sys/vfs/hammer2/hammer2.h

   1 /*
   2  * Copyright (c) 2011-2013 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  * This header file contains structures used internally by the HAMMER2
  38  * implementation.  See hammer2_disk.h for on-disk structures.
  39  */
  40
  41 #ifndef _VFS_HAMMER2_HAMMER2_H_
  42 #define _VFS_HAMMER2_HAMMER2_H_
  43
  44 #include <sys/param.h>
  45 #include <sys/types.h>
  46 #include <sys/kernel.h>
  47 #include <sys/conf.h>
  48 #include <sys/systm.h>
  49 #include <sys/tree.h>
  50 #include <sys/malloc.h>
  51 #include <sys/mount.h>
  52 #include <sys/vnode.h>
  53 #include <sys/proc.h>
  54 #include <sys/mountctl.h>
  55 #include <sys/priv.h>
  56 #include <sys/stat.h>
  57 #include <sys/thread.h>
  58 #include <sys/globaldata.h>
  59 #include <sys/lockf.h>
  60 #include <sys/buf.h>
  61 #include <sys/queue.h>
  62 #include <sys/limits.h>
  63 #include <sys/buf2.h>
  64 #include <sys/signal2.h>
  65 #include <sys/dmsg.h>
  66 #include <sys/mutex.h>
  67 #include <sys/mutex2.h>
  68
  69 #include "hammer2_disk.h"
  70 #include "hammer2_mount.h"
  71 #include "hammer2_ioctl.h"
  72 #include "hammer2_ccms.h"
  73
  74 struct hammer2_chain;
  75 struct hammer2_inode;
  76 struct hammer2_mount;
  77 struct hammer2_pfsmount;
  78 struct hammer2_span;
  79 struct hammer2_state;
  80 struct hammer2_msg;
  81
  82 /*
  83  * The chain structure tracks a portion of the media topology from the
  84  * root (volume) down.  Chains represent volumes, inodes, indirect blocks,
  85  * data blocks, and freemap nodes and leafs.
  86  *
  87  * The chain structure can be multi-homed and its topological recursion
  88  * (chain->core) can be shared amongst several chains.  Chain structures
  89  * are topologically stable once placed in the in-memory topology (they
  90  * don't move around).  Modifications which cross flush synchronization
  91  * boundaries, renames, resizing, or any move of the chain to elsewhere
  92  * in the topology is accomplished via the DELETE-DUPLICATE mechanism.
  93  *
  94  * DELETE-DUPLICATE allows HAMMER2 to track work across flush synchronization
  95  * points without stalling the filesystem or corrupting the flush
  96  * sychronization point.  When necessary a chain will be marked DELETED
  97  * and a new, duplicate chain will be allocated.
  98  *
  99  * This mechanism necessarily requires that we be able to overload chains
 100  * at any given layer in the topology.  Overloading is accomplished via a
 101  * RBTREE recursion through chain->rbtree.
 102  *
 103  * Advantages:
 104  *
 105  *      (1) Fully coherent snapshots can be taken without requiring
 106  *          a pre-flush, resulting in extremely fast (sub-millisecond)
 107  *          snapshots.
 108  *
 109  *      (2) Multiple synchronization points can be in-flight at the same
 110  *          time, representing multiple snapshots or flushes.
 111  *
 112  *      (3) The algorithms needed to keep track of everything are actually
 113  *          not that complex.
 114  *
 115  * Special Considerations:
 116  *
 117  *      A chain is ref-counted on a per-chain basis, but the chain's lock
 118  *      is associated with the shared chain_core and is not per-chain.
 119  *
 120  *      The power-of-2 nature of the media radix tree ensures that there
 121  *      will be no overlaps which straddle edges.
 122  */
 123 RB_HEAD(hammer2_chain_tree, hammer2_chain);
 124 TAILQ_HEAD(h2_flush_deferral_list, hammer2_chain);
 125 TAILQ_HEAD(h2_core_list, hammer2_chain);
 126 TAILQ_HEAD(h2_layer_list, hammer2_chain_layer);
 127
 128 struct hammer2_chain_layer {
 129         int             good;
 130         TAILQ_ENTRY(hammer2_chain_layer) entry;
 131         struct hammer2_chain_tree rbtree;
 132         int             refs;           /* prevent destruction */
 133 };
 134
 135 typedef struct hammer2_chain_layer hammer2_chain_layer_t;
 136
 137 struct hammer2_chain_core {
 138         int             good;
 139         struct ccms_cst cst;
 140         struct h2_core_list ownerq;     /* all chains sharing this core */
 141         struct h2_layer_list layerq;
 142         int             live_zero;      /* blockref array opt */
 143         hammer2_tid_t   update_lo;      /* check update against parent */
 144         hammer2_tid_t   update_hi;      /* check update against parent */
 145         u_int           chain_count;    /* total chains in layers */
 146         u_int           sharecnt;
 147         u_int           flags;
 148         u_int           live_count;     /* live (not deleted) chains in tree */
 149         int             generation;     /* generation number (inserts only) */
 150 };
 151
 152 typedef struct hammer2_chain_core hammer2_chain_core_t;
 153
 154 #define HAMMER2_CORE_UNUSED0001         0x0001
 155 #define HAMMER2_CORE_COUNTEDBREFS       0x0002
 156
 157 /*
 158  * H2 is a copy-on-write filesystem.  In order to allow chains to allocate
 159  * smaller blocks (down to 64-bytes), but improve performance and make
 160  * clustered I/O possible using larger block sizes, the kernel buffer cache
 161  * is abstracted via the hammer2_io structure.
