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;    /* indexed by device offset */
 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         int             act;                    /* activity */
 181 };
 182
 183 typedef struct hammer2_io hammer2_io_t;
 184
 185 /*
 186  * Primary chain structure keeps track of the topology in-memory.
 187  */
 188 struct hammer2_chain {
 189         RB_ENTRY(hammer2_chain) rbnode;         /* node */
 190         TAILQ_ENTRY(hammer2_chain) core_entry;  /* contemporary chains */
 191         hammer2_chain_layer_t   *inlayer;
 192         hammer2_blockref_t      bref;
 193         hammer2_chain_core_t    *core;
 194         hammer2_chain_core_t    *above;
 195         struct hammer2_state    *state;         /* if active cache msg */
 196         struct hammer2_mount    *hmp;
 197         struct hammer2_pfsmount *pmp;           /* can be NULL */
 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         hammer2_media_data_t *data;             /* data pointer shortcut */
 209         TAILQ_ENTRY(hammer2_chain) flush_node;  /* flush deferral list */
 210
 211         int             inode_reason;
 212 };
 213
 214 typedef struct hammer2_chain hammer2_chain_t;
 215
 216 int hammer2_chain_cmp(hammer2_chain_t *chain1, hammer2_chain_t *chain2);
 217 RB_PROTOTYPE(hammer2_chain_tree, hammer2_chain, rbnode, hammer2_chain_cmp);
 218
 219 /*
 220  * Special notes on flags:
 221  *
 222  * INITIAL - This flag allows a chain to be created and for storage to
 223  *           be allocated without having to immediately instantiate the
 224  *           related buffer.  The data is assumed to be all-zeros.  It
 225  *           is primarily used for indirect blocks.
 226  *
 227  * MOVED   - A modified chain becomes MOVED after it flushes.  A chain
 228  *           can also become MOVED if it is moved within the topology
 229  *           (even if not modified).
 230  *
 231  * MODIFIED- The chain's media data has been modified.
 232  */
 233 #define HAMMER2_CHAIN_MODIFIED          0x00000001      /* dirty chain data */
 234 #define HAMMER2_CHAIN_ALLOCATED         0x00000002      /* kmalloc'd chain */
 235 #define HAMMER2_CHAIN_UNUSED0004        0x00000004
 236 #define HAMMER2_CHAIN_FORCECOW          0x00000008      /* force copy-on-wr */
 237 #define HAMMER2_CHAIN_DELETED           0x00000010      /* deleted chain */
 238 #define HAMMER2_CHAIN_INITIAL           0x00000020      /* initial create */
 239 #define HAMMER2_CHAIN_FLUSHED           0x00000040      /* blktable updated */
 240 #define HAMMER2_CHAIN_MOVED             0x00000080      /* bref changed */
 241 #define HAMMER2_CHAIN_IOFLUSH           0x00000100      /* bawrite on put */
 242 #define HAMMER2_CHAIN_DEFERRED          0x00000200      /* on a deferral list */
 243 #define HAMMER2_CHAIN_UNLINKED          0x00000400      /* delete on reclaim */
 244 #define HAMMER2_CHAIN_VOLUMESYNC        0x00000800      /* needs volume sync */
 245 #define HAMMER2_CHAIN_UNUSED01000       0x00001000
 246 #define HAMMER2_CHAIN_MOUNTED           0x00002000      /* PFS is mounted */
 247 #define HAMMER2_CHAIN_ONRBTREE          0x00004000      /* on parent RB tree */
 248 #define HAMMER2_CHAIN_SNAPSHOT          0x00008000      /* snapshot special */
 249 #define HAMMER2_CHAIN_EMBEDDED          0x00010000      /* embedded data */
 250 #define HAMMER2_CHAIN_RELEASE           0x00020000      /* don't keep around */
 251 #define HAMMER2_CHAIN_UNUSED40000       0x00040000
 252 #define HAMMER2_CHAIN_UNUSED80000       0x00080000
 253 #define HAMMER2_CHAIN_DUPLICATED        0x00100000      /* fwd delete-dup */
 254 #define HAMMER2_CHAIN_PFSROOT           0x00200000      /* in pfs->cluster */
 255
 256 /*
 257  * Flags passed to hammer2_chain_lookup() and hammer2_chain_next()
 258  *
 259  * NOTE: MATCHIND allows an indirect block / freemap node to be returned
 260  *       when the passed key range matches the radix.  Remember that key_end
 261  *       is inclusive (e.g. {0x000,0xFFF}, not {0x000,0x1000}).
 262  */
 263 #define HAMMER2_LOOKUP_NOLOCK           0x00000001      /* ref only */
 264 #define HAMMER2_LOOKUP_NODATA           0x00000002      /* data left NULL */
 265 #define HAMMER2_LOOKUP_SHARED           0x00000100
 266 #define HAMMER2_LOOKUP_MATCHIND         0x00000200      /* return all chains */
 267 #define HAMMER2_LOOKUP_UNUSED0400       0x00000400
 268 #define HAMMER2_LOOKUP_ALWAYS           0x00000800      /* resolve data */
 269
 270 /*
 271  * Flags passed to hammer2_chain_modify() and hammer2_chain_resize()
 272  *
 273  * NOTE: OPTDATA allows us to avoid instantiating buffers for INDIRECT
 274  *       blocks in the INITIAL-create state.
