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_tid;     /* check update against parent */
 144         u_int           chain_count;    /* total chains in layers */
 145         u_int           sharecnt;
 146         u_int           flags;
 147         u_int           live_count;     /* live (not deleted) chains in tree */
 148         int             generation;     /* generation number (inserts only) */
 149 };
 150
 151 typedef struct hammer2_chain_core hammer2_chain_core_t;
 152
 153 #define HAMMER2_CORE_UNUSED0001         0x0001
 154 #define HAMMER2_CORE_COUNTEDBREFS       0x0002
 155
 156 struct hammer2_chain {
 157         RB_ENTRY(hammer2_chain) rbnode;         /* node */
 158         TAILQ_ENTRY(hammer2_chain) core_entry;  /* contemporary chains */
 159         hammer2_chain_layer_t   *inlayer;
 160         hammer2_blockref_t      bref;
 161         hammer2_chain_core_t    *core;
 162         hammer2_chain_core_t    *above;
 163         struct hammer2_state    *state;         /* if active cache msg */
 164         struct hammer2_mount    *hmp;
 165         struct hammer2_pfsmount *pmp;           /* can be NULL */
 166
 167         hammer2_tid_t   modify_tid;             /* snapshot/flush filter */
 168         hammer2_tid_t   delete_tid;
 169         hammer2_key_t   data_count;             /* delta's to apply */
 170         hammer2_key_t   inode_count;            /* delta's to apply */
 171         struct buf      *bp;                    /* physical data buffer */
 172         u_int           bytes;                  /* physical data size */
 173         u_int           flags;
 174         u_int           refs;
 175         u_int           lockcnt;
 176         int             debug_reason;
 177         hammer2_media_data_t *data;             /* data pointer shortcut */
 178         TAILQ_ENTRY(hammer2_chain) flush_node;  /* flush deferral list */
 179 };
 180
 181 typedef struct hammer2_chain hammer2_chain_t;
 182
 183 int hammer2_chain_cmp(hammer2_chain_t *chain1, hammer2_chain_t *chain2);
 184 RB_PROTOTYPE(hammer2_chain_tree, hammer2_chain, rbnode, hammer2_chain_cmp);
 185
 186 /*
 187  * Special notes on flags:
 188  *
 189  * INITIAL - This flag allows a chain to be created and for storage to
 190  *           be allocated without having to immediately instantiate the
 191  *           related buffer.  The data is assumed to be all-zeros.  It
 192  *           is primarily used for indirect blocks.
 193  *
 194  * MOVED   - A modified chain becomes MOVED after it flushes.  A chain
 195  *           can also become MOVED if it is moved within the topology
 196  *           (even if not modified).
 197  */
 198 #define HAMMER2_CHAIN_MODIFIED          0x00000001      /* dirty chain data */
 199 #define HAMMER2_CHAIN_ALLOCATED         0x00000002      /* kmalloc'd chain */
 200 #define HAMMER2_CHAIN_DIRTYBP           0x00000004      /* dirty on unlock */
 201 #define HAMMER2_CHAIN_FORCECOW          0x00000008      /* force copy-on-wr */
 202 #define HAMMER2_CHAIN_DELETED           0x00000010      /* deleted chain */
 203 #define HAMMER2_CHAIN_INITIAL           0x00000020      /* initial create */
 204 #define HAMMER2_CHAIN_FLUSHED           0x00000040      /* flush on unlock */
 205 #define HAMMER2_CHAIN_MOVED             0x00000080      /* bref changed */
 206 #define HAMMER2_CHAIN_IOFLUSH           0x00000100      /* bawrite on put */
 207 #define HAMMER2_CHAIN_DEFERRED          0x00000200      /* on a deferral list */
 208 #define HAMMER2_CHAIN_DESTROYED         0x00000400      /* destroying inode */
 209 #define HAMMER2_CHAIN_VOLUMESYNC        0x00000800      /* needs volume sync */
 210 #define HAMMER2_CHAIN_UNUSED1000        0x00001000
 211 #define HAMMER2_CHAIN_MOUNTED           0x00002000      /* PFS is mounted */
 212 #define HAMMER2_CHAIN_ONRBTREE          0x00004000      /* on parent RB tree */
 213 #define HAMMER2_CHAIN_SNAPSHOT          0x00008000      /* snapshot special */
 214 #define HAMMER2_CHAIN_EMBEDDED          0x00010000      /* embedded data */
 215 #define HAMMER2_CHAIN_UNUSED20000       0x00020000
 216 #define HAMMER2_CHAIN_UNUSED40000       0x00040000
 217 #define HAMMER2_CHAIN_UNUSED80000       0x00080000
 218 #define HAMMER2_CHAIN_DUPLICATED        0x00100000      /* fwd delete-dup */
 219 #define HAMMER2_CHAIN_PFSROOT           0x00200000      /* in pfs->cluster */
 220
 221 /*
 222  * Flags passed to hammer2_chain_lookup() and hammer2_chain_next()
 223  *
 224  * NOTE: MATCHIND allows an indirect block / freemap node to be returned
 225  *       when the passed key range matches the radix.  Remember that key_end
 226  *       is inclusive (e.g. {0x000,0xFFF}, not {0x000,0x1000}).
