
 729 /*
 730  * MODIFIED     - Inode is in a modified state, ip->meta may have changes.
 731  * RESIZED      - Inode truncated (any) or inode extended beyond
 732  *                EMBEDDED_BYTES.
 733  */
 734 #define HAMMER2_INODE_MODIFIED          0x0001
 735 #define HAMMER2_INODE_SROOT             0x0002  /* kmalloc special case */
 736 #define HAMMER2_INODE_RENAME_INPROG     0x0004
 737 #define HAMMER2_INODE_ONRBTREE          0x0008
 738 #define HAMMER2_INODE_RESIZED           0x0010  /* requires inode_fsync */
 739 #define HAMMER2_INODE_ISDELETED         0x0020  /* deleted */
 740 #define HAMMER2_INODE_ISUNLINKED        0x0040
 741 #define HAMMER2_INODE_METAGOOD          0x0080  /* inode meta-data good */
 742 #define HAMMER2_INODE_ONSIDEQ           0x0100  /* on side processing queue */
 743 #define HAMMER2_INODE_NOSIDEQ           0x0200  /* disable sideq operation */
 744 #define HAMMER2_INODE_DIRTYDATA         0x0400  /* interlocks inode flush */
 745
 746 int hammer2_inode_cmp(hammer2_inode_t *ip1, hammer2_inode_t *ip2);
 747 RB_PROTOTYPE2(hammer2_inode_tree, hammer2_inode, rbnode, hammer2_inode_cmp,
 748                 hammer2_tid_t);
 749
 750 /*
 751  * inode-unlink side-structure
 752  */
 753 struct hammer2_inode_sideq {
 754         TAILQ_ENTRY(hammer2_inode_sideq) entry;
 755         hammer2_inode_t *ip;
 756 };
 757 TAILQ_HEAD(h2_sideq_list, hammer2_inode_sideq);
 758
 759 typedef struct hammer2_inode_sideq hammer2_inode_sideq_t;
 760
 761 /*
 762  * Transaction management sub-structure under hammer2_pfs
 763  */
 764 struct hammer2_trans {
 765         uint32_t                flags;
 766         uint32_t                sync_wait;
 767         int                     fticks;                 /* FPENDING start */
 768 };
 769
 770 typedef struct hammer2_trans hammer2_trans_t;
 771
 772 #define HAMMER2_TRANS_ISFLUSH           0x80000000      /* flush code */
 773 #define HAMMER2_TRANS_BUFCACHE          0x40000000      /* bio strategy */
 774 #define HAMMER2_TRANS_UNUSED20          0x20000000
 775 #define HAMMER2_TRANS_FPENDING          0x10000000      /* flush pending */
 776 #define HAMMER2_TRANS_WAITING           0x08000000      /* someone waiting */
 777 #define HAMMER2_TRANS_MASK              0x00FFFFFF      /* count mask */
 778
 779 #define HAMMER2_FREEMAP_HEUR_NRADIX     4       /* pwr 2 PBUFRADIX-MINIORADIX */
 780 #define HAMMER2_FREEMAP_HEUR_TYPES      8
 781 #define HAMMER2_FREEMAP_HEUR_SIZE       (HAMMER2_FREEMAP_HEUR_NRADIX * \
 782                                          HAMMER2_FREEMAP_HEUR_TYPES)
 783
 784 #define HAMMER2_DEDUP_HEUR_SIZE         (65536 * 4)
 785 #define HAMMER2_DEDUP_HEUR_MASK         (HAMMER2_DEDUP_HEUR_SIZE - 1)
 786
 787 #define HAMMER2_FLUSH_TOP               0x0001
 788 #define HAMMER2_FLUSH_ALL               0x0002
 789 #define HAMMER2_FLUSH_INODE_STOP        0x0004  /* stop at sub-inode */
 790
 791
 792 /*
 793  * Hammer2 support thread element.
 794  *
 795  * Potentially many support threads can hang off of hammer2, primarily
 796  * off the hammer2_pfs structure.  Typically:
 797  *
 798  * td x Nodes                   A synchronization thread for each node.
 799  * td x Nodes x workers         Worker threads for frontend operations.
 800  * td x 1                       Bioq thread for logical buffer writes.
 801  *
 802  * In addition, the synchronization thread(s) associated with the
 803  * super-root PFS (spmp) for a node is responsible for automatic bulkfree
 804  * and dedup scans.
 805  */
 806 struct hammer2_thread {
 807         struct hammer2_pfs *pmp;
 808         struct hammer2_dev *hmp;
 809         hammer2_xop_list_t xopq;
 810         thread_t        td;
 811         uint32_t        flags;
 812         int             depth;
 813         int             clindex;        /* cluster element index */
 814         int             repidx;
 815         char            *scratch;       /* MAXPHYS */
 816 };
 817
 818 typedef struct hammer2_thread hammer2_thread_t;
 819
 820 #define HAMMER2_THREAD_UNMOUNTING       0x0001  /* unmount request */
 821 #define HAMMER2_THREAD_DEV              0x0002  /* related to dev, not pfs */
 822 #define HAMMER2_THREAD_WAITING          0x0004  /* thread in idle tsleep */
 823 #define HAMMER2_THREAD_REMASTER         0x0008  /* remaster request */
 824 #define HAMMER2_THREAD_STOP             0x0010  /* exit request */
 825 #define HAMMER2_THREAD_FREEZE           0x0020  /* force idle */
 826 #define HAMMER2_THREAD_FROZEN           0x0040  /* thread is frozen */
 827 #define HAMMER2_THREAD_XOPQ             0x0080  /* work pending */
 828 #define HAMMER2_THREAD_STOPPED          0x0100  /* thread has stopped */
 829 #define HAMMER2_THREAD_UNFREEZE         0x0200
 830
 831 #define HAMMER2_THREAD_WAKEUP_MASK      (HAMMER2_THREAD_UNMOUNTING |    \
 832                                          HAMMER2_THREAD_REMASTER |      \
 833                                          HAMMER2_THREAD_STOP |          \
 834                                          HAMMER2_THREAD_FREEZE |        \
 835                                          HAMMER2_THREAD_XOPQ)
 836
 837 /*
 838  * Support structure for dedup heuristic.
 839  */
 840 struct hammer2_dedup {
 841         hammer2_off_t   data_off;
 842         uint64_t        data_crc;
 843         uint32_t        ticks;
 844         uint32_t        unused03;
 845 };
 846
 847 typedef struct hammer2_dedup hammer2_dedup_t;
 848
 849 /*
 850  * hammer2_xop - container for VOP/XOP operation (allocated, not on stack).
 851  *
 852  * This structure is used to distribute a VOP operation across multiple
 853  * nodes.  It provides a rendezvous for concurrent node execution and
 854  * can be detached from the frontend operation to allow the frontend to
 855  * return early.
 856  *
 857  * This structure also sequences operations on up to three inodes.