 275  */
 276 #define HAMMER2_MODIFY_OPTDATA          0x00000002      /* data can be NULL */
 277 #define HAMMER2_MODIFY_NO_MODIFY_TID    0x00000004
 278 #define HAMMER2_MODIFY_ASSERTNOCOPY     0x00000008      /* assert no del-dup */
 279 #define HAMMER2_MODIFY_NOREALLOC        0x00000010
 280 #define HAMMER2_MODIFY_INPLACE          0x00000020      /* don't del-dup */
 281
 282 /*
 283  * Flags passed to hammer2_chain_lock()
 284  */
 285 #define HAMMER2_RESOLVE_NEVER           1
 286 #define HAMMER2_RESOLVE_MAYBE           2
 287 #define HAMMER2_RESOLVE_ALWAYS          3
 288 #define HAMMER2_RESOLVE_MASK            0x0F
 289
 290 #define HAMMER2_RESOLVE_SHARED          0x10    /* request shared lock */
 291 #define HAMMER2_RESOLVE_NOREF           0x20    /* already ref'd on lock */
 292
 293 /*
 294  * Flags passed to hammer2_chain_delete()
 295  */
 296 #define HAMMER2_DELETE_UNUSED0001       0x0001
 297
 298 /*
 299  * Flags passed to hammer2_chain_delete_duplicate()
 300  */
 301 #define HAMMER2_DELDUP_RECORE           0x0001
 302
 303 /*
 304  * Cluster different types of storage together for allocations
 305  */
 306 #define HAMMER2_FREECACHE_INODE         0
 307 #define HAMMER2_FREECACHE_INDIR         1
 308 #define HAMMER2_FREECACHE_DATA          2
 309 #define HAMMER2_FREECACHE_UNUSED3       3
 310 #define HAMMER2_FREECACHE_TYPES         4
 311
 312 /*
 313  * hammer2_freemap_alloc() block preference
 314  */
 315 #define HAMMER2_OFF_NOPREF              ((hammer2_off_t)-1)
 316
 317 /*
 318  * BMAP read-ahead maximum parameters
 319  */
 320 #define HAMMER2_BMAP_COUNT              16      /* max bmap read-ahead */
 321 #define HAMMER2_BMAP_BYTES              (HAMMER2_PBUFSIZE * HAMMER2_BMAP_COUNT)
 322
 323 /*
 324  * Misc
 325  */
 326 #define HAMMER2_FLUSH_DEPTH_LIMIT       10      /* stack recursion limit */
 327
 328 /*
 329  * hammer2_freemap_adjust()
 330  */
 331 #define HAMMER2_FREEMAP_DORECOVER       1
 332 #define HAMMER2_FREEMAP_DOMAYFREE       2
 333 #define HAMMER2_FREEMAP_DOREALFREE      3
 334
 335 /*
 336  * HAMMER2 IN-MEMORY CACHE OF MEDIA STRUCTURES
 337  *
 338  * There is an in-memory representation of all on-media data structure.
 339  * Basically everything is represented by a hammer2_chain structure
 340  * in-memory and other higher-level structures map to chains.
 341  *
 342  * A great deal of data is accessed simply via its buffer cache buffer,
 343  * which is mapped for the duration of the chain's lock.  However, because
 344  * chains may represent blocks smaller than the 16KB minimum we impose
 345  * on buffer cache buffers, we cannot hold related buffer cache buffers
 346  * locked for smaller blocks.  In these situations we kmalloc() a copy
 347  * of the block.
 348  *
 349  * When modifications are made to a chain a new filesystem block must be
 350  * allocated.  Multiple modifications do not necessarily allocate new
 351  * blocks.  However, when a flush occurs a flush synchronization point
 352  * is created and any new modifications made after this point will allocate
 353  * a new block even if the chain is already in a modified state.
 354  *
 355  * The in-memory representation may remain cached (for example in order to
 356  * placemark clustering locks) even after the related data has been
 357  * detached.
 358  *
 359  *                              CORE SHARING
 360  *
 361  * In order to support concurrent flushes a flush synchronization point
 362  * is created represented by a transaction id.  Among other things,
 363  * operations may move filesystem objects from one part of the topology
 364  * to another (for example, if you rename a file or when indirect blocks
 365  * are created or destroyed, and a few other things).  When this occurs
 366  * across a flush synchronization point the flusher needs to be able to
 367  * recurse down BOTH the 'before' version of the topology and the 'after'
 368  * version.
 369  *
 370  * To facilitate this modifications to chains do what is called a
 371  * DELETE-DUPLICATE operation.  Chains are not actually moved in-memory.
