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 struct hammer2_chain {
 158         RB_ENTRY(hammer2_chain) rbnode;         /* node */
 159         TAILQ_ENTRY(hammer2_chain) core_entry;  /* contemporary chains */
 160         hammer2_chain_layer_t   *inlayer;
 161         hammer2_blockref_t      bref;
 162         hammer2_chain_core_t    *core;
 163         hammer2_chain_core_t    *above;
 164         struct hammer2_state    *state;         /* if active cache msg */
 165         struct hammer2_mount    *hmp;
 166         struct hammer2_pfsmount *pmp;           /* can be NULL */
 167
 168         hammer2_tid_t   modify_tid;             /* snapshot/flush filter */
 169         hammer2_tid_t   delete_tid;
 170         hammer2_key_t   data_count;             /* delta's to apply */
 171         hammer2_key_t   inode_count;            /* delta's to apply */
 172         struct buf      *bp;                    /* physical data buffer */
 173         u_int           bytes;                  /* physical data size */
 174         u_int           flags;
 175         u_int           refs;
 176         u_int           lockcnt;
 177         int             debug_reason;
 178         hammer2_media_data_t *data;             /* data pointer shortcut */
 179         TAILQ_ENTRY(hammer2_chain) flush_node;  /* flush deferral list */
 180 };
 181
 182 typedef struct hammer2_chain hammer2_chain_t;
 183
 184 int hammer2_chain_cmp(hammer2_chain_t *chain1, hammer2_chain_t *chain2);
 185 RB_PROTOTYPE(hammer2_chain_tree, hammer2_chain, rbnode, hammer2_chain_cmp);
 186
 187 /*
 188  * Special notes on flags:
 189  *
 190  * INITIAL - This flag allows a chain to be created and for storage to
 191  *           be allocated without having to immediately instantiate the
 192  *           related buffer.  The data is assumed to be all-zeros.  It
 193  *           is primarily used for indirect blocks.
 194  *
 195  * MOVED   - A modified chain becomes MOVED after it flushes.  A chain
 196  *           can also become MOVED if it is moved within the topology
 197  *           (even if not modified).
 198  */
 199 #define HAMMER2_CHAIN_MODIFIED          0x00000001      /* dirty chain data */
 200 #define HAMMER2_CHAIN_ALLOCATED         0x00000002      /* kmalloc'd chain */
 201 #define HAMMER2_CHAIN_DIRTYBP           0x00000004      /* dirty on unlock */
 202 #define HAMMER2_CHAIN_FORCECOW          0x00000008      /* force copy-on-wr */
 203 #define HAMMER2_CHAIN_DELETED           0x00000010      /* deleted chain */
 204 #define HAMMER2_CHAIN_INITIAL           0x00000020      /* initial create */
 205 #define HAMMER2_CHAIN_FLUSHED           0x00000040      /* flush on unlock */
 206 #define HAMMER2_CHAIN_MOVED             0x00000080      /* bref changed */
 207 #define HAMMER2_CHAIN_IOFLUSH           0x00000100      /* bawrite on put */
 208 #define HAMMER2_CHAIN_DEFERRED          0x00000200      /* on a deferral list */
 209 #define HAMMER2_CHAIN_DESTROYED         0x00000400      /* destroying inode */
 210 #define HAMMER2_CHAIN_VOLUMESYNC        0x00000800      /* needs volume sync */
 211 #define HAMMER2_CHAIN_UNUSED01000       0x00001000
 212 #define HAMMER2_CHAIN_MOUNTED           0x00002000      /* PFS is mounted */
 213 #define HAMMER2_CHAIN_ONRBTREE          0x00004000      /* on parent RB tree */
 214 #define HAMMER2_CHAIN_SNAPSHOT          0x00008000      /* snapshot special */
 215 #define HAMMER2_CHAIN_EMBEDDED          0x00010000      /* embedded data */
 216 #define HAMMER2_CHAIN_UNUSED20000       0x00020000
 217 #define HAMMER2_CHAIN_UNUSED40000       0x00040000
 218 #define HAMMER2_CHAIN_UNUSED80000       0x00080000
 219 #define HAMMER2_CHAIN_DUPLICATED        0x00100000      /* fwd delete-dup */
 220 #define HAMMER2_CHAIN_PFSROOT           0x00200000      /* in pfs->cluster */
 221
 222 /*
 223  * Flags passed to hammer2_chain_lookup() and hammer2_chain_next()
 224  *
 225  * NOTE: MATCHIND allows an indirect block / freemap node to be returned
 226  *       when the passed key range matches the radix.  Remember that key_end
 227  *       is inclusive (e.g. {0x000,0xFFF}, not {0x000,0x1000}).