 858  */
 859 typedef void (*hammer2_xop_func_t)(hammer2_thread_t *thr,
 860                                    union hammer2_xop *xop);
 861
 862 struct hammer2_xop_fifo {
 863         TAILQ_ENTRY(hammer2_xop_head) entry;
 864         hammer2_chain_t         *array[HAMMER2_XOPFIFO];
 865         int                     errors[HAMMER2_XOPFIFO];
 866         int                     ri;
 867         int                     wi;
 868         int                     flags;
 869         hammer2_thread_t        *thr;
 870 };
 871
 872 typedef struct hammer2_xop_fifo hammer2_xop_fifo_t;
 873
 874 #define HAMMER2_XOP_FIFO_RUN    0x0001
 875 #define HAMMER2_XOP_FIFO_STALL  0x0002
 876
 877 struct hammer2_xop_head {
 878         hammer2_xop_func_t      func;
 879         hammer2_tid_t           mtid;
 880         struct hammer2_inode    *ip1;
 881         struct hammer2_inode    *ip2;
 882         struct hammer2_inode    *ip3;
 883         uint64_t                run_mask;
 884         uint64_t                chk_mask;
 885         int                     flags;
 886         int                     state;
 887         int                     error;
 888         hammer2_key_t           collect_key;
 889         char                    *name1;
 890         size_t                  name1_len;
 891         char                    *name2;
 892         size_t                  name2_len;
 893         hammer2_xop_fifo_t      collect[HAMMER2_MAXCLUSTER];
 894         hammer2_cluster_t       cluster;        /* help collections */
 895 };
 896
 897 typedef struct hammer2_xop_head hammer2_xop_head_t;
 898
 899 struct hammer2_xop_ipcluster {
 900         hammer2_xop_head_t      head;
 901 };
 902
 903 struct hammer2_xop_strategy {
 904         hammer2_xop_head_t      head;
 905         hammer2_key_t           lbase;
 906         int                     finished;
 907         hammer2_mtx_t           lock;
 908         struct bio              *bio;
 909 };
 910
 911 struct hammer2_xop_readdir {
 912         hammer2_xop_head_t      head;
 913         hammer2_key_t           lkey;
 914 };
 915
 916 struct hammer2_xop_nresolve {
 917         hammer2_xop_head_t      head;
 918         hammer2_key_t           lhc;    /* if name is NULL used lhc */
 919 };
 920
 921 struct hammer2_xop_unlink {
 922         hammer2_xop_head_t      head;
 923         int                     isdir;
 924         int                     dopermanent;
 925 };
 926
 927 #define H2DOPERM_PERMANENT      0x01
 928 #define H2DOPERM_FORCE          0x02
 929 #define H2DOPERM_IGNINO         0x04
 930
 931 struct hammer2_xop_nrename {
 932         hammer2_xop_head_t      head;
 933         hammer2_tid_t           lhc;
 934         int                     ip_key;
 935 };
 936
 937 struct hammer2_xop_scanlhc {
 938         hammer2_xop_head_t      head;
 939         hammer2_key_t           lhc;
 940 };
 941
 942 struct hammer2_xop_scanall {
 943         hammer2_xop_head_t      head;
 944         hammer2_key_t           key_beg;        /* inclusive */
 945         hammer2_key_t           key_end;        /* inclusive */
 946         int                     resolve_flags;
 947         int                     lookup_flags;
 948 };
 949
 950 struct hammer2_xop_lookup {
 951         hammer2_xop_head_t      head;
 952         hammer2_key_t           lhc;
 953 };
 954
 955 struct hammer2_xop_mkdirent {
 956         hammer2_xop_head_t      head;
 957         hammer2_dirent_head_t   dirent;
 958         hammer2_key_t           lhc;
 959 };
 960
 961 struct hammer2_xop_create {
 962         hammer2_xop_head_t      head;
 963         hammer2_inode_meta_t    meta;           /* initial metadata */
 964         hammer2_key_t           lhc;
 965         int                     flags;
 966 };
 967
 968 struct hammer2_xop_destroy {
 969         hammer2_xop_head_t      head;
 970 };
 971
 972 struct hammer2_xop_fsync {
 973         hammer2_xop_head_t      head;
 974         hammer2_inode_meta_t    meta;
 975         hammer2_off_t           osize;
 976         u_int                   ipflags;
 977         int                     clear_directdata;
 978 };
 979
 980 struct hammer2_xop_unlinkall {
 981         hammer2_xop_head_t      head;
 982         hammer2_key_t           key_beg;
 983         hammer2_key_t           key_end;
 984 };
 985
 986 struct hammer2_xop_connect {
 987         hammer2_xop_head_t      head;
 988         hammer2_key_t           lhc;
 989 };
 990
 991 struct hammer2_xop_flush {
 992         hammer2_xop_head_t      head;
 993 };
 994
 995 typedef struct hammer2_xop_readdir hammer2_xop_readdir_t;
 996 typedef struct hammer2_xop_nresolve hammer2_xop_nresolve_t;
 997 typedef struct hammer2_xop_unlink hammer2_xop_unlink_t;
 998 typedef struct hammer2_xop_nrename hammer2_xop_nrename_t;
 999 typedef struct hammer2_xop_ipcluster hammer2_xop_ipcluster_t;
1000 typedef struct hammer2_xop_strategy hammer2_xop_strategy_t;
1001 typedef struct hammer2_xop_mkdirent hammer2_xop_mkdirent_t;
1002 typedef struct hammer2_xop_create hammer2_xop_create_t;
1003 typedef struct hammer2_xop_destroy hammer2_xop_destroy_t;
1004 typedef struct hammer2_xop_fsync hammer2_xop_fsync_t;
1005 typedef struct hammer2_xop_unlinkall hammer2_xop_unlinkall_t;
1006 typedef struct hammer2_xop_scanlhc hammer2_xop_scanlhc_t;
1007 typedef struct hammer2_xop_scanall hammer2_xop_scanall_t;
1008 typedef struct hammer2_xop_lookup hammer2_xop_lookup_t;
1009 typedef struct hammer2_xop_connect hammer2_xop_connect_t;
1010 typedef struct hammer2_xop_flush hammer2_xop_flush_t;
1011
1012 union hammer2_xop {
1013         hammer2_xop_head_t      head;
1014         hammer2_xop_ipcluster_t xop_ipcluster;
1015         hammer2_xop_readdir_t   xop_readdir;
1016         hammer2_xop_nresolve_t  xop_nresolve;
1017         hammer2_xop_unlink_t    xop_unlink;
1018         hammer2_xop_nrename_t   xop_nrename;
1019         hammer2_xop_strategy_t  xop_strategy;
1020         hammer2_xop_mkdirent_t  xop_mkdirent;
1021         hammer2_xop_create_t    xop_create;
1022         hammer2_xop_destroy_t   xop_destroy;
1023         hammer2_xop_fsync_t     xop_fsync;
1024         hammer2_xop_unlinkall_t xop_unlinkall;
1025         hammer2_xop_scanlhc_t   xop_scanlhc;
1026         hammer2_xop_scanall_t   xop_scanall;
1027         hammer2_xop_lookup_t    xop_lookup;
1028         hammer2_xop_flush_t     xop_flush;
1029         hammer2_xop_connect_t   xop_connect;
1030 };
1031
1032 typedef union hammer2_xop hammer2_xop_t;
1033
1034 /*
1035  * hammer2_xop_group - Manage XOP support threads.