 162  */
 163 RB_HEAD(hammer2_io_tree, hammer2_io);
 164
 165 struct hammer2_io {
 166         RB_ENTRY(hammer2_io) rbnode;
 167         struct spinlock spin;
 168         struct hammer2_mount *hmp;
 169         struct buf      *bp;
 170         struct bio      *bio;
 171         off_t           pbase;
 172         int             psize;
 173         void            (*callback)(struct hammer2_io *dio,
 174                                     struct hammer2_chain *chain,
 175                                     void *arg1, off_t arg2);
 176         struct hammer2_chain *arg_c;            /* INPROG I/O only */
 177         void            *arg_p;                 /* INPROG I/O only */
 178         off_t           arg_o;                  /* INPROG I/O only */
 179         int             refs;
 180 };
 181
 182 typedef struct hammer2_io hammer2_io_t;
 183
 184 /*
 185  * Primary chain structure keeps track of the topology in-memory.
 186  */
 187 struct hammer2_chain {
 188         RB_ENTRY(hammer2_chain) rbnode;         /* node */
 189         TAILQ_ENTRY(hammer2_chain) core_entry;  /* contemporary chains */
 190         hammer2_chain_layer_t   *inlayer;
 191         hammer2_blockref_t      bref;
 192         hammer2_chain_core_t    *core;
 193         hammer2_chain_core_t    *above;
 194         struct hammer2_state    *state;         /* if active cache msg */
 195         struct hammer2_mount    *hmp;
 196         struct hammer2_pfsmount *pmp;           /* can be NULL */
 197         struct hammer2_chain    *debug_previous;
 198
 199         hammer2_tid_t   modify_tid;             /* snapshot/flush filter */
 200         hammer2_tid_t   delete_tid;
 201         hammer2_key_t   data_count;             /* delta's to apply */
 202         hammer2_key_t   inode_count;            /* delta's to apply */
 203         hammer2_io_t    *dio;                   /* physical data buffer */
 204         u_int           bytes;                  /* physical data size */
 205         u_int           flags;
 206         u_int           refs;
 207         u_int           lockcnt;
 208         int             debug_reason;
 209         int             src_reason;
 210         int             dst_reason;
 211         hammer2_media_data_t *data;             /* data pointer shortcut */
 212         TAILQ_ENTRY(hammer2_chain) flush_node;  /* flush deferral list */
 213 };
 214
 215 typedef struct hammer2_chain hammer2_chain_t;
 216
 217 int hammer2_chain_cmp(hammer2_chain_t *chain1, hammer2_chain_t *chain2);
 218 RB_PROTOTYPE(hammer2_chain_tree, hammer2_chain, rbnode, hammer2_chain_cmp);
 219
 220 /*
 221  * Special notes on flags:
 222  *
 223  * INITIAL - This flag allows a chain to be created and for storage to
 224  *           be allocated without having to immediately instantiate the
 225  *           related buffer.  The data is assumed to be all-zeros.  It
 226  *           is primarily used for indirect blocks.
 227  *
 228  * MOVED   - A modified chain becomes MOVED after it flushes.  A chain
 229  *           can also become MOVED if it is moved within the topology
 230  *           (even if not modified).
 231  *
 232  * MODIFIED- The chain's media data has been modified.
 233  */
 234 #define HAMMER2_CHAIN_MODIFIED          0x00000001      /* dirty chain data */
 235 #define HAMMER2_CHAIN_ALLOCATED         0x00000002      /* kmalloc'd chain */
 236 #define HAMMER2_CHAIN_UNUSED0004        0x00000004
 237 #define HAMMER2_CHAIN_FORCECOW          0x00000008      /* force copy-on-wr */
 238 #define HAMMER2_CHAIN_DELETED           0x00000010      /* deleted chain */
 239 #define HAMMER2_CHAIN_INITIAL           0x00000020      /* initial create */
 240 #define HAMMER2_CHAIN_FLUSHED           0x00000040      /* blktable updated */
 241 #define HAMMER2_CHAIN_MOVED             0x00000080      /* bref changed */
 242 #define HAMMER2_CHAIN_IOFLUSH           0x00000100      /* bawrite on put */
 243 #define HAMMER2_CHAIN_DEFERRED          0x00000200      /* on a deferral list */
 244 #define HAMMER2_CHAIN_DESTROYED         0x00000400      /* destroying inode */
 245 #define HAMMER2_CHAIN_VOLUMESYNC        0x00000800      /* needs volume sync */
 246 #define HAMMER2_CHAIN_UNUSED01000       0x00001000
 247 #define HAMMER2_CHAIN_MOUNTED           0x00002000      /* PFS is mounted */
 248 #define HAMMER2_CHAIN_ONRBTREE          0x00004000      /* on parent RB tree */
 249 #define HAMMER2_CHAIN_SNAPSHOT          0x00008000      /* snapshot special */
 250 #define HAMMER2_CHAIN_EMBEDDED          0x00010000      /* embedded data */
 251 #define HAMMER2_CHAIN_RELEASE           0x00020000      /* don't keep around */
 252 #define HAMMER2_CHAIN_UNUSED40000       0x00040000
 253 #define HAMMER2_CHAIN_UNUSED80000       0x00080000
 254 #define HAMMER2_CHAIN_DUPLICATED        0x00100000      /* fwd delete-dup */
 255 #define HAMMER2_CHAIN_PFSROOT           0x00200000      /* in pfs->cluster */
 256
 257 /*
 258  * Flags passed to hammer2_chain_lookup() and hammer2_chain_next()
 259  *
 260  * NOTE: MATCHIND allows an indirect block / freemap node to be returned
 261  *       when the passed key range matches the radix.  Remember that key_end
 262  *       is inclusive (e.g. {0x000,0xFFF}, not {0x000,0x1000}).