 228  */
 229 #define HAMMER2_LOOKUP_NOLOCK           0x00000001      /* ref only */
 230 #define HAMMER2_LOOKUP_NODATA           0x00000002      /* data left NULL */
 231 #define HAMMER2_LOOKUP_SHARED           0x00000100
 232 #define HAMMER2_LOOKUP_MATCHIND         0x00000200
 233 #define HAMMER2_LOOKUP_FREEMAP          0x00000400      /* freemap base */
 234 #define HAMMER2_LOOKUP_ALWAYS           0x00000800      /* resolve data */
 235
 236 /*
 237  * Flags passed to hammer2_chain_modify() and hammer2_chain_resize()
 238  *
 239  * NOTE: OPTDATA allows us to avoid instantiating buffers for INDIRECT
 240  *       blocks in the INITIAL-create state.
 241  */
 242 #define HAMMER2_MODIFY_OPTDATA          0x00000002      /* data can be NULL */
 243 #define HAMMER2_MODIFY_NO_MODIFY_TID    0x00000004
 244 #define HAMMER2_MODIFY_ASSERTNOCOPY     0x00000008      /* assert no del-dup */
 245 #define HAMMER2_MODIFY_NOREALLOC        0x00000010
 246 #define HAMMER2_MODIFY_INPLACE          0x00000020      /* don't del-dup */
 247
 248 /*
 249  * Flags passed to hammer2_chain_lock()
 250  */
 251 #define HAMMER2_RESOLVE_NEVER           1
 252 #define HAMMER2_RESOLVE_MAYBE           2
 253 #define HAMMER2_RESOLVE_ALWAYS          3
 254 #define HAMMER2_RESOLVE_MASK            0x0F
 255
 256 #define HAMMER2_RESOLVE_SHARED          0x10    /* request shared lock */
 257 #define HAMMER2_RESOLVE_NOREF           0x20    /* already ref'd on lock */
 258
 259 /*
 260  * Flags passed to hammer2_chain_delete()
 261  */
 262 #define HAMMER2_DELETE_WILLDUP          0x0001  /* no blk free, will be dup */
 263
 264 /*
 265  * Flags passed to hammer2_chain_delete_duplicate()
 266  */
 267 #define HAMMER2_DELDUP_RECORE           0x0001
 268
 269 /*
 270  * Cluster different types of storage together for allocations
 271  */
 272 #define HAMMER2_FREECACHE_INODE         0
 273 #define HAMMER2_FREECACHE_INDIR         1
 274 #define HAMMER2_FREECACHE_DATA          2
 275 #define HAMMER2_FREECACHE_UNUSED3       3
 276 #define HAMMER2_FREECACHE_TYPES         4
 277
 278 /*
 279  * hammer2_freemap_alloc() block preference
 280  */
 281 #define HAMMER2_OFF_NOPREF              ((hammer2_off_t)-1)
 282
 283 /*
 284  * BMAP read-ahead maximum parameters
 285  */
 286 #define HAMMER2_BMAP_COUNT              16      /* max bmap read-ahead */
 287 #define HAMMER2_BMAP_BYTES              (HAMMER2_PBUFSIZE * HAMMER2_BMAP_COUNT)
 288
 289 /*
 290  * Misc
 291  */
 292 #define HAMMER2_FLUSH_DEPTH_LIMIT       10      /* stack recursion limit */
 293
 294 /*
 295  * HAMMER2 IN-MEMORY CACHE OF MEDIA STRUCTURES
 296  *
 297  * There is an in-memory representation of all on-media data structure.
 298  * Basically everything is represented by a hammer2_chain structure
 299  * in-memory and other higher-level structures map to chains.
 300  *
 301  * A great deal of data is accessed simply via its buffer cache buffer,
 302  * which is mapped for the duration of the chain's lock.  However, because
 303  * chains may represent blocks smaller than the 16KB minimum we impose
 304  * on buffer cache buffers, we cannot hold related buffer cache buffers
 305  * locked for smaller blocks.  In these situations we kmalloc() a copy
 306  * of the block.
 307  *
 308  * When modifications are made to a chain a new filesystem block must be
 309  * allocated.  Multiple modifications do not necessarily allocate new
 310  * blocks.  However, when a flush occurs a flush synchronization point
 311  * is created and any new modifications made after this point will allocate
 312  * a new block even if the chain is already in a modified state.
 313  *
 314  * The in-memory representation may remain cached (for example in order to
 315  * placemark clustering locks) even after the related data has been
 316  * detached.
