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         TAILQ_ENTRY(hammer2_chain_layer) entry;
 130         struct hammer2_chain_tree rbtree;
 131         int     refs;           /* prevent destruction */
 132 };
 133
 134 typedef struct hammer2_chain_layer hammer2_chain_layer_t;
 135
 136 struct hammer2_chain_core {
 137         struct ccms_cst cst;
 138         struct h2_core_list ownerq;     /* chain's which own this core */
 139         struct h2_layer_list layerq;
 140         u_int           chain_count;    /* total chains in layers */
 141         u_int           sharecnt;
 142         u_int           flags;
 143         u_int           live_count;     /* live (not deleted) chains in tree */
 144         u_int           live_zero;      /* first empty blockref (index) */
 145 };
 146
 147 typedef struct hammer2_chain_core hammer2_chain_core_t;
 148
 149 #define HAMMER2_CORE_INDIRECT           0x0001
 150 #define HAMMER2_CORE_COUNTEDBREFS       0x0002
 151
 152 struct hammer2_chain {
 153         RB_ENTRY(hammer2_chain) rbnode;         /* node */
 154         TAILQ_ENTRY(hammer2_chain) core_entry;  /* contemporary chains */
 155         hammer2_chain_layer_t   *inlayer;
 156         hammer2_blockref_t      bref;
 157         hammer2_chain_core_t    *core;
 158         hammer2_chain_core_t    *above;
 159         struct hammer2_state    *state;         /* if active cache msg */
 160         struct hammer2_mount    *hmp;
 161         struct hammer2_pfsmount *pmp;           /* can be NULL */
 162
 163         hammer2_tid_t   modify_tid;             /* snapshot/flush filter */
 164         hammer2_tid_t   delete_tid;
 165         hammer2_key_t   data_count;             /* delta's to apply */
 166         hammer2_key_t   inode_count;            /* delta's to apply */
 167         struct buf      *bp;                    /* physical data buffer */
 168         u_int           bytes;                  /* physical data size */
 169         u_int           flags;
 170         u_int           refs;
 171         u_int           lockcnt;
 172         hammer2_media_data_t *data;             /* data pointer shortcut */
 173         TAILQ_ENTRY(hammer2_chain) flush_node;  /* flush deferral list */
 174 };
 175
 176 typedef struct hammer2_chain hammer2_chain_t;
 177
 178 int hammer2_chain_cmp(hammer2_chain_t *chain1, hammer2_chain_t *chain2);
 179 RB_PROTOTYPE(hammer2_chain_tree, hammer2_chain, rbnode, hammer2_chain_cmp);
 180
 181 /*
 182  * Special notes on flags:
 183  *
 184  * INITIAL - This flag allows a chain to be created and for storage to
 185  *           be allocated without having to immediately instantiate the
 186  *           related buffer.  The data is assumed to be all-zeros.  It
 187  *           is primarily used for indirect blocks.
 188  *
 189  * MOVED   - A modified chain becomes MOVED after it flushes.  A chain
 190  *           can also become MOVED if it is moved within the topology
 191  *           (even if not modified).
 192  */
 193 #define HAMMER2_CHAIN_MODIFIED          0x00000001      /* dirty chain data */
 194 #define HAMMER2_CHAIN_ALLOCATED         0x00000002      /* kmalloc'd chain */
 195 #define HAMMER2_CHAIN_DIRTYBP           0x00000004      /* dirty on unlock */
 196 #define HAMMER2_CHAIN_SUBMODIFIED       0x00000008      /* recursive flush */
 197 #define HAMMER2_CHAIN_DELETED           0x00000010      /* deleted chain */
 198 #define HAMMER2_CHAIN_INITIAL           0x00000020      /* initial create */
 199 #define HAMMER2_CHAIN_FLUSHED           0x00000040      /* flush on unlock */
 200 #define HAMMER2_CHAIN_MOVED             0x00000080      /* bref changed */
 201 #define HAMMER2_CHAIN_IOFLUSH           0x00000100      /* bawrite on put */
 202 #define HAMMER2_CHAIN_DEFERRED          0x00000200      /* on a deferral list */
 203 #define HAMMER2_CHAIN_DESTROYED         0x00000400      /* destroying inode */
 204 #define HAMMER2_CHAIN_VOLUMESYNC        0x00000800      /* needs volume sync */
 205 #define HAMMER2_CHAIN_RECYCLE           0x00001000      /* force recycle */
 206 #define HAMMER2_CHAIN_MOUNTED           0x00002000      /* PFS is mounted */
 207 #define HAMMER2_CHAIN_ONRBTREE          0x00004000      /* on parent RB tree */
 208 #define HAMMER2_CHAIN_SNAPSHOT          0x00008000      /* snapshot special */
 209 #define HAMMER2_CHAIN_EMBEDDED          0x00010000      /* embedded data */
 210 #define HAMMER2_CHAIN_HARDLINK          0x00020000      /* converted to hlink */
 211 #define HAMMER2_CHAIN_REPLACE           0x00040000      /* replace bref */
 212
 213 /*
 214  * Flags passed to hammer2_chain_lookup() and hammer2_chain_next()
 215  *
 216  * NOTE: MATCHIND allows an indirect block / freemap node to be returned
 217  *       when the passed key range matches the radix.  Remember that key_end
 218  *       is inclusive (e.g. {0x000,0xFFF}, not {0x000,0x1000}).