 263  */
 264 #define HAMMER2_LOOKUP_NOLOCK           0x00000001      /* ref only */
 265 #define HAMMER2_LOOKUP_NODATA           0x00000002      /* data left NULL */
 266 #define HAMMER2_LOOKUP_SHARED           0x00000100
 267 #define HAMMER2_LOOKUP_MATCHIND         0x00000200      /* return all chains */
 268 #define HAMMER2_LOOKUP_FREEMAP          0x00000400      /* freemap base */
 269 #define HAMMER2_LOOKUP_ALWAYS           0x00000800      /* resolve data */
 270
 271 /*
 272  * Flags passed to hammer2_chain_modify() and hammer2_chain_resize()
 273  *
 274  * NOTE: OPTDATA allows us to avoid instantiating buffers for INDIRECT
 275  *       blocks in the INITIAL-create state.
 276  */
 277 #define HAMMER2_MODIFY_OPTDATA          0x00000002      /* data can be NULL */
 278 #define HAMMER2_MODIFY_NO_MODIFY_TID    0x00000004
 279 #define HAMMER2_MODIFY_ASSERTNOCOPY     0x00000008      /* assert no del-dup */
 280 #define HAMMER2_MODIFY_NOREALLOC        0x00000010
 281 #define HAMMER2_MODIFY_INPLACE          0x00000020      /* don't del-dup */
 282
 283 /*
 284  * Flags passed to hammer2_chain_lock()
 285  */
 286 #define HAMMER2_RESOLVE_NEVER           1
 287 #define HAMMER2_RESOLVE_MAYBE           2
 288 #define HAMMER2_RESOLVE_ALWAYS          3
 289 #define HAMMER2_RESOLVE_MASK            0x0F
 290
 291 #define HAMMER2_RESOLVE_SHARED          0x10    /* request shared lock */
 292 #define HAMMER2_RESOLVE_NOREF           0x20    /* already ref'd on lock */
 293
 294 /*
 295  * Flags passed to hammer2_chain_delete()
 296  */
 297 #define HAMMER2_DELETE_WILLDUP          0x0001  /* no blk free, will be dup */
 298
 299 /*
 300  * Flags passed to hammer2_chain_delete_duplicate()
 301  */
 302 #define HAMMER2_DELDUP_RECORE           0x0001
 303
 304 /*
 305  * Cluster different types of storage together for allocations
 306  */
 307 #define HAMMER2_FREECACHE_INODE         0
 308 #define HAMMER2_FREECACHE_INDIR         1
 309 #define HAMMER2_FREECACHE_DATA          2
 310 #define HAMMER2_FREECACHE_UNUSED3       3
 311 #define HAMMER2_FREECACHE_TYPES         4
 312
 313 /*
 314  * hammer2_freemap_alloc() block preference
 315  */
 316 #define HAMMER2_OFF_NOPREF              ((hammer2_off_t)-1)
 317
 318 /*
 319  * BMAP read-ahead maximum parameters
 320  */
 321 #define HAMMER2_BMAP_COUNT              16      /* max bmap read-ahead */
 322 #define HAMMER2_BMAP_BYTES              (HAMMER2_PBUFSIZE * HAMMER2_BMAP_COUNT)
 323
 324 /*
 325  * Misc
 326  */
 327 #define HAMMER2_FLUSH_DEPTH_LIMIT       10      /* stack recursion limit */
 328
 329 /*
 330  * hammer2_freemap_adjust()
 331  */
 332 #define HAMMER2_FREEMAP_DORECOVER       1
 333 #define HAMMER2_FREEMAP_DOMAYFREE       2
 334 #define HAMMER2_FREEMAP_DOREALFREE      3
 335
 336 /*
 337  * HAMMER2 IN-MEMORY CACHE OF MEDIA STRUCTURES
 338  *
 339  * There is an in-memory representation of all on-media data structure.
 340  * Basically everything is represented by a hammer2_chain structure
 341  * in-memory and other higher-level structures map to chains.
 342  *
 343  * A great deal of data is accessed simply via its buffer cache buffer,
 344  * which is mapped for the duration of the chain's lock.  However, because
 345  * chains may represent blocks smaller than the 16KB minimum we impose
 346  * on buffer cache buffers, we cannot hold related buffer cache buffers
 347  * locked for smaller blocks.  In these situations we kmalloc() a copy
 348  * of the block.
 349  *
 350  * When modifications are made to a chain a new filesystem block must be
 351  * allocated.  Multiple modifications do not necessarily allocate new
 352  * blocks.  However, when a flush occurs a flush synchronization point
 353  * is created and any new modifications made after this point will allocate
 354  * a new block even if the chain is already in a modified state.
 355  *
 356  * The in-memory representation may remain cached (for example in order to
 357  * placemark clustering locks) even after the related data has been
 358  * detached.
 359  *
 360  *                              CORE SHARING
 361  *
 362  * In order to support concurrent flushes a flush synchronization point
 363  * is created represented by a transaction id.  Among other things,
 364  * operations may move filesystem objects from one part of the topology
 365  * to another (for example, if you rename a file or when indirect blocks
 366  * are created or destroyed, and a few other things).  When this occurs
 367  * across a flush synchronization point the flusher needs to be able to
 368  * recurse down BOTH the 'before' version of the topology and the 'after'
 369  * version.
 370  *
 371  * To facilitate this modifications to chains do what is called a
 372  * DELETE-DUPLICATE operation.  Chains are not actually moved in-memory.
 373  * Instead the chain we wish to move is deleted and a new chain is created
 374  * at the target location in the topology.  ANY SUBCHAINS PLACED UNDER THE
 375  * CHAIN BEING MOVED HAVE TO EXIST IN BOTH PLACES.  To make this work
 376  * all sub-chains are managed by the hammer2_chain_core structure.  This
 377  * structure can be multi-homed, meaning that it can have more than one
 378  * chain as its parent.  When a chain is delete-duplicated the chain's core
 379  * becomes shared under both the old and new chain.