 317  *
 318  *                              CORE SHARING
 319  *
 320  * In order to support concurrent flushes a flush synchronization point
 321  * is created represented by a transaction id.  Among other things,
 322  * operations may move filesystem objects from one part of the topology
 323  * to another (for example, if you rename a file or when indirect blocks
 324  * are created or destroyed, and a few other things).  When this occurs
 325  * across a flush synchronization point the flusher needs to be able to
 326  * recurse down BOTH the 'before' version of the topology and the 'after'
 327  * version.
 328  *
 329  * To facilitate this modifications to chains do what is called a
 330  * DELETE-DUPLICATE operation.  Chains are not actually moved in-memory.
 331  * Instead the chain we wish to move is deleted and a new chain is created
 332  * at the target location in the topology.  ANY SUBCHAINS PLACED UNDER THE
 333  * CHAIN BEING MOVED HAVE TO EXIST IN BOTH PLACES.  To make this work
 334  * all sub-chains are managed by the hammer2_chain_core structure.  This
 335  * structure can be multi-homed, meaning that it can have more than one
 336  * chain as its parent.  When a chain is delete-duplicated the chain's core
 337  * becomes shared under both the old and new chain.
 338  *
 339  *                              STALE CHAINS
 340  *
 341  * When a chain is delete-duplicated the old chain typically becomes stale.
 342  * This is detected via the HAMMER2_CHAIN_DUPLICATED flag in chain->flags.
 343  * To avoid executing live filesystem operations on stale chains, the inode
 344  * locking code will follow stale chains via core->ownerq until it finds
 345  * the live chain.  The lock prevents ripups by other threads.  Lookups
 346  * must properly order locking operations to prevent other threads from
 347  * racing the lookup operation and will also follow stale chains when
 348  * required.
 349  */
 350
 351 RB_HEAD(hammer2_inode_tree, hammer2_inode);
 352
 353 /*
 354  * A hammer2 inode.
 355  *
 356  * NOTE: The inode's attribute CST which is also used to lock the inode
 357  *       is embedded in the chain (chain.cst) and aliased w/ attr_cst.
 358  */
 359 struct hammer2_inode {
 360         RB_ENTRY(hammer2_inode) rbnode;         /* inumber lookup (HL) */
 361         ccms_cst_t              topo_cst;       /* directory topology cst */
 362         struct hammer2_pfsmount *pmp;           /* PFS mount */
 363         struct hammer2_inode    *pip;           /* parent inode */
 364         struct vnode            *vp;
 365         hammer2_chain_t         *chain;         /* NOTE: rehomed on rename */
 366         struct lockf            advlock;
 367         hammer2_tid_t           inum;
 368         u_int                   flags;
 369         u_int                   refs;           /* +vpref, +flushref */
 370         uint8_t                 comp_heuristic;
 371         hammer2_off_t           size;
 372         uint64_t                mtime;
 373 };
 374
 375 typedef struct hammer2_inode hammer2_inode_t;
 376
 377 #define HAMMER2_INODE_MODIFIED          0x0001
 378 #define HAMMER2_INODE_SROOT             0x0002  /* kmalloc special case */
 379 #define HAMMER2_INODE_RENAME_INPROG     0x0004
 380 #define HAMMER2_INODE_ONRBTREE          0x0008
 381 #define HAMMER2_INODE_RESIZED           0x0010
 382 #define HAMMER2_INODE_MTIME             0x0020
 383
 384 int hammer2_inode_cmp(hammer2_inode_t *ip1, hammer2_inode_t *ip2);
 385 RB_PROTOTYPE2(hammer2_inode_tree, hammer2_inode, rbnode, hammer2_inode_cmp,
 386                 hammer2_tid_t);
 387
 388 /*
 389  * A hammer2 transaction and flush sequencing structure.
 390  *
 391  * This global structure is tied into hammer2_mount and is used
 392  * to sequence modifying operations and flushes.
 393  *
 394  * (a) Any modifying operations with sync_tid >= flush_tid will stall until
 395  *     all modifying operating with sync_tid < flush_tid complete.
 396  *
 397  *     The flush related to flush_tid stalls until all modifying operations
 398  *     with sync_tid < flush_tid complete.
 399  *
 400  * (b) Once unstalled, modifying operations with sync_tid > flush_tid are
 401  *     allowed to run.  All modifications cause modify/duplicate operations
 402  *     to occur on the related chains.  Note that most INDIRECT blocks will
 403  *     be unaffected because the modifications just overload the RBTREE
 404  *     structurally instead of actually modifying the indirect blocks.