1036  */
1037 struct hammer2_xop_group {
1038         hammer2_thread_t        thrs[HAMMER2_MAXCLUSTER];
1039 };
1040
1041 typedef struct hammer2_xop_group hammer2_xop_group_t;
1042
1043 /*
1044  * flags to hammer2_xop_collect()
1045  */
1046 #define HAMMER2_XOP_COLLECT_NOWAIT      0x00000001
1047 #define HAMMER2_XOP_COLLECT_WAITALL     0x00000002
1048
1049 /*
1050  * flags to hammer2_xop_alloc()
1051  *
1052  * MODIFYING    - This is a modifying transaction, allocate a mtid.
1053  * RECURSE      - Recurse top-level inode (for root flushes)
1054  */
1055 #define HAMMER2_XOP_MODIFYING           0x00000001
1056 #define HAMMER2_XOP_STRATEGY            0x00000002
1057 #define HAMMER2_XOP_INODE_STOP          0x00000004
1058 #define HAMMER2_XOP_VOLHDR              0x00000008
1059
1060 /*
1061  * Global (per partition) management structure, represents a hard block
1062  * device.  Typically referenced by hammer2_chain structures when applicable.
1063  * Typically not used for network-managed elements.
1064  *
1065  * Note that a single hammer2_dev can be indirectly tied to multiple system
1066  * mount points.  There is no direct relationship.  System mounts are
1067  * per-cluster-id, not per-block-device, and a single hard mount might contain
1068  * many PFSs and those PFSs might combine together in various ways to form
1069  * the set of available clusters.
1070  */
1071 struct hammer2_dev {
1072         struct vnode    *devvp;         /* device vnode */
1073         int             ronly;          /* read-only mount */
1074         int             mount_count;    /* number of actively mounted PFSs */
1075         TAILQ_ENTRY(hammer2_dev) mntentry; /* hammer2_mntlist */
1076
1077         struct malloc_type *mchain;
1078         int             nipstacks;
1079         int             maxipstacks;
1080         kdmsg_iocom_t   iocom;          /* volume-level dmsg interface */
1081         struct spinlock io_spin;        /* iotree, iolruq access */
1082         struct hammer2_io_tree iotree;
1083         int             iofree_count;
1084         int             freemap_relaxed;
1085         hammer2_chain_t vchain;         /* anchor chain (topology) */
1086         hammer2_chain_t fchain;         /* anchor chain (freemap) */
1087         struct spinlock list_spin;
1088         struct h2_flush_list flushq;    /* flush seeds */
1089         struct hammer2_pfs *spmp;       /* super-root pmp for transactions */
1090         struct lock     vollk;          /* lockmgr lock */
1091         struct lock     bulklk;         /* bulkfree operation lock */
1092         struct lock     bflock;         /* bulk-free manual function lock */
1093         hammer2_off_t   heur_freemap[HAMMER2_FREEMAP_HEUR_SIZE];
1094         hammer2_dedup_t heur_dedup[HAMMER2_DEDUP_HEUR_SIZE];
1095         int             volhdrno;       /* last volhdrno written */
1096         uint32_t        hflags;         /* HMNT2 flags applicable to device */
1097         hammer2_off_t   free_reserved;  /* nominal free reserved */
1098         hammer2_thread_t bfthr;         /* bulk-free thread */
1099         char            devrepname[64]; /* for kprintf */
1100         hammer2_ioc_bulkfree_t bflast;  /* stats for last bulkfree run */
1101         hammer2_volume_data_t voldata;
1102         hammer2_volume_data_t volsync;  /* synchronized voldata */
1103 };
1104
1105 typedef struct hammer2_dev hammer2_dev_t;
1106
1107 /*
1108  * Helper functions (cluster must be locked for flags to be valid).
1109  */
1110 static __inline
1111 int
1112 hammer2_chain_rdok(hammer2_chain_t *chain)
1113 {
1114         return (chain->error == 0);
1115 }
1116
1117 static __inline
1118 int
1119 hammer2_chain_wrok(hammer2_chain_t *chain)
1120 {
1121         return (chain->error == 0 && chain->hmp->ronly == 0);
1122 }
1123
1124 /*
1125  * Per-cluster management structure.  This structure will be tied to a
1126  * system mount point if the system is mounting the PFS, but is also used
1127  * to manage clusters encountered during the super-root scan or received
1128  * via LNK_SPANs that might not be mounted.
1129  *
1130  * This structure is also used to represent the super-root that hangs off
1131  * of a hard mount point.  The super-root is not really a cluster element.
1132  * In this case the spmp_hmp field will be non-NULL.  It's just easier to do
1133  * this than to special case super-root manipulation in the hammer2_chain*
1134  * code as being only hammer2_dev-related.
1135  *
1136  * pfs_mode and pfs_nmasters are rollup fields which critically describes
1137  * how elements of the cluster act on the cluster.  pfs_mode is only applicable
1138  * when a PFS is mounted by the system.  pfs_nmasters is our best guess as to
1139  * how many masters have been configured for a cluster and is always
1140  * applicable.  pfs_types[] is an array with 1:1 correspondance to the
1141  * iroot cluster and describes the PFS types of the nodes making up the
1142  * cluster.
1143  *
1144  * WARNING! Portions of this structure have deferred initialization.  In
1145  *          particular, if not mounted there will be no wthread.
1146  *          umounted network PFSs will also be missing iroot and numerous
1147  *          other fields will not be initialized prior to mount.
1148  *
1149  *          Synchronization threads are chain-specific and only applicable
1150  *          to local hard PFS entries.  A hammer2_pfs structure may contain
1151  *          more than one when multiple hard PFSs are present on the local
1152  *          machine which require synchronization monitoring.  Most PFSs
1153  *          (such as snapshots) are 1xMASTER PFSs which do not need a
1154  *          synchronization thread.
1155  *
1156  * WARNING! The chains making up pfs->iroot's cluster are accounted for in
1157  *          hammer2_dev->mount_count when the pfs is associated with a mount
1158  *          point.