 372  * Instead the chain we wish to move is deleted and a new chain is created
 373  * at the target location in the topology.  ANY SUBCHAINS PLACED UNDER THE
 374  * CHAIN BEING MOVED HAVE TO EXIST IN BOTH PLACES.  To make this work
 375  * all sub-chains are managed by the hammer2_chain_core structure.  This
 376  * structure can be multi-homed, meaning that it can have more than one
 377  * chain as its parent.  When a chain is delete-duplicated the chain's core
 378  * becomes shared under both the old and new chain.
 379  *
 380  *                              STALE CHAINS
 381  *
 382  * When a chain is delete-duplicated the old chain typically becomes stale.
 383  * This is detected via the HAMMER2_CHAIN_DUPLICATED flag in chain->flags.
 384  * To avoid executing live filesystem operations on stale chains, the inode
 385  * locking code will follow stale chains via core->ownerq until it finds
 386  * the live chain.  The lock prevents ripups by other threads.  Lookups
 387  * must properly order locking operations to prevent other threads from
 388  * racing the lookup operation and will also follow stale chains when
 389  * required.
 390  */
 391
 392 RB_HEAD(hammer2_inode_tree, hammer2_inode);
 393
 394 /*
 395  * A hammer2 inode.
 396  *
 397  * NOTE: The inode's attribute CST which is also used to lock the inode
 398  *       is embedded in the chain (chain.cst) and aliased w/ attr_cst.
 399  */
 400 struct hammer2_inode {
 401         RB_ENTRY(hammer2_inode) rbnode;         /* inumber lookup (HL) */
 402         ccms_cst_t              topo_cst;       /* directory topology cst */
 403         struct hammer2_pfsmount *pmp;           /* PFS mount */
 404         struct hammer2_inode    *pip;           /* parent inode */
 405         struct vnode            *vp;
 406         hammer2_chain_t         *chain;         /* NOTE: rehomed on rename */
 407         struct lockf            advlock;
 408         hammer2_tid_t           inum;
 409         u_int                   flags;
 410         u_int                   refs;           /* +vpref, +flushref */
 411         uint8_t                 comp_heuristic;
 412         hammer2_off_t           size;
 413         uint64_t                mtime;
 414 };
 415
 416 typedef struct hammer2_inode hammer2_inode_t;
 417
 418 #define HAMMER2_INODE_MODIFIED          0x0001
 419 #define HAMMER2_INODE_SROOT             0x0002  /* kmalloc special case */
 420 #define HAMMER2_INODE_RENAME_INPROG     0x0004
 421 #define HAMMER2_INODE_ONRBTREE          0x0008
 422 #define HAMMER2_INODE_RESIZED           0x0010
 423 #define HAMMER2_INODE_MTIME             0x0020
 424
 425 int hammer2_inode_cmp(hammer2_inode_t *ip1, hammer2_inode_t *ip2);
 426 RB_PROTOTYPE2(hammer2_inode_tree, hammer2_inode, rbnode, hammer2_inode_cmp,
 427                 hammer2_tid_t);
 428
 429 /*
 430  * A hammer2 transaction and flush sequencing structure.
 431  *
 432  * This global structure is tied into hammer2_mount and is used
 433  * to sequence modifying operations and flushes.
 434  *
 435  * (a) Any modifying operations with sync_tid >= flush_tid will stall until
 436  *     all modifying operating with sync_tid < flush_tid complete.
 437  *
 438  *     The flush related to flush_tid stalls until all modifying operations
 439  *     with sync_tid < flush_tid complete.
 440  *
 441  * (b) Once unstalled, modifying operations with sync_tid > flush_tid are
 442  *     allowed to run.  All modifications cause modify/duplicate operations
 443  *     to occur on the related chains.  Note that most INDIRECT blocks will
 444  *     be unaffected because the modifications just overload the RBTREE
 445  *     structurally instead of actually modifying the indirect blocks.
 446  *
 447  * (c) The actual flush unstalls and RUNS CONCURRENTLY with (b), but only
 448  *     utilizes the chain structures with sync_tid <= flush_tid.  The
 449  *     flush will modify related indirect blocks and inodes in-place
 450  *     (rather than duplicate) since the adjustments are compatible with
 451  *     (b)'s RBTREE overloading
 452  *
 453  *     SPECIAL NOTE:  Inode modifications have to also propagate along any
 454  *                    modify/duplicate chains.  File writes detect the flush
 455  *                    and force out the conflicting buffer cache buffer(s)
 456  *                    before reusing them.
 457  *
 458  * (d) Snapshots can be made instantly but must be flushed and disconnected
 459  *     from their duplicative source before they can be mounted.  This is
 460  *     because while H2's on-media structure supports forks, its in-memory
 461  *     structure only supports very simple forking for background flushing
 462  *     purposes.
 463  *
 464  * TODO: Flush merging.  When fsync() is called on multiple discrete files
 465  *       concurrently there is no reason to stall the second fsync.
 466  *       The final flush that reaches to root can cover both fsync()s.