 405  *
 406  * (c) The actual flush unstalls and RUNS CONCURRENTLY with (b), but only
 407  *     utilizes the chain structures with sync_tid <= flush_tid.  The
 408  *     flush will modify related indirect blocks and inodes in-place
 409  *     (rather than duplicate) since the adjustments are compatible with
 410  *     (b)'s RBTREE overloading
 411  *
 412  *     SPECIAL NOTE:  Inode modifications have to also propagate along any
 413  *                    modify/duplicate chains.  File writes detect the flush
 414  *                    and force out the conflicting buffer cache buffer(s)
 415  *                    before reusing them.
 416  *
 417  * (d) Snapshots can be made instantly but must be flushed and disconnected
 418  *     from their duplicative source before they can be mounted.  This is
 419  *     because while H2's on-media structure supports forks, its in-memory
 420  *     structure only supports very simple forking for background flushing
 421  *     purposes.
 422  *
 423  * TODO: Flush merging.  When fsync() is called on multiple discrete files
 424  *       concurrently there is no reason to stall the second fsync.
 425  *       The final flush that reaches to root can cover both fsync()s.
 426  *
 427  *     The chains typically terminate as they fly onto the disk.  The flush
 428  *     ultimately reaches the volume header.
 429  */
 430 struct hammer2_trans {
 431         TAILQ_ENTRY(hammer2_trans) entry;
 432         struct hammer2_pfsmount *pmp;
 433         hammer2_tid_t           real_tid;
 434         hammer2_tid_t           sync_tid;
 435         hammer2_tid_t           inode_tid;
 436         thread_t                td;             /* pointer */
 437         int                     flags;
 438         int                     blocked;
 439         uint8_t                 inodes_created;
 440         uint8_t                 dummy[7];
 441 };
 442
 443 typedef struct hammer2_trans hammer2_trans_t;
 444
 445 #define HAMMER2_TRANS_ISFLUSH           0x0001  /* formal flush */
 446 #define HAMMER2_TRANS_UNUSED0002        0x0002
 447 #define HAMMER2_TRANS_BUFCACHE          0x0004  /* from bioq strategy write */
 448 #define HAMMER2_TRANS_NEWINODE          0x0008  /* caller allocating inode */
 449 #define HAMMER2_TRANS_ISALLOCATING      0x0010  /* in allocator */
 450
 451 #define HAMMER2_FREEMAP_HEUR_NRADIX     4       /* pwr 2 PBUFRADIX-MINIORADIX */
 452 #define HAMMER2_FREEMAP_HEUR_TYPES      8
 453 #define HAMMER2_FREEMAP_HEUR            (HAMMER2_FREEMAP_HEUR_NRADIX * \
 454                                          HAMMER2_FREEMAP_HEUR_TYPES)
 455
 456 /*
 457  * Global (per device) mount structure for device (aka vp->v_mount->hmp)
 458  */
 459 TAILQ_HEAD(hammer2_trans_queue, hammer2_trans);
 460
 461 struct hammer2_mount {
 462         struct vnode    *devvp;         /* device vnode */
 463         int             ronly;          /* read-only mount */
 464         int             pmp_count;      /* PFS mounts backed by us */
 465         TAILQ_ENTRY(hammer2_mount) mntentry; /* hammer2_mntlist */
 466
 467         struct malloc_type *mchain;
 468         int             nipstacks;
 469         int             maxipstacks;
 470         hammer2_chain_t vchain;         /* anchor chain (topology) */
 471         hammer2_chain_t fchain;         /* anchor chain (freemap) */
 472         hammer2_inode_t *sroot;         /* super-root localized to media */
 473         struct lock     alloclk;        /* lockmgr lock */
 474         struct lock     voldatalk;      /* lockmgr lock */
 475         struct hammer2_trans_queue transq; /* all in-progress transactions */
 476         hammer2_off_t   heur_freemap[HAMMER2_FREEMAP_HEUR];
 477         int             flushcnt;       /* #of flush trans on the list */
 478
 479         int             volhdrno;       /* last volhdrno written */
 480         hammer2_volume_data_t voldata;
 481         hammer2_volume_data_t volsync;  /* synchronized voldata */
 482 };
 483
 484 typedef struct hammer2_mount hammer2_mount_t;
 485
 486 /*
 487  * HAMMER2 cluster - a device/root associated with a PFS.
 488  *
 489  * A PFS may have several hammer2_cluster's associated with it.
 490  */
 491 #define HAMMER2_MAXCLUSTER      8
 492
 493 struct hammer2_cluster {
 494         int                     nchains;
 495         int                     status;
 496         hammer2_chain_t         *chains[HAMMER2_MAXCLUSTER];
 497 };
 498
 499 typedef struct hammer2_cluster hammer2_cluster_t;
 500
 501 /*
 502  * HAMMER2 PFS mount point structure (aka vp->v_mount->mnt_data).