 380  *
 381  *                              STALE CHAINS
 382  *
 383  * When a chain is delete-duplicated the old chain typically becomes stale.
 384  * This is detected via the HAMMER2_CHAIN_DUPLICATED flag in chain->flags.
 385  * To avoid executing live filesystem operations on stale chains, the inode
 386  * locking code will follow stale chains via core->ownerq until it finds
 387  * the live chain.  The lock prevents ripups by other threads.  Lookups
 388  * must properly order locking operations to prevent other threads from
 389  * racing the lookup operation and will also follow stale chains when
 390  * required.
 391  */
 392
 393 RB_HEAD(hammer2_inode_tree, hammer2_inode);
 394
 395 /*
 396  * A hammer2 inode.
 397  *
 398  * NOTE: The inode's attribute CST which is also used to lock the inode
 399  *       is embedded in the chain (chain.cst) and aliased w/ attr_cst.
 400  */
 401 struct hammer2_inode {
 402         RB_ENTRY(hammer2_inode) rbnode;         /* inumber lookup (HL) */
 403         ccms_cst_t              topo_cst;       /* directory topology cst */
 404         struct hammer2_pfsmount *pmp;           /* PFS mount */
 405         struct hammer2_inode    *pip;           /* parent inode */
 406         struct vnode            *vp;
 407         hammer2_chain_t         *chain;         /* NOTE: rehomed on rename */
 408         struct lockf            advlock;
 409         hammer2_tid_t           inum;
 410         u_int                   flags;
 411         u_int                   refs;           /* +vpref, +flushref */
 412         uint8_t                 comp_heuristic;
 413         hammer2_off_t           size;
 414         uint64_t                mtime;
 415 };
 416
 417 typedef struct hammer2_inode hammer2_inode_t;
 418
 419 #define HAMMER2_INODE_MODIFIED          0x0001
 420 #define HAMMER2_INODE_SROOT             0x0002  /* kmalloc special case */
 421 #define HAMMER2_INODE_RENAME_INPROG     0x0004
 422 #define HAMMER2_INODE_ONRBTREE          0x0008
 423 #define HAMMER2_INODE_RESIZED           0x0010
 424 #define HAMMER2_INODE_MTIME             0x0020
 425
 426 int hammer2_inode_cmp(hammer2_inode_t *ip1, hammer2_inode_t *ip2);
 427 RB_PROTOTYPE2(hammer2_inode_tree, hammer2_inode, rbnode, hammer2_inode_cmp,
 428                 hammer2_tid_t);
 429
 430 /*
 431  * A hammer2 transaction and flush sequencing structure.
 432  *
 433  * This global structure is tied into hammer2_mount and is used
 434  * to sequence modifying operations and flushes.
 435  *
 436  * (a) Any modifying operations with sync_tid >= flush_tid will stall until
 437  *     all modifying operating with sync_tid < flush_tid complete.
 438  *
 439  *     The flush related to flush_tid stalls until all modifying operations
 440  *     with sync_tid < flush_tid complete.
 441  *
 442  * (b) Once unstalled, modifying operations with sync_tid > flush_tid are
 443  *     allowed to run.  All modifications cause modify/duplicate operations
 444  *     to occur on the related chains.  Note that most INDIRECT blocks will
 445  *     be unaffected because the modifications just overload the RBTREE
 446  *     structurally instead of actually modifying the indirect blocks.
 447  *
 448  * (c) The actual flush unstalls and RUNS CONCURRENTLY with (b), but only
 449  *     utilizes the chain structures with sync_tid <= flush_tid.  The
 450  *     flush will modify related indirect blocks and inodes in-place
 451  *     (rather than duplicate) since the adjustments are compatible with
 452  *     (b)'s RBTREE overloading
 453  *
 454  *     SPECIAL NOTE:  Inode modifications have to also propagate along any
 455  *                    modify/duplicate chains.  File writes detect the flush
 456  *                    and force out the conflicting buffer cache buffer(s)
 457  *                    before reusing them.
 458  *
 459  * (d) Snapshots can be made instantly but must be flushed and disconnected
 460  *     from their duplicative source before they can be mounted.  This is
 461  *     because while H2's on-media structure supports forks, its in-memory
 462  *     structure only supports very simple forking for background flushing
 463  *     purposes.
 464  *
 465  * TODO: Flush merging.  When fsync() is called on multiple discrete files
 466  *       concurrently there is no reason to stall the second fsync.
 467  *       The final flush that reaches to root can cover both fsync()s.
 468  *
 469  *     The chains typically terminate as they fly onto the disk.  The flush
 470  *     ultimately reaches the volume header.