 227  */
 228 #define HAMMER2_LOOKUP_NOLOCK           0x00000001      /* ref only */
 229 #define HAMMER2_LOOKUP_NODATA           0x00000002      /* data left NULL */
 230 #define HAMMER2_LOOKUP_SHARED           0x00000100
 231 #define HAMMER2_LOOKUP_MATCHIND         0x00000200
 232 #define HAMMER2_LOOKUP_FREEMAP          0x00000400      /* freemap base */
 233 #define HAMMER2_LOOKUP_ALWAYS           0x00000800      /* resolve data */
 234
 235 /*
 236  * Flags passed to hammer2_chain_modify() and hammer2_chain_resize()
 237  *
 238  * NOTE: OPTDATA allows us to avoid instantiating buffers for INDIRECT
 239  *       blocks in the INITIAL-create state.
 240  */
 241 #define HAMMER2_MODIFY_OPTDATA          0x00000002      /* data can be NULL */
 242 #define HAMMER2_MODIFY_NO_MODIFY_TID    0x00000004
 243 #define HAMMER2_MODIFY_ASSERTNOCOPY     0x00000008      /* assert no del-dup */
 244 #define HAMMER2_MODIFY_NOREALLOC        0x00000010
 245 #define HAMMER2_MODIFY_INPLACE          0x00000020      /* don't del-dup */
 246
 247 /*
 248  * Flags passed to hammer2_chain_lock()
 249  */
 250 #define HAMMER2_RESOLVE_NEVER           1
 251 #define HAMMER2_RESOLVE_MAYBE           2
 252 #define HAMMER2_RESOLVE_ALWAYS          3
 253 #define HAMMER2_RESOLVE_MASK            0x0F
 254
 255 #define HAMMER2_RESOLVE_SHARED          0x10    /* request shared lock */
 256 #define HAMMER2_RESOLVE_NOREF           0x20    /* already ref'd on lock */
 257
 258 /*
 259  * Flags passed to hammer2_chain_delete()
 260  */
 261 #define HAMMER2_DELETE_WILLDUP          0x0001  /* no blk free, will be dup */
 262
 263 /*
 264  * Flags passed to hammer2_chain_delete_duplicate()
 265  */
 266 #define HAMMER2_DELDUP_RECORE           0x0001
 267
 268 /*
 269  * Cluster different types of storage together for allocations
 270  */
 271 #define HAMMER2_FREECACHE_INODE         0
 272 #define HAMMER2_FREECACHE_INDIR         1
 273 #define HAMMER2_FREECACHE_DATA          2
 274 #define HAMMER2_FREECACHE_UNUSED3       3
 275 #define HAMMER2_FREECACHE_TYPES         4
 276
 277 /*
 278  * hammer2_freemap_alloc() block preference
 279  */
 280 #define HAMMER2_OFF_NOPREF              ((hammer2_off_t)-1)
 281
 282 /*
 283  * BMAP read-ahead maximum parameters
 284  */
 285 #define HAMMER2_BMAP_COUNT              16      /* max bmap read-ahead */
 286 #define HAMMER2_BMAP_BYTES              (HAMMER2_PBUFSIZE * HAMMER2_BMAP_COUNT)
 287
 288 /*
 289  * Misc
 290  */
 291 #define HAMMER2_FLUSH_DEPTH_LIMIT       40      /* stack recursion limit */
 292
 293 /*
 294  * HAMMER2 IN-MEMORY CACHE OF MEDIA STRUCTURES
 295  *
 296  * There is an in-memory representation of all on-media data structure.
 297  * Basically everything is represented by a hammer2_chain structure
 298  * in-memory and other higher-level structures map to chains.
 299  *
 300  * A great deal of data is accessed simply via its buffer cache buffer,
 301  * which is mapped for the duration of the chain's lock.  However, because
 302  * chains may represent blocks smaller than the 16KB minimum we impose
 303  * on buffer cache buffers, we cannot hold related buffer cache buffers
 304  * locked for smaller blocks.  In these situations we kmalloc() a copy
 305  * of the block.