 503  * This has a 1:1 correspondence to struct mount (note that the
 504  * hammer2_mount structure has a N:1 correspondence).
 505  *
 506  * This structure represents a cluster mount and not necessarily a
 507  * PFS under a specific device mount (HMP).  The distinction is important
 508  * because the elements backing a cluster mount can change on the fly.
 509  *
 510  * Usually the first element under the cluster represents the original
 511  * user-requested mount that bootstraps the whole mess.  In significant
 512  * setups the original is usually just a read-only media image (or
 513  * representitive file) that simply contains a bootstrap volume header
 514  * listing the configuration.
 515  */
 516 struct hammer2_pfsmount {
 517         struct mount            *mp;
 518         hammer2_cluster_t       cluster;
 519         hammer2_inode_t         *iroot;         /* PFS root inode */
 520         hammer2_off_t           inode_count;    /* copy of inode_count */
 521         ccms_domain_t           ccms_dom;
 522         struct netexport        export;         /* nfs export */
 523         int                     ronly;          /* read-only mount */
 524         struct malloc_type      *minode;
 525         struct malloc_type      *mmsg;
 526         kdmsg_iocom_t           iocom;
 527         struct spinlock         inum_spin;      /* inumber lookup */
 528         struct hammer2_inode_tree inum_tree;
 529         long                    inmem_inodes;
 530         long                    inmem_chains;
 531         int                     inmem_waiting;
 532         thread_t                wthread_td;     /* write thread td */
 533         struct bio_queue_head   wthread_bioq;   /* logical buffer bioq */
 534         struct mtx              wthread_mtx;    /* interlock */
 535         int                     wthread_destroy;/* termination sequencing */
 536 };
 537
 538 typedef struct hammer2_pfsmount hammer2_pfsmount_t;
 539
 540 struct hammer2_cbinfo {
 541         hammer2_chain_t *chain;
 542         void (*func)(hammer2_chain_t *, struct buf *, char *, void *);
 543         void *arg;
 544         size_t boff;
 545 };
 546
 547 typedef struct hammer2_cbinfo hammer2_cbinfo_t;
 548
 549 #if defined(_KERNEL)
 550
 551 MALLOC_DECLARE(M_HAMMER2);
 552
 553 #define VTOI(vp)        ((hammer2_inode_t *)(vp)->v_data)
 554 #define ITOV(ip)        ((ip)->vp)
 555
 556 /*
 557  * Currently locked chains retain the locked buffer cache buffer for
 558  * indirect blocks, and indirect blocks can be one of two sizes.  The
 559  * device buffer has to match the case to avoid deadlocking recursive
 560  * chains that might otherwise try to access different offsets within
 561  * the same device buffer.
 562  */
 563 static __inline
 564 int
 565 hammer2_devblkradix(int radix)
 566 {
 567 #if 1
 568         if (radix <= HAMMER2_LBUFRADIX) {
 569                 return (HAMMER2_LBUFRADIX);
 570         } else {
 571                 return (HAMMER2_PBUFRADIX);
 572         }
 573 #else
 574         return (HAMMER2_PBUFRADIX);
 575 #endif
 576 }
 577
 578 static __inline
 579 size_t
 580 hammer2_devblksize(size_t bytes)
 581 {
 582 #if 1
 583         if (bytes <= HAMMER2_LBUFSIZE) {
 584                 return(HAMMER2_LBUFSIZE);
 585         } else {
 586                 KKASSERT(bytes <= HAMMER2_PBUFSIZE &&
 587                          (bytes ^ (bytes - 1)) == ((bytes << 1) - 1));
 588                 return (HAMMER2_PBUFSIZE);
 589         }
 590 #else
 591         KKASSERT(bytes <= HAMMER2_PBUFSIZE &&
 592                  (bytes ^ (bytes - 1)) == ((bytes << 1) - 1));
 593         return(HAMMER2_PBUFSIZE);
 594 #endif
 595 }
 596
 597
 598 static __inline
 599 hammer2_pfsmount_t *
 600 MPTOPMP(struct mount *mp)
 601 {
 602         return ((hammer2_pfsmount_t *)mp->mnt_data);
 603 }
 604
 605 extern struct vop_ops hammer2_vnode_vops;