 471  */
 472 struct hammer2_trans {
 473         TAILQ_ENTRY(hammer2_trans) entry;
 474         struct hammer2_pfsmount *pmp;           /* might be NULL */
 475         struct hammer2_mount    *hmp_single;    /* if single-targetted */
 476         hammer2_tid_t           sync_tid;
 477         hammer2_tid_t           real_tid;
 478         hammer2_tid_t           inode_tid;
 479         thread_t                td;             /* pointer */
 480         int                     flags;
 481         int                     blocked;
 482         uint8_t                 inodes_created;
 483         uint8_t                 dummy[7];
 484 };
 485
 486 typedef struct hammer2_trans hammer2_trans_t;
 487
 488 #define HAMMER2_TRANS_ISFLUSH           0x0001  /* formal flush */
 489 #define HAMMER2_TRANS_UNUSED0002        0x0002
 490 #define HAMMER2_TRANS_BUFCACHE          0x0004  /* from bioq strategy write */
 491 #define HAMMER2_TRANS_NEWINODE          0x0008  /* caller allocating inode */
 492 #define HAMMER2_TRANS_ISALLOCATING      0x0010  /* in allocator */
 493
 494 #define HAMMER2_FREEMAP_HEUR_NRADIX     4       /* pwr 2 PBUFRADIX-MINIORADIX */
 495 #define HAMMER2_FREEMAP_HEUR_TYPES      8
 496 #define HAMMER2_FREEMAP_HEUR            (HAMMER2_FREEMAP_HEUR_NRADIX * \
 497                                          HAMMER2_FREEMAP_HEUR_TYPES)
 498
 499 /*
 500  * Global (per device) mount structure for device (aka vp->v_mount->hmp)
 501  */
 502 TAILQ_HEAD(hammer2_trans_queue, hammer2_trans);
 503
 504 struct hammer2_mount {
 505         struct vnode    *devvp;         /* device vnode */
 506         int             ronly;          /* read-only mount */
 507         int             pmp_count;      /* PFS mounts backed by us */
 508         TAILQ_ENTRY(hammer2_mount) mntentry; /* hammer2_mntlist */
 509
 510         struct malloc_type *mchain;
 511         int             nipstacks;
 512         int             maxipstacks;
 513         struct spinlock io_spin;        /* iotree access */
 514         struct hammer2_io_tree iotree;
 515         int             iofree_count;
 516         hammer2_chain_t vchain;         /* anchor chain (topology) */
 517         hammer2_chain_t fchain;         /* anchor chain (freemap) */
 518         hammer2_inode_t *sroot;         /* super-root localized to media */
 519         struct lock     alloclk;        /* lockmgr lock */
 520         struct lock     voldatalk;      /* lockmgr lock */
 521         struct hammer2_trans_queue transq; /* all in-progress transactions */
 522         hammer2_off_t   heur_freemap[HAMMER2_FREEMAP_HEUR];
 523         int             flushcnt;       /* #of flush trans on the list */
 524
 525         int             volhdrno;       /* last volhdrno written */
 526         hammer2_volume_data_t voldata;
 527         hammer2_volume_data_t volsync;  /* synchronized voldata */
 528 };
 529
 530 typedef struct hammer2_mount hammer2_mount_t;
 531
 532 /*
 533  * HAMMER2 cluster - a device/root associated with a PFS.
 534  *
 535  * A PFS may have several hammer2_cluster's associated with it.
 536  */
 537 #define HAMMER2_MAXCLUSTER      8
 538
 539 struct hammer2_cluster {
 540         int                     nchains;
 541         int                     status;
 542         hammer2_chain_t         *chains[HAMMER2_MAXCLUSTER];
 543 };
 544
 545 typedef struct hammer2_cluster hammer2_cluster_t;
 546
 547 /*
 548  * HAMMER2 PFS mount point structure (aka vp->v_mount->mnt_data).
 549  * This has a 1:1 correspondence to struct mount (note that the
 550  * hammer2_mount structure has a N:1 correspondence).
 551  *
 552  * This structure represents a cluster mount and not necessarily a
 553  * PFS under a specific device mount (HMP).  The distinction is important
 554  * because the elements backing a cluster mount can change on the fly.
 555  *
 556  * Usually the first element under the cluster represents the original
 557  * user-requested mount that bootstraps the whole mess.  In significant
 558  * setups the original is usually just a read-only media image (or
 559  * representitive file) that simply contains a bootstrap volume header
 560  * listing the configuration.
 561  */
 562 struct hammer2_pfsmount {
 563         struct mount            *mp;
 564         hammer2_cluster_t       cluster;
 565         hammer2_inode_t         *iroot;         /* PFS root inode */
 566         hammer2_off_t           inode_count;    /* copy of inode_count */
 567         ccms_domain_t           ccms_dom;
 568         struct netexport        export;         /* nfs export */
 569         int                     ronly;          /* read-only mount */
 570         struct malloc_type      *minode;
 571         struct malloc_type      *mmsg;
 572         kdmsg_iocom_t           iocom;
 573         struct spinlock         inum_spin;      /* inumber lookup */
 574         struct hammer2_inode_tree inum_tree;
 575         long                    inmem_inodes;
 576         long                    inmem_chains;
 577         int                     inmem_waiting;
 578         int                     count_lwinprog; /* logical write in prog */
 579         thread_t                wthread_td;     /* write thread td */
 580         struct bio_queue_head   wthread_bioq;   /* logical buffer bioq */
 581         struct mtx              wthread_mtx;    /* interlock */
 582         int                     wthread_destroy;/* termination sequencing */
 583 };
 584
 585 typedef struct hammer2_pfsmount hammer2_pfsmount_t;
 586
 587 #define HAMMER2_LWINPROG_WAITING        0x80000000
 588 #define HAMMER2_LWINPROG_MASK           0x7FFFFFFF
 589
 590 #if defined(_KERNEL)
 591
 592 MALLOC_DECLARE(M_HAMMER2);
 593
 594 #define VTOI(vp)        ((hammer2_inode_t *)(vp)->v_data)
 595 #define ITOV(ip)        ((ip)->vp)
 596
 597 /*
 598  * Currently locked chains retain the locked buffer cache buffer for
 599  * indirect blocks, and indirect blocks can be one of two sizes.  The
 600  * device buffer has to match the case to avoid deadlocking recursive
 601  * chains that might otherwise try to access different offsets within
 602  * the same device buffer.