 306  *
 307  * When modifications are made to a chain a new filesystem block must be
 308  * allocated.  Multiple modifications do not necessarily allocate new
 309  * blocks.  However, when a flush occurs a flush synchronization point
 310  * is created and any new modifications made after this point will allocate
 311  * a new block even if the chain is already in a modified state.
 312  *
 313  * The in-memory representation may remain cached (for example in order to
 314  * placemark clustering locks) even after the related data has been
 315  * detached.
 316  *
 317  *                              CORE SHARING
 318  *
 319  * In order to support concurrent flushes a flush synchronization point
 320  * is created represented by a transaction id.  Among other things,
 321  * operations may move filesystem objects from one part of the topology
 322  * to another (for example, if you rename a file or when indirect blocks
 323  * are created or destroyed, and a few other things).  When this occurs
 324  * across a flush synchronization point the flusher needs to be able to
 325  * recurse down BOTH the 'before' version of the topology and the 'after'
 326  * version.
 327  *
 328  * To facilitate this modifications to chains do what is called a
 329  * DELETE-DUPLICATE operation.  Chains are not actually moved in-memory.
 330  * Instead the chain we wish to move is deleted and a new chain is created
 331  * at the target location in the topology.  ANY SUBCHAINS PLACED UNDER THE
 332  * CHAIN BEING MOVED HAVE TO EXIST IN BOTH PLACES.  To make this work
 333  * all sub-chains are managed by the hammer2_chain_core structure.  This
 334  * structure can be multi-homed, meaning that it can have more than one
 335  * chain as its parent.  When a chain is delete-duplicated the chain's core
 336  * becomes shared under both the old and new chain.
 337  *
 338  *                              STALE CHAINS
 339  *
 340  * When a chain is delete-duplicated the old chain typically becomes stale.
 341  * This is detected via the HAMMER2_CHAIN_DUPLICATED flag in chain->flags.
 342  * To avoid executing live filesystem operations on stale chains, the inode
 343  * locking code will follow stale chains via core->ownerq until it finds
 344  * the live chain.  The lock prevents ripups by other threads.  Lookups
 345  * must properly order locking operations to prevent other threads from
 346  * racing the lookup operation and will also follow stale chains when
 347  * required.
 348  */
 349
 350 RB_HEAD(hammer2_inode_tree, hammer2_inode);
 351
 352 /*
 353  * A hammer2 inode.
 354  *
 355  * NOTE: The inode's attribute CST which is also used to lock the inode
 356  *       is embedded in the chain (chain.cst) and aliased w/ attr_cst.
 357  */
 358 struct hammer2_inode {
 359         RB_ENTRY(hammer2_inode) rbnode;         /* inumber lookup (HL) */
 360         ccms_cst_t              topo_cst;       /* directory topology cst */
 361         struct hammer2_pfsmount *pmp;           /* PFS mount */
 362         struct hammer2_inode    *pip;           /* parent inode */
 363         struct vnode            *vp;
 364         hammer2_chain_t         *chain;         /* NOTE: rehomed on rename */
 365         struct lockf            advlock;
 366         hammer2_tid_t           inum;
 367         u_int                   flags;
 368         u_int                   refs;           /* +vpref, +flushref */
 369         uint8_t                 comp_heuristic;
 370         hammer2_off_t           size;
 371         uint64_t                mtime;
 372 };
 373
 374 typedef struct hammer2_inode hammer2_inode_t;
 375
 376 #define HAMMER2_INODE_MODIFIED          0x0001
 377 #define HAMMER2_INODE_SROOT             0x0002  /* kmalloc special case */
 378 #define HAMMER2_INODE_RENAME_INPROG     0x0004
 379 #define HAMMER2_INODE_ONRBTREE          0x0008
 380 #define HAMMER2_INODE_RESIZED           0x0010
 381 #define HAMMER2_INODE_MTIME             0x0020
 382
 383 int hammer2_inode_cmp(hammer2_inode_t *ip1, hammer2_inode_t *ip2);
 384 RB_PROTOTYPE2(hammer2_inode_tree, hammer2_inode, rbnode, hammer2_inode_cmp,
 385                 hammer2_tid_t);
 386
 387 /*
 388  * A hammer2 transaction and flush sequencing structure.
 389  *
 390  * This global structure is tied into hammer2_mount and is used
 391  * to sequence modifying operations and flushes.
 392  *
 393  * (a) Any modifying operations with sync_tid >= flush_tid will stall until
 394  *     all modifying operating with sync_tid < flush_tid complete.
 395  *
 396  *     The flush related to flush_tid stalls until all modifying operations
 397  *     with sync_tid < flush_tid complete.