 219  */
 220 #define HAMMER2_LOOKUP_NOLOCK           0x00000001      /* ref only */
 221 #define HAMMER2_LOOKUP_NODATA           0x00000002      /* data left NULL */
 222 #define HAMMER2_LOOKUP_SHARED           0x00000100
 223 #define HAMMER2_LOOKUP_MATCHIND         0x00000200
 224 #define HAMMER2_LOOKUP_FREEMAP          0x00000400      /* freemap base */
 225 #define HAMMER2_LOOKUP_ALWAYS           0x00000800      /* resolve data */
 226
 227 /*
 228  * Flags passed to hammer2_chain_modify() and hammer2_chain_resize()
 229  *
 230  * NOTE: OPTDATA allows us to avoid instantiating buffers for INDIRECT
 231  *       blocks in the INITIAL-create state.
 232  */
 233 #define HAMMER2_MODIFY_OPTDATA          0x00000002      /* data can be NULL */
 234 #define HAMMER2_MODIFY_NO_MODIFY_TID    0x00000004
 235 #define HAMMER2_MODIFY_ASSERTNOCOPY     0x00000008
 236 #define HAMMER2_MODIFY_NOREALLOC        0x00000010
 237
 238 /*
 239  * Flags passed to hammer2_chain_lock()
 240  */
 241 #define HAMMER2_RESOLVE_NEVER           1
 242 #define HAMMER2_RESOLVE_MAYBE           2
 243 #define HAMMER2_RESOLVE_ALWAYS          3
 244 #define HAMMER2_RESOLVE_MASK            0x0F
 245
 246 #define HAMMER2_RESOLVE_SHARED          0x10    /* request shared lock */
 247 #define HAMMER2_RESOLVE_NOREF           0x20    /* already ref'd on lock */
 248
 249 /*
 250  * Flags passed to hammer2_chain_delete()
 251  */
 252 #define HAMMER2_DELETE_WILLDUP          0x0001  /* no blk free, will be dup */
 253
 254 /*
 255  * Flags passed to hammer2_chain_delete_duplicate()
 256  */
 257 #define HAMMER2_DELDUP_RECORE           0x0001
 258
 259 /*
 260  * Cluster different types of storage together for allocations
 261  */
 262 #define HAMMER2_FREECACHE_INODE         0
 263 #define HAMMER2_FREECACHE_INDIR         1
 264 #define HAMMER2_FREECACHE_DATA          2
 265 #define HAMMER2_FREECACHE_UNUSED3       3
 266 #define HAMMER2_FREECACHE_TYPES         4
 267
 268 /*
 269  * hammer2_freemap_alloc() block preference
 270  */
 271 #define HAMMER2_OFF_NOPREF              ((hammer2_off_t)-1)
 272
 273 /*
 274  * BMAP read-ahead maximum parameters
 275  */
 276 #define HAMMER2_BMAP_COUNT              16      /* max bmap read-ahead */
 277 #define HAMMER2_BMAP_BYTES              (HAMMER2_PBUFSIZE * HAMMER2_BMAP_COUNT)
 278
 279 /*
 280  * Misc
 281  */
 282 #define HAMMER2_FLUSH_DEPTH_LIMIT       40      /* stack recursion limit */
 283
 284 /*
 285  * HAMMER2 IN-MEMORY CACHE OF MEDIA STRUCTURES
 286  *
 287  * There is an in-memory representation of all on-media data structure.