 603  */
 604 static __inline
 605 int
 606 hammer2_devblkradix(int radix)
 607 {
 608         if (radix <= HAMMER2_LBUFRADIX) {
 609                 return (HAMMER2_LBUFRADIX);
 610         } else {
 611                 return (HAMMER2_PBUFRADIX);
 612         }
 613 }
 614
 615 static __inline
 616 size_t
 617 hammer2_devblksize(size_t bytes)
 618 {
 619         if (bytes <= HAMMER2_LBUFSIZE) {
 620                 return(HAMMER2_LBUFSIZE);
 621         } else {
 622                 KKASSERT(bytes <= HAMMER2_PBUFSIZE &&
 623                          (bytes ^ (bytes - 1)) == ((bytes << 1) - 1));
 624                 return (HAMMER2_PBUFSIZE);
 625         }
 626 }
 627
 628
 629 static __inline
 630 hammer2_pfsmount_t *
 631 MPTOPMP(struct mount *mp)
 632 {
 633         return ((hammer2_pfsmount_t *)mp->mnt_data);
 634 }
 635
 636 extern struct vop_ops hammer2_vnode_vops;
 637 extern struct vop_ops hammer2_spec_vops;
 638 extern struct vop_ops hammer2_fifo_vops;
 639
 640 extern int hammer2_debug;
 641 extern int hammer2_cluster_enable;
 642 extern int hammer2_hardlink_enable;
 643 extern int hammer2_flush_pipe;
 644 extern long hammer2_iod_file_read;
 645 extern long hammer2_iod_meta_read;
 646 extern long hammer2_iod_indr_read;
 647 extern long hammer2_iod_fmap_read;
 648 extern long hammer2_iod_volu_read;
 649 extern long hammer2_iod_file_write;
 650 extern long hammer2_iod_meta_write;
 651 extern long hammer2_iod_indr_write;
 652 extern long hammer2_iod_fmap_write;
 653 extern long hammer2_iod_volu_write;
 654 extern long hammer2_ioa_file_read;
 655 extern long hammer2_ioa_meta_read;
 656 extern long hammer2_ioa_indr_read;
 657 extern long hammer2_ioa_fmap_read;
 658 extern long hammer2_ioa_volu_read;
 659 extern long hammer2_ioa_file_write;
 660 extern long hammer2_ioa_meta_write;
 661 extern long hammer2_ioa_indr_write;
 662 extern long hammer2_ioa_fmap_write;
 663 extern long hammer2_ioa_volu_write;
 664
 665 extern struct objcache *cache_buffer_read;
 666 extern struct objcache *cache_buffer_write;
 667
 668 extern int destroy;
 669 extern int write_thread_wakeup;
 670
 671 extern mtx_t thread_protect;
 672
 673 /*
 674  * hammer2_subr.c
 675  */
 676 #define hammer2_icrc32(buf, size)       iscsi_crc32((buf), (size))
 677 #define hammer2_icrc32c(buf, size, crc) iscsi_crc32_ext((buf), (size), (crc))
 678
 679 hammer2_chain_t *hammer2_inode_lock_ex(hammer2_inode_t *ip);
 680 hammer2_chain_t *hammer2_inode_lock_sh(hammer2_inode_t *ip);
 681 void hammer2_inode_unlock_ex(hammer2_inode_t *ip, hammer2_chain_t *chain);
 682 void hammer2_inode_unlock_sh(hammer2_inode_t *ip, hammer2_chain_t *chain);
 683 void hammer2_chain_refactor(hammer2_chain_t **chainp);
 684 void hammer2_voldata_lock(hammer2_mount_t *hmp);
 685 void hammer2_voldata_unlock(hammer2_mount_t *hmp, int modify);
 686 ccms_state_t hammer2_inode_lock_temp_release(hammer2_inode_t *ip);
 687 void hammer2_inode_lock_temp_restore(hammer2_inode_t *ip, ccms_state_t ostate);
 688 ccms_state_t hammer2_inode_lock_upgrade(hammer2_inode_t *ip);
 689 void hammer2_inode_lock_downgrade(hammer2_inode_t *ip, ccms_state_t ostate);
 690
 691 void hammer2_mount_exlock(hammer2_mount_t *hmp);
 692 void hammer2_mount_shlock(hammer2_mount_t *hmp);
 693 void hammer2_mount_unlock(hammer2_mount_t *hmp);
 694
 695 int hammer2_get_dtype(hammer2_chain_t *chain);
 696 int hammer2_get_vtype(hammer2_chain_t *chain);
 697 u_int8_t hammer2_get_obj_type(enum vtype vtype);
 698 void hammer2_time_to_timespec(u_int64_t xtime, struct timespec *ts);
 699 u_int64_t hammer2_timespec_to_time(struct timespec *ts);
 700 u_int32_t hammer2_to_unix_xid(uuid_t *uuid);
 701 void hammer2_guid_to_uuid(uuid_t *uuid, u_int32_t guid);
 702
 703 hammer2_key_t hammer2_dirhash(const unsigned char *name, size_t len);
 704 int hammer2_getradix(size_t bytes);
 705
 706 int hammer2_calc_logical(hammer2_inode_t *ip, hammer2_off_t uoff,
 707                         hammer2_key_t *lbasep, hammer2_key_t *leofp);
 708 int hammer2_calc_physical(hammer2_inode_t *ip, hammer2_key_t lbase);
 709 void hammer2_update_time(uint64_t *timep);
 710
 711 /*
 712  * hammer2_inode.c
 713  */
 714 struct vnode *hammer2_igetv(hammer2_inode_t *ip, int *errorp);
 715
 716 void hammer2_inode_lock_nlinks(hammer2_inode_t *ip);
 717 void hammer2_inode_unlock_nlinks(hammer2_inode_t *ip);
 718 hammer2_inode_t *hammer2_inode_lookup(hammer2_pfsmount_t *pmp,
 719                         hammer2_tid_t inum);
 720 hammer2_inode_t *hammer2_inode_get(hammer2_pfsmount_t *pmp,
 721                         hammer2_inode_t *dip, hammer2_chain_t *chain);
 722 void hammer2_inode_free(hammer2_inode_t *ip);
 723 void hammer2_inode_ref(hammer2_inode_t *ip);
 724 void hammer2_inode_drop(hammer2_inode_t *ip);
 725 void hammer2_inode_repoint(hammer2_inode_t *ip, hammer2_inode_t *pip,
 726                         hammer2_chain_t *chain);
 727
 728 hammer2_inode_t *hammer2_inode_create(hammer2_trans_t *trans,
 729                         hammer2_inode_t *dip,