 398  *
 399  * (b) Once unstalled, modifying operations with sync_tid > flush_tid are
 400  *     allowed to run.  All modifications cause modify/duplicate operations
 401  *     to occur on the related chains.  Note that most INDIRECT blocks will
 402  *     be unaffected because the modifications just overload the RBTREE
 403  *     structurally instead of actually modifying the indirect blocks.
 404  *
 405  * (c) The actual flush unstalls and RUNS CONCURRENTLY with (b), but only
 406  *     utilizes the chain structures with sync_tid <= flush_tid.  The
 407  *     flush will modify related indirect blocks and inodes in-place
 408  *     (rather than duplicate) since the adjustments are compatible with
 409  *     (b)'s RBTREE overloading
 410  *
 411  *     SPECIAL NOTE:  Inode modifications have to also propagate along any
 412  *                    modify/duplicate chains.  File writes detect the flush
 413  *                    and force out the conflicting buffer cache buffer(s)
 414  *                    before reusing them.
 415  *
 416  * (d) Snapshots can be made instantly but must be flushed and disconnected
 417  *     from their duplicative source before they can be mounted.  This is
 418  *     because while H2's on-media structure supports forks, its in-memory
 419  *     structure only supports very simple forking for background flushing
 420  *     purposes.
 421  *
 422  * TODO: Flush merging.  When fsync() is called on multiple discrete files
 423  *       concurrently there is no reason to stall the second fsync.
 424  *       The final flush that reaches to root can cover both fsync()s.
 425  *
 426  *     The chains typically terminate as they fly onto the disk.  The flush
 427  *     ultimately reaches the volume header.
 428  */
 429 struct hammer2_trans {
 430         TAILQ_ENTRY(hammer2_trans) entry;
 431         struct hammer2_pfsmount *pmp;
 432         hammer2_tid_t           real_tid;
 433         hammer2_tid_t           sync_tid;
 434         hammer2_tid_t           inode_tid;
 435         thread_t                td;             /* pointer */
 436         int                     flags;
 437         int                     blocked;
 438         uint8_t                 inodes_created;
 439         uint8_t                 dummy[7];
 440 };
 441
 442 typedef struct hammer2_trans hammer2_trans_t;
 443
 444 #define HAMMER2_TRANS_ISFLUSH           0x0001  /* formal flush */
 445 #define HAMMER2_TRANS_RESTRICTED        0x0002  /* snapshot flush restrict */
 446 #define HAMMER2_TRANS_BUFCACHE          0x0004  /* from bioq strategy write */
 447 #define HAMMER2_TRANS_NEWINODE          0x0008  /* caller allocating inode */
 448 #define HAMMER2_TRANS_ISALLOCATING      0x0010  /* in allocator */
 449
 450 #define HAMMER2_FREEMAP_HEUR_NRADIX     4       /* pwr 2 PBUFRADIX-MINIORADIX */
 451 #define HAMMER2_FREEMAP_HEUR_TYPES      8
 452 #define HAMMER2_FREEMAP_HEUR            (HAMMER2_FREEMAP_HEUR_NRADIX * \
 453                                          HAMMER2_FREEMAP_HEUR_TYPES)
 454
 455 /*
 456  * Global (per device) mount structure for device (aka vp->v_mount->hmp)
 457  */
 458 TAILQ_HEAD(hammer2_trans_queue, hammer2_trans);
 459
 460 struct hammer2_mount {
 461         struct vnode    *devvp;         /* device vnode */
 462         int             ronly;          /* read-only mount */
 463         int             pmp_count;      /* PFS mounts backed by us */
 464         TAILQ_ENTRY(hammer2_mount) mntentry; /* hammer2_mntlist */
 465
 466         struct malloc_type *mchain;
 467         int             nipstacks;
 468         int             maxipstacks;
 469         hammer2_chain_t vchain;         /* anchor chain (topology) */
 470         hammer2_chain_t fchain;         /* anchor chain (freemap) */
 471         hammer2_inode_t *sroot;         /* super-root localized to media */
 472         struct lock     alloclk;        /* lockmgr lock */
 473         struct lock     voldatalk;      /* lockmgr lock */
 474         struct hammer2_trans_queue transq; /* all in-progress transactions */
 475         hammer2_off_t   heur_freemap[HAMMER2_FREEMAP_HEUR];
 476         int             flushcnt;       /* #of flush trans on the list */
 477
 478         int             volhdrno;       /* last volhdrno written */
 479         hammer2_volume_data_t voldata;
 480         hammer2_volume_data_t volsync;  /* synchronized voldata */
 481 };
 482
 483 typedef struct hammer2_mount hammer2_mount_t;
 484
 485 /*
 486  * HAMMER2 cluster - a device/root associated with a PFS.
 487  *
 488  * A PFS may have several hammer2_cluster's associated with it.