 606 extern struct vop_ops hammer2_spec_vops;
 607 extern struct vop_ops hammer2_fifo_vops;
 608
 609 extern int hammer2_debug;
 610 extern int hammer2_cluster_enable;
 611 extern int hammer2_hardlink_enable;
 612 extern long hammer2_iod_file_read;
 613 extern long hammer2_iod_meta_read;
 614 extern long hammer2_iod_indr_read;
 615 extern long hammer2_iod_fmap_read;
 616 extern long hammer2_iod_volu_read;
 617 extern long hammer2_iod_file_write;
 618 extern long hammer2_iod_meta_write;
 619 extern long hammer2_iod_indr_write;
 620 extern long hammer2_iod_fmap_write;
 621 extern long hammer2_iod_volu_write;
 622 extern long hammer2_ioa_file_read;
 623 extern long hammer2_ioa_meta_read;
 624 extern long hammer2_ioa_indr_read;
 625 extern long hammer2_ioa_fmap_read;
 626 extern long hammer2_ioa_volu_read;
 627 extern long hammer2_ioa_file_write;
 628 extern long hammer2_ioa_meta_write;
 629 extern long hammer2_ioa_indr_write;
 630 extern long hammer2_ioa_fmap_write;
 631 extern long hammer2_ioa_volu_write;
 632
 633 extern struct objcache *cache_buffer_read;
 634 extern struct objcache *cache_buffer_write;
 635
 636 extern int destroy;
 637 extern int write_thread_wakeup;
 638
 639 extern mtx_t thread_protect;
 640
 641 /*
 642  * hammer2_subr.c
 643  */
 644 #define hammer2_icrc32(buf, size)       iscsi_crc32((buf), (size))
 645 #define hammer2_icrc32c(buf, size, crc) iscsi_crc32_ext((buf), (size), (crc))
 646
 647 hammer2_chain_t *hammer2_inode_lock_ex(hammer2_inode_t *ip);
 648 hammer2_chain_t *hammer2_inode_lock_sh(hammer2_inode_t *ip);
 649 void hammer2_inode_unlock_ex(hammer2_inode_t *ip, hammer2_chain_t *chain);
 650 void hammer2_inode_unlock_sh(hammer2_inode_t *ip, hammer2_chain_t *chain);
 651 void hammer2_chain_refactor(hammer2_chain_t **chainp);
 652 void hammer2_voldata_lock(hammer2_mount_t *hmp);
 653 void hammer2_voldata_unlock(hammer2_mount_t *hmp, int modify);
 654 ccms_state_t hammer2_inode_lock_temp_release(hammer2_inode_t *ip);
 655 void hammer2_inode_lock_temp_restore(hammer2_inode_t *ip, ccms_state_t ostate);
 656 ccms_state_t hammer2_inode_lock_upgrade(hammer2_inode_t *ip);
 657 void hammer2_inode_lock_downgrade(hammer2_inode_t *ip, ccms_state_t ostate);
 658
 659 void hammer2_mount_exlock(hammer2_mount_t *hmp);
 660 void hammer2_mount_shlock(hammer2_mount_t *hmp);
 661 void hammer2_mount_unlock(hammer2_mount_t *hmp);
 662
 663 int hammer2_get_dtype(hammer2_chain_t *chain);
 664 int hammer2_get_vtype(hammer2_chain_t *chain);
 665 u_int8_t hammer2_get_obj_type(enum vtype vtype);
 666 void hammer2_time_to_timespec(u_int64_t xtime, struct timespec *ts);
 667 u_int64_t hammer2_timespec_to_time(struct timespec *ts);
 668 u_int32_t hammer2_to_unix_xid(uuid_t *uuid);
 669 void hammer2_guid_to_uuid(uuid_t *uuid, u_int32_t guid);
 670
 671 hammer2_key_t hammer2_dirhash(const unsigned char *name, size_t len);
 672 int hammer2_getradix(size_t bytes);
 673
 674 int hammer2_calc_logical(hammer2_inode_t *ip, hammer2_off_t uoff,
 675                         hammer2_key_t *lbasep, hammer2_key_t *leofp);
 676 int hammer2_calc_physical(hammer2_inode_t *ip, hammer2_key_t lbase);
 677 void hammer2_update_time(uint64_t *timep);
 678
 679 /*
 680  * hammer2_inode.c
 681  */
 682 struct vnode *hammer2_igetv(hammer2_inode_t *ip, int *errorp);
 683
 684 void hammer2_inode_lock_nlinks(hammer2_inode_t *ip);
 685 void hammer2_inode_unlock_nlinks(hammer2_inode_t *ip);
 686 hammer2_inode_t *hammer2_inode_lookup(hammer2_pfsmount_t *pmp,
 687                         hammer2_tid_t inum);
 688 hammer2_inode_t *hammer2_inode_get(hammer2_pfsmount_t *pmp,
 689                         hammer2_inode_t *dip, hammer2_chain_t *chain);
 690 void hammer2_inode_free(hammer2_inode_t *ip);
 691 void hammer2_inode_ref(hammer2_inode_t *ip);
 692 void hammer2_inode_drop(hammer2_inode_t *ip);
 693 void hammer2_inode_repoint(hammer2_inode_t *ip, hammer2_inode_t *pip,
 694                         hammer2_chain_t *chain);