 467  *
 468  *     The chains typically terminate as they fly onto the disk.  The flush
 469  *     ultimately reaches the volume header.
 470  */
 471 struct hammer2_trans {
 472         TAILQ_ENTRY(hammer2_trans) entry;
 473         struct hammer2_pfsmount *pmp;           /* might be NULL */
 474         struct hammer2_mount    *hmp_single;    /* if single-targetted */
 475         hammer2_tid_t           sync_tid;
 476         hammer2_tid_t           real_tid;
 477         hammer2_tid_t           inode_tid;
 478         thread_t                td;             /* pointer */
 479         int                     flags;
 480         int                     blocked;
 481         uint8_t                 inodes_created;
 482         uint8_t                 dummy[7];
 483 };
 484
 485 typedef struct hammer2_trans hammer2_trans_t;
 486
 487 #define HAMMER2_TRANS_ISFLUSH           0x0001  /* formal flush */
 488 #define HAMMER2_TRANS_UNUSED0002        0x0002
 489 #define HAMMER2_TRANS_BUFCACHE          0x0004  /* from bioq strategy write */
 490 #define HAMMER2_TRANS_NEWINODE          0x0008  /* caller allocating inode */
 491 #define HAMMER2_TRANS_ISALLOCATING      0x0010  /* in allocator */
 492
 493 #define HAMMER2_FREEMAP_HEUR_NRADIX     4       /* pwr 2 PBUFRADIX-MINIORADIX */
 494 #define HAMMER2_FREEMAP_HEUR_TYPES      8
 495 #define HAMMER2_FREEMAP_HEUR            (HAMMER2_FREEMAP_HEUR_NRADIX * \
 496                                          HAMMER2_FREEMAP_HEUR_TYPES)
 497
 498 /*
 499  * Global (per device) mount structure for device (aka vp->v_mount->hmp)
 500  */
 501 TAILQ_HEAD(hammer2_trans_queue, hammer2_trans);
 502
 503 struct hammer2_mount {
 504         struct vnode    *devvp;         /* device vnode */
 505         int             ronly;          /* read-only mount */
 506         int             pmp_count;      /* PFS mounts backed by us */
 507         TAILQ_ENTRY(hammer2_mount) mntentry; /* hammer2_mntlist */
 508
 509         struct malloc_type *mchain;
 510         int             nipstacks;
 511         int             maxipstacks;
 512         struct spinlock io_spin;        /* iotree access */
 513         struct hammer2_io_tree iotree;
 514         int             iofree_count;
 515         hammer2_chain_t vchain;         /* anchor chain (topology) */
 516         hammer2_chain_t fchain;         /* anchor chain (freemap) */
 517         hammer2_inode_t *sroot;         /* super-root localized to media */
 518         struct lock     alloclk;        /* lockmgr lock */
 519         struct lock     voldatalk;      /* lockmgr lock */
 520         struct hammer2_trans_queue transq; /* all in-progress transactions */
 521         hammer2_off_t   heur_freemap[HAMMER2_FREEMAP_HEUR];
 522         int             flushcnt;       /* #of flush trans on the list */
 523
 524         int             volhdrno;       /* last volhdrno written */
 525         hammer2_volume_data_t voldata;
 526         hammer2_volume_data_t volsync;  /* synchronized voldata */
 527 };
 528
 529 typedef struct hammer2_mount hammer2_mount_t;
 530
 531 /*
 532  * HAMMER2 cluster - a device/root associated with a PFS.
 533  *
 534  * A PFS may have several hammer2_cluster's associated with it.
 535  */
 536 #define HAMMER2_MAXCLUSTER      8
 537
 538 struct hammer2_cluster {
 539         int                     nchains;
 540         int                     status;
 541         hammer2_chain_t         *chains[HAMMER2_MAXCLUSTER];
 542 };
 543
 544 typedef struct hammer2_cluster hammer2_cluster_t;
 545
 546 /*
 547  * HAMMER2 PFS mount point structure (aka vp->v_mount->mnt_data).
 548  * This has a 1:1 correspondence to struct mount (note that the
 549  * hammer2_mount structure has a N:1 correspondence).
 550  *
 551  * This structure represents a cluster mount and not necessarily a
 552  * PFS under a specific device mount (HMP).  The distinction is important
 553  * because the elements backing a cluster mount can change on the fly.
 554  *
 555  * Usually the first element under the cluster represents the original
 556  * user-requested mount that bootstraps the whole mess.  In significant
 557  * setups the original is usually just a read-only media image (or
 558  * representitive file) that simply contains a bootstrap volume header
 559  * listing the configuration.