 489  */
 490 #define HAMMER2_MAXCLUSTER      8
 491
 492 struct hammer2_cluster {
 493         int                     nchains;
 494         int                     status;
 495         hammer2_chain_t         *chains[HAMMER2_MAXCLUSTER];
 496 };
 497
 498 typedef struct hammer2_cluster hammer2_cluster_t;
 499
 500 /*
 501  * HAMMER2 PFS mount point structure (aka vp->v_mount->mnt_data).
 502  * This has a 1:1 correspondence to struct mount (note that the
 503  * hammer2_mount structure has a N:1 correspondence).
 504  *
 505  * This structure represents a cluster mount and not necessarily a
 506  * PFS under a specific device mount (HMP).  The distinction is important
 507  * because the elements backing a cluster mount can change on the fly.
 508  *
 509  * Usually the first element under the cluster represents the original
 510  * user-requested mount that bootstraps the whole mess.  In significant
 511  * setups the original is usually just a read-only media image (or
 512  * representitive file) that simply contains a bootstrap volume header
 513  * listing the configuration.
 514  */
 515 struct hammer2_pfsmount {
 516         struct mount            *mp;
 517         hammer2_cluster_t       cluster;
 518         hammer2_inode_t         *iroot;         /* PFS root inode */
 519         hammer2_off_t           inode_count;    /* copy of inode_count */
 520         ccms_domain_t           ccms_dom;
 521         struct netexport        export;         /* nfs export */
 522         int                     ronly;          /* read-only mount */
 523         struct malloc_type      *minode;
 524         struct malloc_type      *mmsg;
 525         kdmsg_iocom_t           iocom;
 526         struct spinlock         inum_spin;      /* inumber lookup */
 527         struct hammer2_inode_tree inum_tree;
 528         long                    inmem_inodes;
 529         long                    inmem_chains;
 530         int                     inmem_waiting;
 531         thread_t                wthread_td;     /* write thread td */
 532         struct bio_queue_head   wthread_bioq;   /* logical buffer bioq */
 533         struct mtx              wthread_mtx;    /* interlock */
 534         int                     wthread_destroy;/* termination sequencing */
 535 };
 536
 537 typedef struct hammer2_pfsmount hammer2_pfsmount_t;
 538
 539 struct hammer2_cbinfo {
 540         hammer2_chain_t *chain;
 541         void (*func)(hammer2_chain_t *, struct buf *, char *, void *);
 542         void *arg;
 543         size_t boff;
 544 };
 545
 546 typedef struct hammer2_cbinfo hammer2_cbinfo_t;
 547
 548 #if defined(_KERNEL)
 549
 550 MALLOC_DECLARE(M_HAMMER2);
 551
 552 #define VTOI(vp)        ((hammer2_inode_t *)(vp)->v_data)
 553 #define ITOV(ip)        ((ip)->vp)
 554
 555 /*
 556  * Currently locked chains retain the locked buffer cache buffer for
 557  * indirect blocks, and indirect blocks can be one of two sizes.  The
 558  * device buffer has to match the case to avoid deadlocking recursive
 559  * chains that might otherwise try to access different offsets within
 560  * the same device buffer.
 561  */
 562 static __inline
 563 int
 564 hammer2_devblkradix(int radix)
 565 {
 566 #if 1
 567         if (radix <= HAMMER2_LBUFRADIX) {
 568                 return (HAMMER2_LBUFRADIX);
 569         } else {
 570                 return (HAMMER2_PBUFRADIX);
 571         }
 572 #else
 573         return (HAMMER2_PBUFRADIX);
 574 #endif
 575 }
 576
 577 static __inline
 578 size_t
 579 hammer2_devblksize(size_t bytes)
 580 {
 581 #if 1
 582         if (bytes <= HAMMER2_LBUFSIZE) {
 583                 return(HAMMER2_LBUFSIZE);
 584         } else {
 585                 KKASSERT(bytes <= HAMMER2_PBUFSIZE &&
 586                          (bytes ^ (bytes - 1)) == ((bytes << 1) - 1));
 587                 return (HAMMER2_PBUFSIZE);
 588         }
 589 #else
 590         KKASSERT(bytes <= HAMMER2_PBUFSIZE &&
 591                  (bytes ^ (bytes - 1)) == ((bytes << 1) - 1));