1159  */
1160 struct hammer2_pfs {
1161         struct mount            *mp;
1162         TAILQ_ENTRY(hammer2_pfs) mntentry;      /* hammer2_pfslist */
1163         uuid_t                  pfs_clid;
1164         hammer2_dev_t           *spmp_hmp;      /* only if super-root pmp */
1165         hammer2_dev_t           *force_local;   /* only if 'local' mount */
1166         hammer2_inode_t         *iroot;         /* PFS root inode */
1167         uint8_t                 pfs_types[HAMMER2_MAXCLUSTER];
1168         char                    *pfs_names[HAMMER2_MAXCLUSTER];
1169         hammer2_dev_t           *pfs_hmps[HAMMER2_MAXCLUSTER];
1170         hammer2_trans_t         trans;
1171         struct lock             lock;           /* PFS lock for certain ops */
1172         struct lock             lock_nlink;     /* rename and nlink lock */
1173         struct netexport        export;         /* nfs export */
1174         int                     speedup_ticks;  /* speedup_syncer() helper */
1175         int                     ronly;          /* read-only mount */
1176         int                     hflags;         /* pfs-specific mount flags */
1177         struct malloc_type      *minode;
1178         struct malloc_type      *mmsg;
1179         struct spinlock         inum_spin;      /* inumber lookup */
1180         struct hammer2_inode_tree inum_tree;    /* (not applicable to spmp) */
1181         long                    inum_count;     /* #of inodes in inum_tree */
1182         struct spinlock         lru_spin;       /* inumber lookup */
1183         struct hammer2_chain_list lru_list;     /* basis for LRU tests */
1184         int                     lru_count;      /* #of chains on LRU */
1185         int                     flags;
1186         hammer2_tid_t           modify_tid;     /* modify transaction id */
1187         hammer2_tid_t           inode_tid;      /* inode allocator */
1188         uint8_t                 pfs_nmasters;   /* total masters */
1189         uint8_t                 pfs_mode;       /* operating mode PFSMODE */
1190         uint8_t                 unused01;
1191         uint8_t                 unused02;
1192         int                     free_ticks;     /* free_* calculations */
1193         long                    inmem_inodes;
1194         hammer2_off_t           free_reserved;
1195         hammer2_off_t           free_nominal;
1196         uint32_t                inmem_dirty_chains;
1197         int                     count_lwinprog; /* logical write in prog */
1198         struct spinlock         list_spin;
1199         struct h2_sideq_list    sideq;          /* last-close dirty/unlink */
1200         long                    sideq_count;
1201         hammer2_thread_t        sync_thrs[HAMMER2_MAXCLUSTER];
1202         uint32_t                cluster_flags;  /* cached cluster flags */
1203         int                     has_xop_threads;
1204         struct spinlock         xop_spin;       /* xop sequencer */
1205         hammer2_xop_group_t     xop_groups[HAMMER2_XOPGROUPS];
1206 };
1207
1208 typedef struct hammer2_pfs hammer2_pfs_t;
1209
1210 TAILQ_HEAD(hammer2_pfslist, hammer2_pfs);
1211
1212 #define HAMMER2_PMPF_SPMP       0x00000001
1213
1214 /*
1215  * NOTE: The LRU list contains at least all the chains with refs == 0
1216  *       that can be recycled, and may contain additional chains which
1217  *       cannot.
1218  */
1219 #define HAMMER2_LRU_LIMIT               4096
1220
1221 #define HAMMER2_DIRTYCHAIN_WAITING      0x80000000
1222 #define HAMMER2_DIRTYCHAIN_MASK         0x7FFFFFFF
1223
1224 #define HAMMER2_LWINPROG_WAITING        0x80000000
1225 #define HAMMER2_LWINPROG_WAITING0       0x40000000
1226 #define HAMMER2_LWINPROG_MASK           0x3FFFFFFF
1227
1228 /*
1229  * hammer2_cluster_check
1230  */
1231 #define HAMMER2_CHECK_NULL      0x00000001
1232
1233 /*
1234  * Misc
1235  */
1236 #if defined(_KERNEL) || defined(_KERNEL_STRUCTURES)
1237 #define VTOI(vp)        ((hammer2_inode_t *)(vp)->v_data)
1238 #endif
1239
1240 #if defined(_KERNEL)
1241
1242 MALLOC_DECLARE(M_HAMMER2);
1243
1244 #define ITOV(ip)        ((ip)->vp)
1245
1246 /*
1247  * Currently locked chains retain the locked buffer cache buffer for
1248  * indirect blocks, and indirect blocks can be one of two sizes.  The
1249  * device buffer has to match the case to avoid deadlocking recursive
1250  * chains that might otherwise try to access different offsets within
1251  * the same device buffer.
1252  */
1253 static __inline
1254 int
1255 hammer2_devblkradix(int radix)
1256 {
1257 #if 0
1258         if (radix <= HAMMER2_LBUFRADIX) {
1259                 return (HAMMER2_LBUFRADIX);
1260         } else {
1261                 return (HAMMER2_PBUFRADIX);
1262         }
1263 #endif
1264         return (HAMMER2_PBUFRADIX);
1265 }
1266
1267 /*
1268  * XXX almost time to remove this.  DIO uses PBUFSIZE exclusively now.