 560  */
 561 struct hammer2_pfsmount {
 562         struct mount            *mp;
 563         hammer2_cluster_t       cluster;
 564         hammer2_inode_t         *iroot;         /* PFS root inode */
 565         hammer2_inode_t         *ihidden;       /* PFS hidden directory */
 566         struct lock             lock;           /* PFS lock for certain ops */
 567         hammer2_off_t           inode_count;    /* copy of inode_count */
 568         ccms_domain_t           ccms_dom;
 569         struct netexport        export;         /* nfs export */
 570         int                     ronly;          /* read-only mount */
 571         struct malloc_type      *minode;
 572         struct malloc_type      *mmsg;
 573         kdmsg_iocom_t           iocom;
 574         struct spinlock         inum_spin;      /* inumber lookup */
 575         struct hammer2_inode_tree inum_tree;
 576         long                    inmem_inodes;
 577         long                    inmem_dirty_chains;
 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_DIRTYCHAIN_WAITING      0x80000000
 588 #define HAMMER2_DIRTYCHAIN_MASK         0x7FFFFFFF
 589
 590 #define HAMMER2_LWINPROG_WAITING        0x80000000
 591 #define HAMMER2_LWINPROG_MASK           0x7FFFFFFF
 592
 593 #if defined(_KERNEL)
 594
 595 MALLOC_DECLARE(M_HAMMER2);
 596
 597 #define VTOI(vp)        ((hammer2_inode_t *)(vp)->v_data)
 598 #define ITOV(ip)        ((ip)->vp)
 599
 600 /*
 601  * Currently locked chains retain the locked buffer cache buffer for
 602  * indirect blocks, and indirect blocks can be one of two sizes.  The
 603  * device buffer has to match the case to avoid deadlocking recursive
 604  * chains that might otherwise try to access different offsets within
 605  * the same device buffer.
 606  */
 607 static __inline
 608 int
 609 hammer2_devblkradix(int radix)
 610 {
 611         if (radix <= HAMMER2_LBUFRADIX) {
 612                 return (HAMMER2_LBUFRADIX);
 613         } else {
 614                 return (HAMMER2_PBUFRADIX);
 615         }
 616 }
 617
 618 static __inline
 619 size_t
 620 hammer2_devblksize(size_t bytes)
 621 {
 622         if (bytes <= HAMMER2_LBUFSIZE) {
 623                 return(HAMMER2_LBUFSIZE);
 624         } else {
 625                 KKASSERT(bytes <= HAMMER2_PBUFSIZE &&
 626                          (bytes ^ (bytes - 1)) == ((bytes << 1) - 1));
 627                 return (HAMMER2_PBUFSIZE);
 628         }
 629 }
 630
 631
 632 static __inline
 633 hammer2_pfsmount_t *
 634 MPTOPMP(struct mount *mp)
 635 {
 636         return ((hammer2_pfsmount_t *)mp->mnt_data);
 637 }
 638
 639 extern struct vop_ops hammer2_vnode_vops;
 640 extern struct vop_ops hammer2_spec_vops;
 641 extern struct vop_ops hammer2_fifo_vops;
 642
 643 extern int hammer2_debug;
 644 extern int hammer2_cluster_enable;
 645 extern int hammer2_hardlink_enable;
 646 extern int hammer2_flush_pipe;
 647 extern long hammer2_limit_dirty_chains;
 648 extern long hammer2_iod_file_read;
 649 extern long hammer2_iod_meta_read;
 650 extern long hammer2_iod_indr_read;
 651 extern long hammer2_iod_fmap_read;
 652 extern long hammer2_iod_volu_read;
 653 extern long hammer2_iod_file_write;
 654 extern long hammer2_iod_meta_write;
 655 extern long hammer2_iod_indr_write;
 656 extern long hammer2_iod_fmap_write;
 657 extern long hammer2_iod_volu_write;
 658 extern long hammer2_ioa_file_read;
 659 extern long hammer2_ioa_meta_read;
 660 extern long hammer2_ioa_indr_read;
 661 extern long hammer2_ioa_fmap_read;
 662 extern long hammer2_ioa_volu_read;
 663 extern long hammer2_ioa_file_write;
 664 extern long hammer2_ioa_meta_write;
 665 extern long hammer2_ioa_indr_write;
 666 extern long hammer2_ioa_fmap_write;
 667 extern long hammer2_ioa_volu_write;
 668
 669 extern struct objcache *cache_buffer_read;
 670 extern struct objcache *cache_buffer_write;
 671
 672 extern int destroy;
 673 extern int write_thread_wakeup;
 674
 675 extern mtx_t thread_protect;
 676
 677 /*
 678  * hammer2_subr.c
 679  */
 680 #define hammer2_icrc32(buf, size)       iscsi_crc32((buf), (size))
 681 #define hammer2_icrc32c(buf, size, crc) iscsi_crc32_ext((buf), (size), (crc))
 682
 683 hammer2_chain_t *hammer2_inode_lock_ex(hammer2_inode_t *ip);
 684 hammer2_chain_t *hammer2_inode_lock_sh(hammer2_inode_t *ip);