 592         return(HAMMER2_PBUFSIZE);
 593 #endif
 594 }
 595
 596
 597 static __inline
 598 hammer2_pfsmount_t *
 599 MPTOPMP(struct mount *mp)
 600 {
 601         return ((hammer2_pfsmount_t *)mp->mnt_data);
 602 }
 603
 604 extern struct vop_ops hammer2_vnode_vops;
 605 extern struct vop_ops hammer2_spec_vops;
 606 extern struct vop_ops hammer2_fifo_vops;
 607
 608 extern int hammer2_debug;
 609 extern int hammer2_cluster_enable;
 610 extern int hammer2_hardlink_enable;
 611 extern long hammer2_iod_file_read;
 612 extern long hammer2_iod_meta_read;
 613 extern long hammer2_iod_indr_read;
 614 extern long hammer2_iod_fmap_read;
 615 extern long hammer2_iod_volu_read;
 616 extern long hammer2_iod_file_write;
 617 extern long hammer2_iod_meta_write;
 618 extern long hammer2_iod_indr_write;
 619 extern long hammer2_iod_fmap_write;
 620 extern long hammer2_iod_volu_write;
 621 extern long hammer2_ioa_file_read;
 622 extern long hammer2_ioa_meta_read;
 623 extern long hammer2_ioa_indr_read;
 624 extern long hammer2_ioa_fmap_read;
 625 extern long hammer2_ioa_volu_read;
 626 extern long hammer2_ioa_file_write;
 627 extern long hammer2_ioa_meta_write;
 628 extern long hammer2_ioa_indr_write;
 629 extern long hammer2_ioa_fmap_write;
 630 extern long hammer2_ioa_volu_write;
 631
 632 extern struct objcache *cache_buffer_read;
 633 extern struct objcache *cache_buffer_write;
 634
 635 extern int destroy;
 636 extern int write_thread_wakeup;
 637
 638 extern mtx_t thread_protect;
 639
 640 /*
 641  * hammer2_subr.c
 642  */
 643 #define hammer2_icrc32(buf, size)       iscsi_crc32((buf), (size))
 644 #define hammer2_icrc32c(buf, size, crc) iscsi_crc32_ext((buf), (size), (crc))
 645
 646 hammer2_chain_t *hammer2_inode_lock_ex(hammer2_inode_t *ip);
 647 hammer2_chain_t *hammer2_inode_lock_sh(hammer2_inode_t *ip);
 648 void hammer2_inode_unlock_ex(hammer2_inode_t *ip, hammer2_chain_t *chain);
 649 void hammer2_inode_unlock_sh(hammer2_inode_t *ip, hammer2_chain_t *chain);
 650 void hammer2_chain_refactor(hammer2_chain_t **chainp);
 651 void hammer2_voldata_lock(hammer2_mount_t *hmp);
 652 void hammer2_voldata_unlock(hammer2_mount_t *hmp, int modify);
 653 ccms_state_t hammer2_inode_lock_temp_release(hammer2_inode_t *ip);
 654 void hammer2_inode_lock_temp_restore(hammer2_inode_t *ip, ccms_state_t ostate);
 655 ccms_state_t hammer2_inode_lock_upgrade(hammer2_inode_t *ip);
 656 void hammer2_inode_lock_downgrade(hammer2_inode_t *ip, ccms_state_t ostate);
 657
 658 void hammer2_mount_exlock(hammer2_mount_t *hmp);
 659 void hammer2_mount_shlock(hammer2_mount_t *hmp);
 660 void hammer2_mount_unlock(hammer2_mount_t *hmp);
 661
 662 int hammer2_get_dtype(hammer2_chain_t *chain);
 663 int hammer2_get_vtype(hammer2_chain_t *chain);
 664 u_int8_t hammer2_get_obj_type(enum vtype vtype);
 665 void hammer2_time_to_timespec(u_int64_t xtime, struct timespec *ts);
 666 u_int64_t hammer2_timespec_to_time(struct timespec *ts);
 667 u_int32_t hammer2_to_unix_xid(uuid_t *uuid);
 668 void hammer2_guid_to_uuid(uuid_t *uuid, u_int32_t guid);
 669
 670 hammer2_key_t hammer2_dirhash(const unsigned char *name, size_t len);
 671 int hammer2_getradix(size_t bytes);
 672
 673 int hammer2_calc_logical(hammer2_inode_t *ip, hammer2_off_t uoff,
 674                         hammer2_key_t *lbasep, hammer2_key_t *leofp);
 675 int hammer2_calc_physical(hammer2_inode_t *ip, hammer2_key_t lbase);
 676 void hammer2_update_time(uint64_t *timep);
 677
 678 /*
 679  * hammer2_inode.c
 680  */
 681 struct vnode *hammer2_igetv(hammer2_inode_t *ip, int *errorp);
 682
 683 void hammer2_inode_lock_nlinks(hammer2_inode_t *ip);
 684 void hammer2_inode_unlock_nlinks(hammer2_inode_t *ip);
 685 hammer2_inode_t *hammer2_inode_lookup(hammer2_pfsmount_t *pmp,
 686                         hammer2_tid_t inum);
 687 hammer2_inode_t *hammer2_inode_get(hammer2_pfsmount_t *pmp,
 688                         hammer2_inode_t *dip, hammer2_chain_t *chain);
 689 void hammer2_inode_free(hammer2_inode_t *ip);