1269  */
1270 static __inline
1271 size_t
1272 hammer2_devblksize(size_t bytes)
1273 {
1274 #if 0
1275         if (bytes <= HAMMER2_LBUFSIZE) {
1276                 return(HAMMER2_LBUFSIZE);
1277         } else {
1278                 KKASSERT(bytes <= HAMMER2_PBUFSIZE &&
1279                          (bytes ^ (bytes - 1)) == ((bytes << 1) - 1));
1280                 return (HAMMER2_PBUFSIZE);
1281         }
1282 #endif
1283         return (HAMMER2_PBUFSIZE);
1284 }
1285
1286
1287 static __inline
1288 hammer2_pfs_t *
1289 MPTOPMP(struct mount *mp)
1290 {
1291         return ((hammer2_pfs_t *)mp->mnt_data);
1292 }
1293
1294 #define HAMMER2_DEDUP_FRAG      (HAMMER2_PBUFSIZE / 64)
1295 #define HAMMER2_DEDUP_FRAGRADIX (HAMMER2_PBUFRADIX - 6)
1296
1297 static __inline
1298 uint64_t
1299 hammer2_dedup_mask(hammer2_io_t *dio, hammer2_off_t data_off, u_int bytes)
1300 {
1301         int bbeg;
1302         int bits;
1303         uint64_t mask;
1304
1305         bbeg = (int)((data_off & ~HAMMER2_OFF_MASK_RADIX) - dio->pbase) >>
1306                HAMMER2_DEDUP_FRAGRADIX;
1307         bits = (int)((bytes + (HAMMER2_DEDUP_FRAG - 1)) >>
1308                HAMMER2_DEDUP_FRAGRADIX);
1309         mask = ((uint64_t)1 << bbeg) - 1;
1310         if (bbeg + bits == 64)
1311                 mask = (uint64_t)-1;
1312         else
1313                 mask = ((uint64_t)1 << (bbeg + bits)) - 1;
1314
1315         mask &= ~(((uint64_t)1 << bbeg) - 1);
1316
1317         return mask;
1318 }
1319
1320 static __inline
1321 int
1322 hammer2_error_to_errno(int error)
1323 {
1324         if (error) {
1325                 if (error & HAMMER2_ERROR_EIO)
1326                         error = EIO;
1327                 else if (error & HAMMER2_ERROR_CHECK)
1328                         error = EDOM;
1329                 else if (error & HAMMER2_ERROR_ABORTED)
1330                         error = EINTR;
1331                 else if (error & HAMMER2_ERROR_BADBREF)
1332                         error = EIO;
1333                 else if (error & HAMMER2_ERROR_ENOSPC)
1334                         error = ENOSPC;
1335                 else if (error & HAMMER2_ERROR_ENOENT)
1336                         error = ENOENT;
1337                 else if (error & HAMMER2_ERROR_ENOTEMPTY)
1338                         error = ENOTEMPTY;
1339                 else if (error & HAMMER2_ERROR_EAGAIN)
1340                         error = EAGAIN;
1341                 else if (error & HAMMER2_ERROR_ENOTDIR)
1342                         error = ENOTDIR;
1343                 else if (error & HAMMER2_ERROR_EISDIR)
1344                         error = EISDIR;
1345                 else if (error & HAMMER2_ERROR_EINPROGRESS)
1346                         error = EINPROGRESS;
1347                 else if (error & HAMMER2_ERROR_EEXIST)
1348                         error = EEXIST;
1349                 else
1350                         error = EDOM;
1351         }
1352         return error;
1353 }
1354
1355 static __inline
1356 int
1357 hammer2_errno_to_error(int error)
1358 {
1359         switch(error) {
1360         case 0:
1361                 return 0;
1362         case EIO:
1363                 return HAMMER2_ERROR_EIO;
1364         case EINVAL:
1365         default:
1366                 return HAMMER2_ERROR_EINVAL;
1367         }
1368 }
1369
1370 extern struct vop_ops hammer2_vnode_vops;
1371 extern struct vop_ops hammer2_spec_vops;
1372 extern struct vop_ops hammer2_fifo_vops;
1373 extern struct hammer2_pfslist hammer2_pfslist;
1374 extern struct lock hammer2_mntlk;
1375
1376
1377 extern int hammer2_debug;
1378 extern int hammer2_cluster_meta_read;
1379 extern int hammer2_cluster_data_read;
1380 extern int hammer2_dedup_enable;
1381 extern int hammer2_always_compress;
1382 extern int hammer2_inval_enable;
1383 extern int hammer2_flush_pipe;
1384 extern int hammer2_dio_count;
1385 extern int hammer2_dio_limit;
1386 extern int hammer2_bulkfree_tps;
1387 extern long hammer2_chain_allocs;
1388 extern long hammer2_chain_frees;
1389 extern long hammer2_limit_dirty_chains;
1390 extern long hammer2_count_modified_chains;
1391 extern long hammer2_iod_invals;
1392 extern long hammer2_iod_file_read;
1393 extern long hammer2_iod_meta_read;
1394 extern long hammer2_iod_indr_read;
1395 extern long hammer2_iod_fmap_read;
1396 extern long hammer2_iod_volu_read;
1397 extern long hammer2_iod_file_write;
1398 extern long hammer2_iod_file_wembed;
1399 extern long hammer2_iod_file_wzero;
1400 extern long hammer2_iod_file_wdedup;
1401 extern long hammer2_iod_meta_write;
1402 extern long hammer2_iod_indr_write;
1403 extern long hammer2_iod_fmap_write;
1404 extern long hammer2_iod_volu_write;
1405
1406 extern long hammer2_process_xxhash64;
1407 extern long hammer2_process_icrc32;
1408
1409 extern struct objcache *cache_buffer_read;
1410 extern struct objcache *cache_buffer_write;
1411 extern struct objcache *cache_xops;
1412
1413 /*
1414  * hammer2_subr.c
1415  */
1416 #define hammer2_icrc32(buf, size)       iscsi_crc32((buf), (size))
1417 #define hammer2_icrc32c(buf, size, crc) iscsi_crc32_ext((buf), (size), (crc))
1418
1419 int hammer2_signal_check(time_t *timep);
1420 const char *hammer2_error_str(int error);
1421
1422 void hammer2_inode_lock(hammer2_inode_t *ip, int how);
1423 void hammer2_inode_unlock(hammer2_inode_t *ip);
1424 hammer2_chain_t *hammer2_inode_chain(hammer2_inode_t *ip, int clindex, int how);
1425 hammer2_chain_t *hammer2_inode_chain_and_parent(hammer2_inode_t *ip,
1426                         int clindex, hammer2_chain_t **parentp, int how);
1427 hammer2_mtx_state_t hammer2_inode_lock_temp_release(hammer2_inode_t *ip);
1428 void hammer2_inode_lock_temp_restore(hammer2_inode_t *ip,
1429                         hammer2_mtx_state_t ostate);
1430 int hammer2_inode_lock_upgrade(hammer2_inode_t *ip);
1431 void hammer2_inode_lock_downgrade(hammer2_inode_t *ip, int);
1432
1433 void hammer2_dev_exlock(hammer2_dev_t *hmp);
1434 void hammer2_dev_shlock(hammer2_dev_t *hmp);
1435 void hammer2_dev_unlock(hammer2_dev_t *hmp);
1436
1437 int hammer2_get_dtype(uint8_t type);
1438 int hammer2_get_vtype(uint8_t type);
1439 uint8_t hammer2_get_obj_type(enum vtype vtype);
1440 void hammer2_time_to_timespec(uint64_t xtime, struct timespec *ts);
1441 uint64_t hammer2_timespec_to_time(const struct timespec *ts);
1442 uint32_t hammer2_to_unix_xid(const uuid_t *uuid);