 685 void hammer2_inode_unlock_ex(hammer2_inode_t *ip, hammer2_chain_t *chain);
 686 void hammer2_inode_unlock_sh(hammer2_inode_t *ip, hammer2_chain_t *chain);
 687 void hammer2_chain_refactor(hammer2_chain_t **chainp);
 688 void hammer2_voldata_lock(hammer2_mount_t *hmp);
 689 void hammer2_voldata_unlock(hammer2_mount_t *hmp, int modify);
 690 ccms_state_t hammer2_inode_lock_temp_release(hammer2_inode_t *ip);
 691 void hammer2_inode_lock_temp_restore(hammer2_inode_t *ip, ccms_state_t ostate);
 692 ccms_state_t hammer2_inode_lock_upgrade(hammer2_inode_t *ip);
 693 void hammer2_inode_lock_downgrade(hammer2_inode_t *ip, ccms_state_t ostate);
 694
 695 void hammer2_mount_exlock(hammer2_mount_t *hmp);
 696 void hammer2_mount_shlock(hammer2_mount_t *hmp);
 697 void hammer2_mount_unlock(hammer2_mount_t *hmp);
 698
 699 int hammer2_get_dtype(hammer2_chain_t *chain);
 700 int hammer2_get_vtype(hammer2_chain_t *chain);
 701 u_int8_t hammer2_get_obj_type(enum vtype vtype);
 702 void hammer2_time_to_timespec(u_int64_t xtime, struct timespec *ts);
 703 u_int64_t hammer2_timespec_to_time(struct timespec *ts);
 704 u_int32_t hammer2_to_unix_xid(uuid_t *uuid);
 705 void hammer2_guid_to_uuid(uuid_t *uuid, u_int32_t guid);
 706
 707 hammer2_key_t hammer2_dirhash(const unsigned char *name, size_t len);
 708 int hammer2_getradix(size_t bytes);
 709
 710 int hammer2_calc_logical(hammer2_inode_t *ip, hammer2_off_t uoff,
 711                         hammer2_key_t *lbasep, hammer2_key_t *leofp);
 712 int hammer2_calc_physical(hammer2_inode_t *ip, hammer2_key_t lbase);
 713 void hammer2_update_time(uint64_t *timep);
 714
 715 /*
 716  * hammer2_inode.c
 717  */
 718 struct vnode *hammer2_igetv(hammer2_inode_t *ip, int *errorp);
 719
 720 void hammer2_inode_lock_nlinks(hammer2_inode_t *ip);
 721 void hammer2_inode_unlock_nlinks(hammer2_inode_t *ip);
 722 hammer2_inode_t *hammer2_inode_lookup(hammer2_pfsmount_t *pmp,
 723                         hammer2_tid_t inum);
 724 hammer2_inode_t *hammer2_inode_get(hammer2_pfsmount_t *pmp,
 725                         hammer2_inode_t *dip, hammer2_chain_t *chain);
 726 void hammer2_inode_free(hammer2_inode_t *ip);
 727 void hammer2_inode_ref(hammer2_inode_t *ip);
 728 void hammer2_inode_drop(hammer2_inode_t *ip);
 729 void hammer2_inode_repoint(hammer2_inode_t *ip, hammer2_inode_t *pip,
 730                         hammer2_chain_t *chain);
 731
 732 hammer2_inode_t *hammer2_inode_create(hammer2_trans_t *trans,
 733                         hammer2_inode_t *dip,
 734                         struct vattr *vap, struct ucred *cred,
 735                         const uint8_t *name, size_t name_len,
 736                         hammer2_chain_t **chainp, int *errorp);
 737 int hammer2_inode_connect(hammer2_trans_t *trans,
 738                         hammer2_chain_t **chainp, int hlink,
 739                         hammer2_inode_t *dip, hammer2_chain_t **dchainp,
 740                         const uint8_t *name, size_t name_len,
 741                         hammer2_key_t key);
 742 hammer2_inode_t *hammer2_inode_common_parent(hammer2_inode_t *fdip,
 743                         hammer2_inode_t *tdip);
 744 void hammer2_inode_fsync(hammer2_trans_t *trans, hammer2_inode_t *ip,
 745                         hammer2_chain_t **parentp);
 746 int hammer2_unlink_file(hammer2_trans_t *trans, hammer2_inode_t *dip,
 747                         const uint8_t *name, size_t name_len, int isdir,
 748                         int *hlinkp, struct nchandle *nch);
 749 int hammer2_hardlink_consolidate(hammer2_trans_t *trans,
 750                         hammer2_inode_t *ip, hammer2_chain_t **chainp,
 751                         hammer2_inode_t *cdip, hammer2_chain_t **cdchainp,
 752                         int nlinks);
 753 int hammer2_hardlink_deconsolidate(hammer2_trans_t *trans, hammer2_inode_t *dip,
 754                         hammer2_chain_t **chainp, hammer2_chain_t **ochainp);
 755 int hammer2_hardlink_find(hammer2_inode_t *dip,
 756                         hammer2_chain_t **chainp, hammer2_chain_t **ochainp);
 757 void hammer2_inode_install_hidden(hammer2_pfsmount_t *pmp);
 758
 759 /*
 760  * hammer2_chain.c
 761  */
 762 void hammer2_modify_volume(hammer2_mount_t *hmp);
 763 hammer2_chain_t *hammer2_chain_alloc(hammer2_mount_t *hmp,