 690 void hammer2_inode_ref(hammer2_inode_t *ip);
 691 void hammer2_inode_drop(hammer2_inode_t *ip);
 692 void hammer2_inode_repoint(hammer2_inode_t *ip, hammer2_inode_t *pip,
 693                         hammer2_chain_t *chain);
 694
 695 hammer2_inode_t *hammer2_inode_create(hammer2_trans_t *trans,
 696                         hammer2_inode_t *dip,
 697                         struct vattr *vap, struct ucred *cred,
 698                         const uint8_t *name, size_t name_len,
 699                         hammer2_chain_t **chainp, int *errorp);
 700 int hammer2_inode_connect(hammer2_trans_t *trans, int hlink,
 701                         hammer2_inode_t *dip, hammer2_chain_t **chainp,
 702                         const uint8_t *name, size_t name_len);
 703 hammer2_inode_t *hammer2_inode_common_parent(hammer2_inode_t *fdip,
 704                         hammer2_inode_t *tdip);
 705 void hammer2_inode_fsync(hammer2_trans_t *trans, hammer2_inode_t *ip,
 706                         hammer2_chain_t **parentp);
 707 int hammer2_unlink_file(hammer2_trans_t *trans, hammer2_inode_t *dip,
 708                         const uint8_t *name, size_t name_len, int isdir,
 709                         int *hlinkp);
 710 int hammer2_hardlink_consolidate(hammer2_trans_t *trans, hammer2_inode_t *ip,
 711                         hammer2_chain_t **chainp,
 712                         hammer2_inode_t *tdip, int linkcnt);
 713 int hammer2_hardlink_deconsolidate(hammer2_trans_t *trans, hammer2_inode_t *dip,
 714                         hammer2_chain_t **chainp, hammer2_chain_t **ochainp);
 715 int hammer2_hardlink_find(hammer2_inode_t *dip,
 716                         hammer2_chain_t **chainp, hammer2_chain_t **ochainp);
 717
 718 /*
 719  * hammer2_chain.c
 720  */
 721 void hammer2_modify_volume(hammer2_mount_t *hmp);
 722 hammer2_chain_t *hammer2_chain_alloc(hammer2_mount_t *hmp, hammer2_pfsmount_t *pmp,
 723                                 hammer2_trans_t *trans, hammer2_blockref_t *bref);
 724 void hammer2_chain_core_alloc(hammer2_trans_t *trans, hammer2_chain_t *nchain,
 725                                 hammer2_chain_t *ochain);
 726 void hammer2_chain_ref(hammer2_chain_t *chain);
 727 void hammer2_chain_drop(hammer2_chain_t *chain);
 728 int hammer2_chain_lock(hammer2_chain_t *chain, int how);
 729 void hammer2_chain_load_async(hammer2_chain_t *chain,
 730                                 void (*func)(hammer2_chain_t *, struct buf *,
 731                                              char *, void *),
 732                                 void *arg);
 733 void hammer2_chain_moved(hammer2_chain_t *chain);
 734 void hammer2_chain_modify(hammer2_trans_t *trans,
 735                                 hammer2_chain_t **chainp, int flags);
 736 hammer2_inode_data_t *hammer2_chain_modify_ip(hammer2_trans_t *trans,
 737                                 hammer2_inode_t *ip, hammer2_chain_t **chainp,
 738                                 int flags);
 739 void hammer2_chain_resize(hammer2_trans_t *trans, hammer2_inode_t *ip,
 740                                 hammer2_chain_t *parent,
 741                                 hammer2_chain_t **chainp,
 742                                 int nradix, int flags);
 743 void hammer2_chain_unlock(hammer2_chain_t *chain);
 744 void hammer2_chain_wait(hammer2_chain_t *chain);
 745 hammer2_chain_t *hammer2_chain_get(hammer2_chain_t *parent,
 746                                 hammer2_blockref_t *bref);
 747 hammer2_chain_t *hammer2_chain_lookup_init(hammer2_chain_t *parent, int flags);
 748 void hammer2_chain_lookup_done(hammer2_chain_t *parent);
 749 hammer2_chain_t *hammer2_chain_lookup(hammer2_chain_t **parentp,
 750                                 hammer2_key_t *key_nextp,
 751                                 hammer2_key_t key_beg, hammer2_key_t key_end,
 752                                 int *cache_indexp, int flags);
 753 hammer2_chain_t *hammer2_chain_next(hammer2_chain_t **parentp,
 754                                 hammer2_chain_t *chain,