1443 void hammer2_guid_to_uuid(uuid_t *uuid, uint32_t guid);
1444 void hammer2_trans_manage_init(hammer2_pfs_t *pmp);
1445
1446 hammer2_key_t hammer2_dirhash(const unsigned char *name, size_t len);
1447 int hammer2_getradix(size_t bytes);
1448
1449 int hammer2_calc_logical(hammer2_inode_t *ip, hammer2_off_t uoff,
1450                         hammer2_key_t *lbasep, hammer2_key_t *leofp);
1451 int hammer2_calc_physical(hammer2_inode_t *ip, hammer2_key_t lbase);
1452 void hammer2_update_time(uint64_t *timep);
1453 void hammer2_adjreadcounter(hammer2_blockref_t *bref, size_t bytes);
1454
1455 /*
1456  * hammer2_inode.c
1457  */
1458 struct vnode *hammer2_igetv(hammer2_inode_t *ip, int *errorp);
1459 hammer2_inode_t *hammer2_inode_lookup(hammer2_pfs_t *pmp,
1460                         hammer2_tid_t inum);
1461 hammer2_inode_t *hammer2_inode_get(hammer2_pfs_t *pmp, hammer2_inode_t *dip,
1462                         hammer2_cluster_t *cluster, int idx);
1463 void hammer2_inode_free(hammer2_inode_t *ip);
1464 void hammer2_inode_ref(hammer2_inode_t *ip);
1465 void hammer2_inode_drop(hammer2_inode_t *ip);
1466 void hammer2_inode_repoint(hammer2_inode_t *ip, hammer2_inode_t *pip,
1467                         hammer2_cluster_t *cluster);
1468 void hammer2_inode_repoint_one(hammer2_inode_t *ip, hammer2_cluster_t *cluster,
1469                         int idx);
1470 void hammer2_inode_modify(hammer2_inode_t *ip);
1471 void hammer2_inode_run_sideq(hammer2_pfs_t *pmp, int doall);
1472
1473 hammer2_inode_t *hammer2_inode_create(hammer2_inode_t *dip,
1474                         hammer2_inode_t *pip,
1475                         struct vattr *vap, struct ucred *cred,
1476                         const uint8_t *name, size_t name_len, hammer2_key_t lhc,
1477                         hammer2_key_t inum, uint8_t type, uint8_t target_type,
1478                         int flags, int *errorp);
1479 int hammer2_inode_chain_sync(hammer2_inode_t *ip);
1480 int hammer2_inode_chain_flush(hammer2_inode_t *ip);
1481 int hammer2_inode_unlink_finisher(hammer2_inode_t *ip, int isopen);
1482 int hammer2_dirent_create(hammer2_inode_t *dip, const char *name,
1483                         size_t name_len, hammer2_key_t inum, uint8_t type);
1484
1485 /*
1486  * hammer2_chain.c
1487  */
1488 void hammer2_voldata_lock(hammer2_dev_t *hmp);
1489 void hammer2_voldata_unlock(hammer2_dev_t *hmp);
1490 void hammer2_voldata_modify(hammer2_dev_t *hmp);
1491 hammer2_chain_t *hammer2_chain_alloc(hammer2_dev_t *hmp,
1492                                 hammer2_pfs_t *pmp,
1493                                 hammer2_blockref_t *bref);
1494 void hammer2_chain_core_init(hammer2_chain_t *chain);
1495 void hammer2_chain_ref(hammer2_chain_t *chain);
1496 void hammer2_chain_ref_hold(hammer2_chain_t *chain);
1497 void hammer2_chain_drop(hammer2_chain_t *chain);
1498 void hammer2_chain_drop_unhold(hammer2_chain_t *chain);
1499 int hammer2_chain_lock(hammer2_chain_t *chain, int how);
1500 void hammer2_chain_lock_unhold(hammer2_chain_t *chain, int how);
1501 void hammer2_chain_load_data(hammer2_chain_t *chain);
1502 const hammer2_media_data_t *hammer2_chain_rdata(hammer2_chain_t *chain);
1503 hammer2_media_data_t *hammer2_chain_wdata(hammer2_chain_t *chain);
1504
1505 int hammer2_chain_inode_find(hammer2_pfs_t *pmp, hammer2_key_t inum,
1506                                 int clindex, int flags,
1507                                 hammer2_chain_t **parentp,
1508                                 hammer2_chain_t **chainp);
1509 int hammer2_chain_modify(hammer2_chain_t *chain, hammer2_tid_t mtid,
1510                                 hammer2_off_t dedup_off, int flags);
1511 int hammer2_chain_modify_ip(hammer2_inode_t *ip, hammer2_chain_t *chain,
1512                                 hammer2_tid_t mtid, int flags);
1513 int hammer2_chain_resize(hammer2_chain_t *chain,
1514                                 hammer2_tid_t mtid, hammer2_off_t dedup_off,
1515                                 int nradix, int flags);
1516 void hammer2_chain_unlock(hammer2_chain_t *chain);
1517 void hammer2_chain_unlock_hold(hammer2_chain_t *chain);
1518 void hammer2_chain_wait(hammer2_chain_t *chain);
1519 hammer2_chain_t *hammer2_chain_get(hammer2_chain_t *parent, int generation,
1520                                 hammer2_blockref_t *bref, int how);
1521 hammer2_chain_t *hammer2_chain_lookup_init(hammer2_chain_t *parent, int flags);
1522 void hammer2_chain_lookup_done(hammer2_chain_t *parent);
1523 hammer2_chain_t *hammer2_chain_getparent(hammer2_chain_t *chain, int flags);
1524 hammer2_chain_t *hammer2_chain_repparent(hammer2_chain_t **chainp, int flags);
1525 hammer2_chain_t *hammer2_chain_lookup(hammer2_chain_t **parentp,
1526                                 hammer2_key_t *key_nextp,
1527                                 hammer2_key_t key_beg, hammer2_key_t key_end,
1528                                 int *errorp, int flags);
1529 hammer2_chain_t *hammer2_chain_next(hammer2_chain_t **parentp,
1530                                 hammer2_chain_t *chain,
1531                                 hammer2_key_t *key_nextp,
1532                                 hammer2_key_t key_beg, hammer2_key_t key_end,
1533                                 int *errorp, int flags);
1534 int hammer2_chain_scan(hammer2_chain_t *parent,
1535                                 hammer2_chain_t **chainp,
1536                                 hammer2_blockref_t *bref,
1537                                 int *firstp, int flags);
1538
1539 int hammer2_chain_create(hammer2_chain_t **parentp, hammer2_chain_t **chainp,
1540                                 hammer2_pfs_t *pmp, int methods,
1541                                 hammer2_key_t key, int keybits,
1542                                 int type, size_t bytes, hammer2_tid_t mtid,
1543                                 hammer2_off_t dedup_off, int flags);
1544 void hammer2_chain_rename(hammer2_chain_t **parentp,
1545                                 hammer2_chain_t *chain,
1546                                 hammer2_tid_t mtid, int flags);
1547 int hammer2_chain_delete(hammer2_chain_t *parent, hammer2_chain_t *chain,
1548                                 hammer2_tid_t mtid, int flags);
1549 int hammer2_chain_indirect_maintenance(hammer2_chain_t *parent,
1550                                 hammer2_chain_t *chain);
1551 void hammer2_chain_setflush(hammer2_chain_t *chain);