 764                                 hammer2_pfsmount_t *pmp,
 765                                 hammer2_trans_t *trans,
 766                                 hammer2_blockref_t *bref);
 767 void hammer2_chain_core_alloc(hammer2_trans_t *trans, hammer2_chain_t *nchain,
 768                                 hammer2_chain_t *ochain);
 769 void hammer2_chain_ref(hammer2_chain_t *chain);
 770 void hammer2_chain_drop(hammer2_chain_t *chain);
 771 int hammer2_chain_lock(hammer2_chain_t *chain, int how);
 772 void hammer2_chain_load_async(hammer2_chain_t *chain,
 773                                 void (*func)(hammer2_io_t *dio,
 774                                              hammer2_chain_t *chain,
 775                                              void *arg_p, off_t arg_o),
 776                                 void *arg_p, off_t arg_o);
 777 void hammer2_chain_moved(hammer2_chain_t *chain);
 778 void hammer2_chain_modify(hammer2_trans_t *trans,
 779                                 hammer2_chain_t **chainp, int flags);
 780 hammer2_inode_data_t *hammer2_chain_modify_ip(hammer2_trans_t *trans,
 781                                 hammer2_inode_t *ip, hammer2_chain_t **chainp,
 782                                 int flags);
 783 void hammer2_chain_resize(hammer2_trans_t *trans, hammer2_inode_t *ip,
 784                                 hammer2_chain_t *parent,
 785                                 hammer2_chain_t **chainp,
 786                                 int nradix, int flags);
 787 void hammer2_chain_unlock(hammer2_chain_t *chain);
 788 void hammer2_chain_wait(hammer2_chain_t *chain);
 789 hammer2_chain_t *hammer2_chain_get(hammer2_chain_t *parent,
 790                                 hammer2_blockref_t *bref, int generation);
 791 hammer2_chain_t *hammer2_chain_lookup_init(hammer2_chain_t *parent, int flags);
 792 void hammer2_chain_lookup_done(hammer2_chain_t *parent);
 793 hammer2_chain_t *hammer2_chain_lookup(hammer2_chain_t **parentp,
 794                                 hammer2_key_t *key_nextp,
 795                                 hammer2_key_t key_beg, hammer2_key_t key_end,
 796                                 int *cache_indexp, int flags);
 797 hammer2_chain_t *hammer2_chain_next(hammer2_chain_t **parentp,
 798                                 hammer2_chain_t *chain,
 799                                 hammer2_key_t *key_nextp,
 800                                 hammer2_key_t key_beg, hammer2_key_t key_end,
 801                                 int *cache_indexp, int flags);
 802 hammer2_chain_t *hammer2_chain_scan(hammer2_chain_t *parent,
 803                                 hammer2_chain_t *chain,
 804                                 int *cache_indexp, int flags);
 805
 806 int hammer2_chain_create(hammer2_trans_t *trans,
 807                                 hammer2_chain_t **parentp,
 808                                 hammer2_chain_t **chainp,
 809                                 hammer2_key_t key, int keybits,
 810                                 int type, size_t bytes);
 811 void hammer2_chain_duplicate(hammer2_trans_t *trans, hammer2_chain_t **parentp,
 812                                 hammer2_chain_t **chainp,
 813                                 hammer2_blockref_t *bref, int snapshot,
 814                                 int duplicate_reason);
 815 int hammer2_chain_snapshot(hammer2_trans_t *trans, hammer2_chain_t **chainp,
 816                                 hammer2_ioc_pfs_t *pfs);
 817 void hammer2_chain_delete(hammer2_trans_t *trans, hammer2_chain_t *chain,
 818                                 int flags);
 819 void hammer2_chain_delete_duplicate(hammer2_trans_t *trans,
 820                                 hammer2_chain_t **chainp, int flags);
 821 void hammer2_chain_flush(hammer2_trans_t *trans, hammer2_chain_t **chainp);
 822 void hammer2_chain_commit(hammer2_trans_t *trans, hammer2_chain_t *chain);
 823 void hammer2_chain_setsubmod(hammer2_trans_t *trans, hammer2_chain_t *chain);
 824
 825 void hammer2_chain_memory_wait(hammer2_pfsmount_t *pmp);
 826 void hammer2_chain_memory_inc(hammer2_pfsmount_t *pmp);
 827 void hammer2_chain_memory_wakeup(hammer2_pfsmount_t *pmp);
 828 void hammer2_chain_countbrefs(hammer2_chain_t *chain,
 829                                 hammer2_blockref_t *base, int count);