 730                         struct vattr *vap, struct ucred *cred,
 731                         const uint8_t *name, size_t name_len,
 732                         hammer2_chain_t **chainp, int *errorp);
 733 int hammer2_inode_connect(hammer2_trans_t *trans, int hlink,
 734                         hammer2_inode_t *dip, hammer2_chain_t **chainp,
 735                         const uint8_t *name, size_t name_len);
 736 hammer2_inode_t *hammer2_inode_common_parent(hammer2_inode_t *fdip,
 737                         hammer2_inode_t *tdip);
 738 void hammer2_inode_fsync(hammer2_trans_t *trans, hammer2_inode_t *ip,
 739                         hammer2_chain_t **parentp);
 740 int hammer2_unlink_file(hammer2_trans_t *trans, hammer2_inode_t *dip,
 741                         const uint8_t *name, size_t name_len, int isdir,
 742                         int *hlinkp);
 743 int hammer2_hardlink_consolidate(hammer2_trans_t *trans, hammer2_inode_t *ip,
 744                         hammer2_chain_t **chainp,
 745                         hammer2_inode_t *tdip, int linkcnt);
 746 int hammer2_hardlink_deconsolidate(hammer2_trans_t *trans, hammer2_inode_t *dip,
 747                         hammer2_chain_t **chainp, hammer2_chain_t **ochainp);
 748 int hammer2_hardlink_find(hammer2_inode_t *dip,
 749                         hammer2_chain_t **chainp, hammer2_chain_t **ochainp);
 750
 751 /*
 752  * hammer2_chain.c
 753  */
 754 void hammer2_modify_volume(hammer2_mount_t *hmp);
 755 hammer2_chain_t *hammer2_chain_alloc(hammer2_mount_t *hmp,
 756                                 hammer2_pfsmount_t *pmp,
 757                                 hammer2_trans_t *trans,
 758                                 hammer2_blockref_t *bref);
 759 void hammer2_chain_core_alloc(hammer2_trans_t *trans, hammer2_chain_t *nchain,
 760                                 hammer2_chain_t *ochain);
 761 void hammer2_chain_ref(hammer2_chain_t *chain);
 762 void hammer2_chain_drop(hammer2_chain_t *chain);
 763 int hammer2_chain_lock(hammer2_chain_t *chain, int how);
 764 void hammer2_chain_load_async(hammer2_chain_t *chain,
 765                                 void (*func)(hammer2_io_t *dio,
 766                                              hammer2_chain_t *chain,
 767                                              void *arg_p, off_t arg_o),
 768                                 void *arg_p, off_t arg_o);
 769 void hammer2_chain_moved(hammer2_chain_t *chain);
 770 void hammer2_chain_modify(hammer2_trans_t *trans,
 771                                 hammer2_chain_t **chainp, int flags);
 772 hammer2_inode_data_t *hammer2_chain_modify_ip(hammer2_trans_t *trans,
 773                                 hammer2_inode_t *ip, hammer2_chain_t **chainp,
 774                                 int flags);
 775 void hammer2_chain_resize(hammer2_trans_t *trans, hammer2_inode_t *ip,
 776                                 hammer2_chain_t *parent,
 777                                 hammer2_chain_t **chainp,
 778                                 int nradix, int flags);
 779 void hammer2_chain_unlock(hammer2_chain_t *chain);
 780 void hammer2_chain_wait(hammer2_chain_t *chain);
 781 hammer2_chain_t *hammer2_chain_get(hammer2_chain_t *parent,
 782                                 hammer2_blockref_t *bref, int generation);
 783 hammer2_chain_t *hammer2_chain_lookup_init(hammer2_chain_t *parent, int flags);
 784 void hammer2_chain_lookup_done(hammer2_chain_t *parent);
 785 hammer2_chain_t *hammer2_chain_lookup(hammer2_chain_t **parentp,
 786                                 hammer2_key_t *key_nextp,
 787                                 hammer2_key_t key_beg, hammer2_key_t key_end,
 788                                 int *cache_indexp, int flags);
 789 hammer2_chain_t *hammer2_chain_next(hammer2_chain_t **parentp,
 790                                 hammer2_chain_t *chain,
 791                                 hammer2_key_t *key_nextp,
 792                                 hammer2_key_t key_beg, hammer2_key_t key_end,
 793                                 int *cache_indexp, int flags);
 794 hammer2_chain_t *hammer2_chain_scan(hammer2_chain_t *parent,
 795                                 hammer2_chain_t *chain,
 796                                 int *cache_indexp, int flags);
 797
 798 int hammer2_chain_create(hammer2_trans_t *trans,
 799                                 hammer2_chain_t **parentp,
 800                                 hammer2_chain_t **chainp,
 801                                 hammer2_key_t key, int keybits,
 802                                 int type, size_t bytes);
 803 void hammer2_chain_duplicate(hammer2_trans_t *trans, hammer2_chain_t **parentp,
 804                                 hammer2_chain_t **chainp,
 805                                 hammer2_blockref_t *bref, int snapshot,
 806                                 int duplicate_reason);
 807 int hammer2_chain_snapshot(hammer2_trans_t *trans, hammer2_chain_t **chainp,
 808                                 hammer2_ioc_pfs_t *pfs);
 809 void hammer2_chain_delete(hammer2_trans_t *trans, hammer2_chain_t *chain,