1552 void hammer2_chain_countbrefs(hammer2_chain_t *chain,
1553                                 hammer2_blockref_t *base, int count);
1554 hammer2_chain_t *hammer2_chain_bulksnap(hammer2_dev_t *hmp);
1555 void hammer2_chain_bulkdrop(hammer2_chain_t *copy);
1556
1557 void hammer2_chain_setcheck(hammer2_chain_t *chain, void *bdata);
1558 int hammer2_chain_testcheck(hammer2_chain_t *chain, void *bdata);
1559 int hammer2_chain_dirent_test(hammer2_chain_t *chain, const char *name,
1560                                 size_t name_len);
1561
1562 void hammer2_pfs_memory_wait(hammer2_inode_t *ip, int always_moderate);
1563 void hammer2_pfs_memory_inc(hammer2_pfs_t *pmp);
1564 void hammer2_pfs_memory_wakeup(hammer2_pfs_t *pmp);
1565
1566 void hammer2_base_delete(hammer2_chain_t *parent,
1567                                 hammer2_blockref_t *base, int count,
1568                                 hammer2_chain_t *chain);
1569 void hammer2_base_insert(hammer2_chain_t *parent,
1570                                 hammer2_blockref_t *base, int count,
1571                                 hammer2_chain_t *chain,
1572                                 hammer2_blockref_t *elm);
1573
1574 /*
1575  * hammer2_flush.c
1576  */
1577 int hammer2_flush(hammer2_chain_t *chain, int istop);
1578 void hammer2_delayed_flush(hammer2_chain_t *chain);
1579
1580 /*
1581  * hammer2_trans.c
1582  */
1583 void hammer2_trans_init(hammer2_pfs_t *pmp, uint32_t flags);
1584 hammer2_tid_t hammer2_trans_sub(hammer2_pfs_t *pmp);
1585 void hammer2_trans_done(hammer2_pfs_t *pmp, int quicksideq);
1586 hammer2_tid_t hammer2_trans_newinum(hammer2_pfs_t *pmp);
1587 void hammer2_trans_assert_strategy(hammer2_pfs_t *pmp);
1588 void hammer2_dedup_record(hammer2_chain_t *chain, hammer2_io_t *dio,
1589                                 char *data);
1590
1591 /*
1592  * hammer2_ioctl.c
1593  */
1594 int hammer2_ioctl(hammer2_inode_t *ip, u_long com, void *data,
1595                                 int fflag, struct ucred *cred);
1596
1597 /*
1598  * hammer2_io.c
1599  */
1600 void hammer2_io_putblk(hammer2_io_t **diop);
1601 void hammer2_io_inval(hammer2_io_t *dio, hammer2_off_t data_off, u_int bytes);
1602 void hammer2_io_cleanup(hammer2_dev_t *hmp, struct hammer2_io_tree *tree);
1603 char *hammer2_io_data(hammer2_io_t *dio, off_t lbase);
1604 void hammer2_io_bkvasync(hammer2_io_t *dio);
1605 hammer2_io_t *hammer2_io_getblk(hammer2_dev_t *hmp, int btype, off_t lbase,
1606                                 int lsize, int op);
1607 void hammer2_io_dedup_set(hammer2_dev_t *hmp, hammer2_blockref_t *bref);
1608 void hammer2_io_dedup_delete(hammer2_dev_t *hmp, uint8_t btype,
1609                                 hammer2_off_t data_off, u_int bytes);
1610 void hammer2_io_dedup_assert(hammer2_dev_t *hmp, hammer2_off_t data_off,
1611                                 u_int bytes);
1612 void hammer2_io_callback(struct bio *bio);
1613 int hammer2_io_new(hammer2_dev_t *hmp, int btype, off_t lbase, int lsize,
1614                                 hammer2_io_t **diop);
1615 int hammer2_io_newnz(hammer2_dev_t *hmp, int btype, off_t lbase, int lsize,
1616                                 hammer2_io_t **diop);
1617 int hammer2_io_bread(hammer2_dev_t *hmp, int btype, off_t lbase, int lsize,
1618                                 hammer2_io_t **diop);
1619 hammer2_io_t *hammer2_io_getquick(hammer2_dev_t *hmp, off_t lbase, int lsize);
1620 void hammer2_io_bawrite(hammer2_io_t **diop);
1621 void hammer2_io_bdwrite(hammer2_io_t **diop);
1622 int hammer2_io_bwrite(hammer2_io_t **diop);
1623 void hammer2_io_setdirty(hammer2_io_t *dio);
1624 void hammer2_io_brelse(hammer2_io_t **diop);
1625 void hammer2_io_bqrelse(hammer2_io_t **diop);
1626
1627 /*
1628  * hammer2_thread.c
1629  */
1630 void hammer2_thr_signal(hammer2_thread_t *thr, uint32_t flags);
1631 void hammer2_thr_signal2(hammer2_thread_t *thr,
1632                         uint32_t pflags, uint32_t nflags);
1633 void hammer2_thr_wait(hammer2_thread_t *thr, uint32_t flags);
1634 void hammer2_thr_wait_neg(hammer2_thread_t *thr, uint32_t flags);
1635 int hammer2_thr_wait_any(hammer2_thread_t *thr, uint32_t flags, int timo);
1636 void hammer2_thr_create(hammer2_thread_t *thr,
1637                         hammer2_pfs_t *pmp, hammer2_dev_t *hmp,
1638                         const char *id, int clindex, int repidx,
1639                         void (*func)(void *arg));
1640 void hammer2_thr_delete(hammer2_thread_t *thr);
1641 void hammer2_thr_remaster(hammer2_thread_t *thr);
1642 void hammer2_thr_freeze_async(hammer2_thread_t *thr);
1643 void hammer2_thr_freeze(hammer2_thread_t *thr);
1644 void hammer2_thr_unfreeze(hammer2_thread_t *thr);
1645 int hammer2_thr_break(hammer2_thread_t *thr);
1646 void hammer2_primary_xops_thread(void *arg);
1647
1648 /*
1649  * hammer2_thread.c (XOP API)
1650  */
1651 void hammer2_xop_group_init(hammer2_pfs_t *pmp, hammer2_xop_group_t *xgrp);
1652 void *hammer2_xop_alloc(hammer2_inode_t *ip, int flags);
1653 void hammer2_xop_setname(hammer2_xop_head_t *xop,
1654                                 const char *name, size_t name_len);
1655 void hammer2_xop_setname2(hammer2_xop_head_t *xop,
1656                                 const char *name, size_t name_len);
1657 size_t hammer2_xop_setname_inum(hammer2_xop_head_t *xop, hammer2_key_t inum);
1658 void hammer2_xop_setip2(hammer2_xop_head_t *xop, hammer2_inode_t *ip2);
1659 void hammer2_xop_setip3(hammer2_xop_head_t *xop, hammer2_inode_t *ip3);
1660 void hammer2_xop_reinit(hammer2_xop_head_t *xop);
1661 void hammer2_xop_helper_create(hammer2_pfs_t *pmp);
1662 void hammer2_xop_helper_cleanup(hammer2_pfs_t *pmp);
1663 void hammer2_xop_start(hammer2_xop_head_t *xop, hammer2_xop_func_t func);
1664 void hammer2_xop_start_except(hammer2_xop_head_t *xop, hammer2_xop_func_t func,
1665                                 int notidx);
1666 int hammer2_xop_collect(hammer2_xop_head_t *xop, int flags);
1667 void hammer2_xop_retire(hammer2_xop_head_t *xop, uint64_t mask);
1668 int hammer2_xop_active(hammer2_xop_head_t *xop);
1669 int hammer2_xop_feed(hammer2_xop_head_t *xop, hammer2_chain_t *chain,
1670                                 int clindex, int error);
1671
1672 /*
1673  * hammer2_synchro.c
1674  */
1675 void hammer2_primary_sync_thread(void *arg);
1676
1677 /*
1678  * XOP backends in hammer2_xops.c, primarily for VNOPS.  Other XOP backends
1679  * may be integrated into other source files.