 830 void hammer2_chain_layer_check_locked(hammer2_mount_t *hmp,
 831                                 hammer2_chain_core_t *core);
 832
 833 int hammer2_base_find(hammer2_chain_t *chain,
 834                                 hammer2_blockref_t *base, int count,
 835                                 int *cache_indexp, hammer2_key_t *key_nextp,
 836                                 hammer2_key_t key_beg, hammer2_key_t key_end);
 837 void hammer2_base_delete(hammer2_trans_t *trans, hammer2_chain_t *chain,
 838                                 hammer2_blockref_t *base, int count,
 839                                 int *cache_indexp, hammer2_chain_t *child);
 840 void hammer2_base_insert(hammer2_trans_t *trans, hammer2_chain_t *chain,
 841                                 hammer2_blockref_t *base, int count,
 842                                 int *cache_indexp, hammer2_chain_t *child);
 843
 844 /*
 845  * hammer2_trans.c
 846  */
 847 void hammer2_trans_init(hammer2_trans_t *trans, hammer2_pfsmount_t *pmp,
 848                                 hammer2_mount_t *hmp, int flags);
 849 void hammer2_trans_clear_invfsync(hammer2_trans_t *trans);
 850 void hammer2_trans_done(hammer2_trans_t *trans);
 851
 852 /*
 853  * hammer2_ioctl.c
 854  */
 855 int hammer2_ioctl(hammer2_inode_t *ip, u_long com, void *data,
 856                                 int fflag, struct ucred *cred);
 857
 858 /*
 859  * hammer2_io.c
 860  */
 861 hammer2_io_t *hammer2_io_getblk(hammer2_mount_t *hmp, off_t lbase,
 862                                 int lsize, int *ownerp);
 863 void hammer2_io_putblk(hammer2_io_t **diop);
 864 void hammer2_io_cleanup(hammer2_mount_t *hmp, struct hammer2_io_tree *tree);
 865 char *hammer2_io_data(hammer2_io_t *dio, off_t lbase);
 866 int hammer2_io_new(hammer2_mount_t *hmp, off_t lbase, int lsize,
 867                                 hammer2_io_t **diop);
 868 int hammer2_io_newnz(hammer2_mount_t *hmp, off_t lbase, int lsize,
 869                                 hammer2_io_t **diop);
 870 int hammer2_io_newq(hammer2_mount_t *hmp, off_t lbase, int lsize,
 871                                 hammer2_io_t **diop);
 872 int hammer2_io_bread(hammer2_mount_t *hmp, off_t lbase, int lsize,
 873                                 hammer2_io_t **diop);
 874 void hammer2_io_breadcb(hammer2_mount_t *hmp, off_t lbase, int lsize,
 875                                 void (*callback)(hammer2_io_t *dio,
 876                                                  hammer2_chain_t *arg_c,
 877                                                  void *arg_p, off_t arg_o),
 878                                 hammer2_chain_t *arg_c,
 879                                 void *arg_p, off_t arg_o);
 880 void hammer2_io_bawrite(hammer2_io_t **diop);
 881 void hammer2_io_bdwrite(hammer2_io_t **diop);
 882 int hammer2_io_bwrite(hammer2_io_t **diop);
 883 void hammer2_io_setdirty(hammer2_io_t *dio);
 884 void hammer2_io_setinval(hammer2_io_t *dio, u_int bytes);
 885 void hammer2_io_brelse(hammer2_io_t **diop);
 886 void hammer2_io_bqrelse(hammer2_io_t **diop);
 887 int hammer2_io_isdirty(hammer2_io_t *dio);
 888
 889 /*
 890  * hammer2_msgops.c
 891  */
 892 int hammer2_msg_dbg_rcvmsg(kdmsg_msg_t *msg);
 893 int hammer2_msg_adhoc_input(kdmsg_msg_t *msg);
 894
 895 /*
 896  * hammer2_vfsops.c
 897  */
 898 void hammer2_clusterctl_wakeup(kdmsg_iocom_t *iocom);
 899 void hammer2_volconf_update(hammer2_pfsmount_t *pmp, int index);
 900 void hammer2_cluster_reconnect(hammer2_pfsmount_t *pmp, struct file *fp);
 901 void hammer2_dump_chain(hammer2_chain_t *chain, int tab, int *countp);
 902 void hammer2_bioq_sync(hammer2_pfsmount_t *pmp);
 903 int hammer2_vfs_sync(struct mount *mp, int waitflags);
 904 void hammer2_lwinprog_ref(hammer2_pfsmount_t *pmp);
 905 void hammer2_lwinprog_drop(hammer2_pfsmount_t *pmp);
 906 void hammer2_lwinprog_wait(hammer2_pfsmount_t *pmp);
 907
 908 /*
 909  * hammer2_freemap.c
 910  */
 911 int hammer2_freemap_alloc(hammer2_trans_t *trans, hammer2_chain_t *chain,
 912                                 size_t bytes);
 913 void hammer2_freemap_adjust(hammer2_trans_t *trans, hammer2_mount_t *hmp,
 914                                 hammer2_blockref_t *bref, int how);
 915
 916
 917 #endif /* !_KERNEL */
 918 #endif /* !_VFS_HAMMER2_HAMMER2_H_ */