 695
 696 hammer2_inode_t *hammer2_inode_create(hammer2_trans_t *trans,
 697                         hammer2_inode_t *dip,
 698                         struct vattr *vap, struct ucred *cred,
 699                         const uint8_t *name, size_t name_len,
 700                         hammer2_chain_t **chainp, int *errorp);
 701 int hammer2_inode_connect(hammer2_trans_t *trans, int hlink,
 702                         hammer2_inode_t *dip, hammer2_chain_t **chainp,
 703                         const uint8_t *name, size_t name_len);
 704 hammer2_inode_t *hammer2_inode_common_parent(hammer2_inode_t *fdip,
 705                         hammer2_inode_t *tdip);
 706 void hammer2_inode_fsync(hammer2_trans_t *trans, hammer2_inode_t *ip,
 707                         hammer2_chain_t **parentp);
 708 int hammer2_unlink_file(hammer2_trans_t *trans, hammer2_inode_t *dip,
 709                         const uint8_t *name, size_t name_len, int isdir,
 710                         int *hlinkp);
 711 int hammer2_hardlink_consolidate(hammer2_trans_t *trans, hammer2_inode_t *ip,
 712                         hammer2_chain_t **chainp,
 713                         hammer2_inode_t *tdip, int linkcnt);
 714 int hammer2_hardlink_deconsolidate(hammer2_trans_t *trans, hammer2_inode_t *dip,
 715                         hammer2_chain_t **chainp, hammer2_chain_t **ochainp);
 716 int hammer2_hardlink_find(hammer2_inode_t *dip,
 717                         hammer2_chain_t **chainp, hammer2_chain_t **ochainp);
 718
 719 /*
 720  * hammer2_chain.c
 721  */
 722 void hammer2_modify_volume(hammer2_mount_t *hmp);
 723 hammer2_chain_t *hammer2_chain_alloc(hammer2_mount_t *hmp, hammer2_pfsmount_t *pmp,
 724                                 hammer2_trans_t *trans, hammer2_blockref_t *bref);
 725 void hammer2_chain_core_alloc(hammer2_trans_t *trans, hammer2_chain_t *nchain,
 726                                 hammer2_chain_t *ochain);
 727 void hammer2_chain_ref(hammer2_chain_t *chain);
 728 void hammer2_chain_drop(hammer2_chain_t *chain);
 729 int hammer2_chain_lock(hammer2_chain_t *chain, int how);
 730 void hammer2_chain_load_async(hammer2_chain_t *chain,
 731                                 void (*func)(hammer2_chain_t *, struct buf *,
 732                                              char *, void *),
 733                                 void *arg);
 734 void hammer2_chain_moved(hammer2_chain_t *chain);
 735 void hammer2_chain_modify(hammer2_trans_t *trans,
 736                                 hammer2_chain_t **chainp, int flags);
 737 hammer2_inode_data_t *hammer2_chain_modify_ip(hammer2_trans_t *trans,
 738                                 hammer2_inode_t *ip, hammer2_chain_t **chainp,
 739                                 int flags);
 740 void hammer2_chain_resize(hammer2_trans_t *trans, hammer2_inode_t *ip,
 741                                 hammer2_chain_t *parent,
 742                                 hammer2_chain_t **chainp,
 743                                 int nradix, int flags);
 744 void hammer2_chain_unlock(hammer2_chain_t *chain);
 745 void hammer2_chain_wait(hammer2_chain_t *chain);
 746 hammer2_chain_t *hammer2_chain_get(hammer2_chain_t *parent,
 747                                 hammer2_blockref_t *bref);
 748 hammer2_chain_t *hammer2_chain_lookup_init(hammer2_chain_t *parent, int flags);
 749 void hammer2_chain_lookup_done(hammer2_chain_t *parent);
 750 hammer2_chain_t *hammer2_chain_lookup(hammer2_chain_t **parentp,
 751                                 hammer2_key_t *key_nextp,
 752                                 hammer2_key_t key_beg, hammer2_key_t key_end,
 753                                 int *cache_indexp, int flags);
 754 hammer2_chain_t *hammer2_chain_next(hammer2_chain_t **parentp,
 755                                 hammer2_chain_t *chain,
 756                                 hammer2_key_t *key_nextp,