1680  */
1681 void hammer2_xop_ipcluster(hammer2_thread_t *thr, hammer2_xop_t *xop);
1682 void hammer2_xop_readdir(hammer2_thread_t *thr, hammer2_xop_t *xop);
1683 void hammer2_xop_nresolve(hammer2_thread_t *thr, hammer2_xop_t *xop);
1684 void hammer2_xop_unlink(hammer2_thread_t *thr, hammer2_xop_t *xop);
1685 void hammer2_xop_nrename(hammer2_thread_t *thr, hammer2_xop_t *xop);
1686 void hammer2_xop_scanlhc(hammer2_thread_t *thr, hammer2_xop_t *xop);
1687 void hammer2_xop_scanall(hammer2_thread_t *thr, hammer2_xop_t *xop);
1688 void hammer2_xop_lookup(hammer2_thread_t *thr, hammer2_xop_t *xop);
1689 void hammer2_inode_xop_mkdirent(hammer2_thread_t *thr, hammer2_xop_t *xop);
1690 void hammer2_inode_xop_create(hammer2_thread_t *thr, hammer2_xop_t *xop);
1691 void hammer2_inode_xop_destroy(hammer2_thread_t *thr, hammer2_xop_t *xop);
1692 void hammer2_inode_xop_chain_sync(hammer2_thread_t *thr, hammer2_xop_t *xop);
1693 void hammer2_inode_xop_unlinkall(hammer2_thread_t *thr, hammer2_xop_t *xop);
1694 void hammer2_inode_xop_connect(hammer2_thread_t *thr, hammer2_xop_t *xop);
1695 void hammer2_inode_xop_flush(hammer2_thread_t *thr, hammer2_xop_t *xop);
1696
1697 /*
1698  * hammer2_msgops.c
1699  */
1700 int hammer2_msg_dbg_rcvmsg(kdmsg_msg_t *msg);
1701 int hammer2_msg_adhoc_input(kdmsg_msg_t *msg);
1702
1703 /*
1704  * hammer2_vfsops.c
1705  */
1706 void hammer2_volconf_update(hammer2_dev_t *hmp, int index);
1707 void hammer2_dump_chain(hammer2_chain_t *chain, int tab, int *countp, char pfx,
1708                                 u_int flags);
1709 int hammer2_vfs_sync(struct mount *mp, int waitflags);
1710 int hammer2_vfs_enospace(hammer2_inode_t *ip, off_t bytes, struct ucred *cred);
1711
1712 hammer2_pfs_t *hammer2_pfsalloc(hammer2_chain_t *chain,
1713                                 const hammer2_inode_data_t *ripdata,
1714                                 hammer2_tid_t modify_tid,
1715                                 hammer2_dev_t *force_local);
1716 void hammer2_pfsdealloc(hammer2_pfs_t *pmp, int clindex, int destroying);
1717 int hammer2_vfs_vget(struct mount *mp, struct vnode *dvp,
1718                                 ino_t ino, struct vnode **vpp);
1719
1720 void hammer2_lwinprog_ref(hammer2_pfs_t *pmp);
1721 void hammer2_lwinprog_drop(hammer2_pfs_t *pmp);
1722 void hammer2_lwinprog_wait(hammer2_pfs_t *pmp, int pipe);
1723
1724 /*
1725  * hammer2_freemap.c
1726  */
1727 int hammer2_freemap_alloc(hammer2_chain_t *chain, size_t bytes);
1728 void hammer2_freemap_adjust(hammer2_dev_t *hmp,
1729                                 hammer2_blockref_t *bref, int how);
1730
1731 /*
1732  * hammer2_cluster.c
1733  */
1734 uint8_t hammer2_cluster_type(hammer2_cluster_t *cluster);
1735 const hammer2_media_data_t *hammer2_cluster_rdata(hammer2_cluster_t *cluster);
1736 hammer2_media_data_t *hammer2_cluster_wdata(hammer2_cluster_t *cluster);
1737 hammer2_cluster_t *hammer2_cluster_from_chain(hammer2_chain_t *chain);
1738 void hammer2_cluster_bref(hammer2_cluster_t *cluster, hammer2_blockref_t *bref);
1739 hammer2_cluster_t *hammer2_cluster_alloc(hammer2_pfs_t *pmp,
1740                                 hammer2_blockref_t *bref);
1741 void hammer2_cluster_ref(hammer2_cluster_t *cluster);
1742 void hammer2_cluster_drop(hammer2_cluster_t *cluster);
1743 void hammer2_cluster_lock(hammer2_cluster_t *cluster, int how);
1744 int hammer2_cluster_check(hammer2_cluster_t *cluster, hammer2_key_t lokey,
1745                         int flags);
1746 void hammer2_cluster_resolve(hammer2_cluster_t *cluster);
1747 void hammer2_cluster_forcegood(hammer2_cluster_t *cluster);
1748 void hammer2_cluster_unlock(hammer2_cluster_t *cluster);
1749
1750 void hammer2_bulkfree_init(hammer2_dev_t *hmp);
1751 void hammer2_bulkfree_uninit(hammer2_dev_t *hmp);
1752 int hammer2_bulkfree_pass(hammer2_dev_t *hmp, hammer2_chain_t *vchain,
1753                         struct hammer2_ioc_bulkfree *bfi);
1754
1755 /*
1756  * hammer2_iocom.c
1757  */
1758 void hammer2_iocom_init(hammer2_dev_t *hmp);
1759 void hammer2_iocom_uninit(hammer2_dev_t *hmp);
1760 void hammer2_cluster_reconnect(hammer2_dev_t *hmp, struct file *fp);
1761
1762 /*
1763  * hammer2_strategy.c
1764  */
1765 int hammer2_vop_strategy(struct vop_strategy_args *ap);
1766 int hammer2_vop_bmap(struct vop_bmap_args *ap);
1767 void hammer2_write_thread(void *arg);
1768 void hammer2_bioq_sync(hammer2_pfs_t *pmp);
1769 void hammer2_dedup_clear(hammer2_dev_t *hmp);
1770
1771 #endif /* !_KERNEL */
1772 #endif /* !_VFS_HAMMER2_HAMMER2_H_ */