 755                                 hammer2_key_t *key_nextp,
 756                                 hammer2_key_t key_beg, hammer2_key_t key_end,
 757                                 int *cache_indexp, int flags);
 758
 759 int hammer2_chain_create(hammer2_trans_t *trans,
 760                                 hammer2_chain_t **parentp,
 761                                 hammer2_chain_t **chainp,
 762                                 hammer2_key_t key, int keybits,
 763                                 int type, size_t bytes);
 764 void hammer2_chain_duplicate(hammer2_trans_t *trans, hammer2_chain_t **parentp,
 765                                 hammer2_chain_t **chainp,
 766                                 hammer2_blockref_t *bref, int snapshot);
 767 int hammer2_chain_snapshot(hammer2_trans_t *trans, hammer2_chain_t **chainp,
 768                                 hammer2_ioc_pfs_t *pfs);
 769 void hammer2_chain_delete(hammer2_trans_t *trans, hammer2_chain_t *chain,
 770                                 int flags);
 771 void hammer2_chain_delete_duplicate(hammer2_trans_t *trans,
 772                                 hammer2_chain_t **chainp, int flags);
 773 void hammer2_chain_flush(hammer2_trans_t *trans, hammer2_chain_t **chainp);
 774 void hammer2_chain_commit(hammer2_trans_t *trans, hammer2_chain_t *chain);
 775 void hammer2_chain_setsubmod(hammer2_trans_t *trans, hammer2_chain_t *chain);
 776
 777 void hammer2_chain_memory_wait(hammer2_pfsmount_t *pmp);
 778 void hammer2_chain_memory_wakeup(hammer2_pfsmount_t *pmp);
 779 void hammer2_chain_countbrefs(hammer2_chain_t *chain,
 780                                 hammer2_blockref_t *base, int count);
 781 void hammer2_chain_layer_check_locked(hammer2_mount_t *hmp,
 782                                 hammer2_chain_core_t *core);
 783
 784 int hammer2_base_find(hammer2_chain_t *chain,
 785                                 hammer2_blockref_t *base, int count,
 786                                 int *cache_indexp, hammer2_key_t *key_nextp,
 787                                 hammer2_key_t key_beg, hammer2_key_t key_end);
 788 void hammer2_base_delete(hammer2_chain_t *chain,
 789                                 hammer2_blockref_t *base, int count,
 790                                 int *cache_indexp, hammer2_chain_t *child);
 791 void hammer2_base_insert(hammer2_chain_t *chain,
 792                                 hammer2_blockref_t *base, int count,
 793                                 int *cache_indexp, hammer2_chain_t *child);
 794
 795 /*
 796  * hammer2_trans.c
 797  */
 798 void hammer2_trans_init(hammer2_trans_t *trans,
 799                         hammer2_pfsmount_t *pmp, int flags);
 800 void hammer2_trans_clear_invfsync(hammer2_trans_t *trans);
 801 void hammer2_trans_done(hammer2_trans_t *trans);
 802
 803 /*
 804  * hammer2_ioctl.c
 805  */
 806 int hammer2_ioctl(hammer2_inode_t *ip, u_long com, void *data,
 807                                 int fflag, struct ucred *cred);
 808
 809 /*
 810  * hammer2_msgops.c
 811  */
 812 int hammer2_msg_dbg_rcvmsg(kdmsg_msg_t *msg);
 813 int hammer2_msg_adhoc_input(kdmsg_msg_t *msg);
 814
 815 /*
 816  * hammer2_vfsops.c
 817  */
 818 void hammer2_clusterctl_wakeup(kdmsg_iocom_t *iocom);
 819 void hammer2_volconf_update(hammer2_pfsmount_t *pmp, int index);
 820 void hammer2_cluster_reconnect(hammer2_pfsmount_t *pmp, struct file *fp);
 821 void hammer2_dump_chain(hammer2_chain_t *chain, int tab, int *countp);
 822 void hammer2_bioq_sync(hammer2_pfsmount_t *pmp);
 823 int hammer2_vfs_sync(struct mount *mp, int waitflags);
 824
 825 /*
 826  * hammer2_freemap.c
 827  */
 828 int hammer2_freemap_alloc(hammer2_trans_t *trans, hammer2_mount_t *hmp,
 829                                 hammer2_blockref_t *bref, size_t bytes);
 830 void hammer2_freemap_free(hammer2_trans_t *trans, hammer2_mount_t *hmp,
 831                                 hammer2_blockref_t *bref, int how);
 832
 833
 834 #endif /* !_KERNEL */
 835 #endif /* !_VFS_HAMMER2_HAMMER2_H_ */