 810                                 int flags);
 811 void hammer2_chain_delete_duplicate(hammer2_trans_t *trans,
 812                                 hammer2_chain_t **chainp, int flags);
 813 void hammer2_chain_flush(hammer2_trans_t *trans, hammer2_chain_t **chainp);
 814 void hammer2_chain_commit(hammer2_trans_t *trans, hammer2_chain_t *chain);
 815 void hammer2_chain_setsubmod(hammer2_trans_t *trans, hammer2_chain_t *chain);
 816
 817 void hammer2_chain_memory_wait(hammer2_pfsmount_t *pmp);
 818 void hammer2_chain_memory_wakeup(hammer2_pfsmount_t *pmp);
 819 void hammer2_chain_countbrefs(hammer2_chain_t *chain,
 820                                 hammer2_blockref_t *base, int count);
 821 void hammer2_chain_layer_check_locked(hammer2_mount_t *hmp,
 822                                 hammer2_chain_core_t *core);
 823
 824 int hammer2_base_find(hammer2_chain_t *chain,
 825                                 hammer2_blockref_t *base, int count,
 826                                 int *cache_indexp, hammer2_key_t *key_nextp,
 827                                 hammer2_key_t key_beg, hammer2_key_t key_end);
 828 void hammer2_base_delete(hammer2_trans_t *trans, hammer2_chain_t *chain,
 829                                 hammer2_blockref_t *base, int count,
 830                                 int *cache_indexp, hammer2_chain_t *child);
 831 void hammer2_base_insert(hammer2_trans_t *trans, hammer2_chain_t *chain,
 832                                 hammer2_blockref_t *base, int count,
 833                                 int *cache_indexp, hammer2_chain_t *child);
 834
 835 /*
 836  * hammer2_trans.c
 837  */
 838 void hammer2_trans_init(hammer2_trans_t *trans, hammer2_pfsmount_t *pmp,
 839                                 hammer2_mount_t *hmp, int flags);
 840 void hammer2_trans_clear_invfsync(hammer2_trans_t *trans);
 841 void hammer2_trans_done(hammer2_trans_t *trans);
 842
 843 /*
 844  * hammer2_ioctl.c
 845  */
 846 int hammer2_ioctl(hammer2_inode_t *ip, u_long com, void *data,
 847                                 int fflag, struct ucred *cred);
 848
 849 /*
 850  * hammer2_io.c
 851  */
 852 hammer2_io_t *hammer2_io_getblk(hammer2_mount_t *hmp, off_t lbase,
 853                                 int lsize, int *ownerp);
 854 void hammer2_io_putblk(hammer2_io_t **diop);
 855 void hammer2_io_cleanup(hammer2_mount_t *hmp, struct hammer2_io_tree *tree);
 856 char *hammer2_io_data(hammer2_io_t *dio, off_t lbase);
 857 int hammer2_io_new(hammer2_mount_t *hmp, off_t lbase, int lsize,
 858                                 hammer2_io_t **diop);
 859 int hammer2_io_newnz(hammer2_mount_t *hmp, off_t lbase, int lsize,
 860                                 hammer2_io_t **diop);
 861 int hammer2_io_newq(hammer2_mount_t *hmp, off_t lbase, int lsize,
 862                                 hammer2_io_t **diop);
 863 int hammer2_io_bread(hammer2_mount_t *hmp, off_t lbase, int lsize,
 864                                 hammer2_io_t **diop);
 865 void hammer2_io_breadcb(hammer2_mount_t *hmp, off_t lbase, int lsize,
 866                                 void (*callback)(hammer2_io_t *dio,
 867                                                  hammer2_chain_t *arg_c,
 868                                                  void *arg_p, off_t arg_o),
 869                                 hammer2_chain_t *arg_c,
 870                                 void *arg_p, off_t arg_o);
 871 void hammer2_io_bawrite(hammer2_io_t **diop);
 872 void hammer2_io_bdwrite(hammer2_io_t **diop);
 873 int hammer2_io_bwrite(hammer2_io_t **diop);
 874 void hammer2_io_setdirty(hammer2_io_t *dio);
 875 void hammer2_io_setinval(hammer2_io_t *dio, u_int bytes);
 876 void hammer2_io_brelse(hammer2_io_t **diop);
 877 void hammer2_io_bqrelse(hammer2_io_t **diop);
 878 int hammer2_io_isdirty(hammer2_io_t *dio);
 879
 880 /*
 881  * hammer2_msgops.c
 882  */
 883 int hammer2_msg_dbg_rcvmsg(kdmsg_msg_t *msg);
 884 int hammer2_msg_adhoc_input(kdmsg_msg_t *msg);
 885
 886 /*
 887  * hammer2_vfsops.c
 888  */
 889 void hammer2_clusterctl_wakeup(kdmsg_iocom_t *iocom);
 890 void hammer2_volconf_update(hammer2_pfsmount_t *pmp, int index);
 891 void hammer2_cluster_reconnect(hammer2_pfsmount_t *pmp, struct file *fp);
 892 void hammer2_dump_chain(hammer2_chain_t *chain, int tab, int *countp);
 893 void hammer2_bioq_sync(hammer2_pfsmount_t *pmp);
 894 int hammer2_vfs_sync(struct mount *mp, int waitflags);
 895 void hammer2_lwinprog_ref(hammer2_pfsmount_t *pmp);
 896 void hammer2_lwinprog_drop(hammer2_pfsmount_t *pmp);
 897 void hammer2_lwinprog_wait(hammer2_pfsmount_t *pmp);
 898
 899 /*
 900  * hammer2_freemap.c
 901  */
 902 int hammer2_freemap_alloc(hammer2_trans_t *trans, hammer2_chain_t *chain,
 903                                 size_t bytes);
 904 void hammer2_freemap_adjust(hammer2_trans_t *trans, hammer2_mount_t *hmp,
 905                                 hammer2_blockref_t *bref, int how);
 906
 907
 908 #endif /* !_KERNEL */
 909 #endif /* !_VFS_HAMMER2_HAMMER2_H_ */