 757                                 hammer2_key_t key_beg, hammer2_key_t key_end,
 758                                 int *cache_indexp, int flags);
 759
 760 int hammer2_chain_create(hammer2_trans_t *trans,
 761                                 hammer2_chain_t **parentp,
 762                                 hammer2_chain_t **chainp,
 763                                 hammer2_key_t key, int keybits,
 764                                 int type, size_t bytes);
 765 void hammer2_chain_duplicate(hammer2_trans_t *trans, hammer2_chain_t **parentp,
 766                                 hammer2_chain_t **chainp,
 767                                 hammer2_blockref_t *bref, int snapshot);
 768 int hammer2_chain_snapshot(hammer2_trans_t *trans, hammer2_chain_t **chainp,
 769                                 hammer2_ioc_pfs_t *pfs);
 770 void hammer2_chain_delete(hammer2_trans_t *trans, hammer2_chain_t *chain,
 771                                 int flags);
 772 void hammer2_chain_delete_duplicate(hammer2_trans_t *trans,
 773                                 hammer2_chain_t **chainp, int flags);
 774 void hammer2_chain_flush(hammer2_trans_t *trans, hammer2_chain_t **chainp);
 775 void hammer2_chain_commit(hammer2_trans_t *trans, hammer2_chain_t *chain);
 776 void hammer2_chain_setsubmod(hammer2_trans_t *trans, hammer2_chain_t *chain);
 777
 778 void hammer2_chain_memory_wait(hammer2_pfsmount_t *pmp);
 779 void hammer2_chain_memory_wakeup(hammer2_pfsmount_t *pmp);
 780 void hammer2_chain_countbrefs(hammer2_chain_t *chain,
 781                                 hammer2_blockref_t *base, int count);
 782 void hammer2_chain_layer_check_locked(hammer2_mount_t *hmp,
 783                                 hammer2_chain_core_t *core);
 784
 785 int hammer2_base_find(hammer2_chain_t *chain,
 786                                 hammer2_blockref_t *base, int count,
 787                                 int *cache_indexp, hammer2_key_t *key_nextp,
 788                                 hammer2_key_t key_beg, hammer2_key_t key_end);
 789 void hammer2_base_delete(hammer2_chain_t *chain,
 790                                 hammer2_blockref_t *base, int count,
 791                                 int *cache_indexp, hammer2_chain_t *child);
 792 void hammer2_base_insert(hammer2_chain_t *chain,
 793                                 hammer2_blockref_t *base, int count,
 794                                 int *cache_indexp, hammer2_chain_t *child);
 795
 796 /*
 797  * hammer2_trans.c
 798  */
 799 void hammer2_trans_init(hammer2_trans_t *trans,
 800                         hammer2_pfsmount_t *pmp, int flags);
 801 void hammer2_trans_clear_invfsync(hammer2_trans_t *trans);
 802 void hammer2_trans_done(hammer2_trans_t *trans);
 803
 804 /*
 805  * hammer2_ioctl.c
 806  */
 807 int hammer2_ioctl(hammer2_inode_t *ip, u_long com, void *data,
 808                                 int fflag, struct ucred *cred);
 809
 810 /*
 811  * hammer2_msgops.c
 812  */
 813 int hammer2_msg_dbg_rcvmsg(kdmsg_msg_t *msg);
 814 int hammer2_msg_adhoc_input(kdmsg_msg_t *msg);
 815
 816 /*
 817  * hammer2_vfsops.c
 818  */
 819 void hammer2_clusterctl_wakeup(kdmsg_iocom_t *iocom);
 820 void hammer2_volconf_update(hammer2_pfsmount_t *pmp, int index);
 821 void hammer2_cluster_reconnect(hammer2_pfsmount_t *pmp, struct file *fp);
 822 void hammer2_dump_chain(hammer2_chain_t *chain, int tab, int *countp);
 823 void hammer2_bioq_sync(hammer2_pfsmount_t *pmp);
 824 int hammer2_vfs_sync(struct mount *mp, int waitflags);
 825
 826 /*
 827  * hammer2_freemap.c
 828  */
 829 int hammer2_freemap_alloc(hammer2_trans_t *trans, hammer2_mount_t *hmp,
 830                                 hammer2_blockref_t *bref, size_t bytes);
 831 void hammer2_freemap_free(hammer2_trans_t *trans, hammer2_mount_t *hmp,
 832                                 hammer2_blockref_t *bref, int how);
 833
 834
 835 #endif /* !_KERNEL */
 836 #endif /* !_VFS_HAMMER2_HAMMER2_H_ */