 288  *
 289  * When accessed read-only the data will be mapped to the related buffer
 290  * cache buffer.
 291  *
 292  * When accessed read-write (marked modified) a kmalloc()'d copy of the
 293  * is created which can then be modified.  The copy is destroyed when a
 294  * filesystem block is allocated to replace it.
 295  *
 296  * Active inodes (those with vnodes attached) will maintain the kmalloc()'d
 297  * copy for both the read-only and the read-write case.  The combination of
 298  * (bp) and (data) determines whether (data) was allocated or not.
 299  *
 300  * The in-memory representation may remain cached (for example in order to
 301  * placemark clustering locks) even after the related data has been
 302  * detached.
 303  */
 304
 305 RB_HEAD(hammer2_inode_tree, hammer2_inode);
 306
 307 /*
 308  * A hammer2 inode.
 309  *
 310  * NOTE: The inode's attribute CST which is also used to lock the inode
 311  *       is embedded in the chain (chain.cst) and aliased w/ attr_cst.
 312  */
 313 struct hammer2_inode {
 314         RB_ENTRY(hammer2_inode) rbnode;         /* inumber lookup (HL) */
 315         ccms_cst_t              topo_cst;       /* directory topology cst */
 316         struct hammer2_pfsmount *pmp;           /* PFS mount */
 317         struct hammer2_inode    *pip;           /* parent inode */
 318         struct vnode            *vp;
 319         hammer2_chain_t         *chain;         /* NOTE: rehomed on rename */
 320         struct lockf            advlock;
 321         hammer2_tid_t           inum;
 322         u_int                   flags;
 323         u_int                   refs;           /* +vpref, +flushref */
 324         uint8_t                 comp_heuristic;
 325         hammer2_off_t           size;
 326         uint64_t                mtime;
 327 };
 328
 329 typedef struct hammer2_inode hammer2_inode_t;
 330
 331 #define HAMMER2_INODE_MODIFIED          0x0001
 332 #define HAMMER2_INODE_SROOT             0x0002  /* kmalloc special case */
 333 #define HAMMER2_INODE_RENAME_INPROG     0x0004
 334 #define HAMMER2_INODE_ONRBTREE          0x0008
 335 #define HAMMER2_INODE_RESIZED           0x0010
 336 #define HAMMER2_INODE_MTIME             0x0020
 337
 338 int hammer2_inode_cmp(hammer2_inode_t *ip1, hammer2_inode_t *ip2);
 339 RB_PROTOTYPE2(hammer2_inode_tree, hammer2_inode, rbnode, hammer2_inode_cmp,
 340                 hammer2_tid_t);
 341
 342 /*
 343  * A hammer2 transaction and flush sequencing structure.
 344  *
 345  * This global structure is tied into hammer2_mount and is used
 346  * to sequence modifying operations and flushes.
 347  *
 348  * (a) Any modifying operations with sync_tid >= flush_tid will stall until
 349  *     all modifying operating with sync_tid < flush_tid complete.
 350  *
 351  *     The flush related to flush_tid stalls until all modifying operations
 352  *     with sync_tid < flush_tid complete.
 353  *
 354  * (b) Once unstalled, modifying operations with sync_tid > flush_tid are
 355  *     allowed to run.  All modifications cause modify/duplicate operations
 356  *     to occur on the related chains.  Note that most INDIRECT blocks will
 357  *     be unaffected because the modifications just overload the RBTREE
 358  *     structurally instead of actually modifying the indirect blocks.
 359  *
 360  * (c) The actual flush unstalls and RUNS CONCURRENTLY with (b), but only
 361  *     utilizes the chain structures with sync_tid <= flush_tid.  The
 362  *     flush will modify related indirect blocks and inodes in-place
 363  *     (rather than duplicate) since the adjustments are compatible with
 364  *     (b)'s RBTREE overloading
 365  *
 366  *     SPECIAL NOTE:  Inode modifications have to also propagate along any
 367  *                    modify/duplicate chains.  File writes detect the flush
 368  *                    and force out the conflicting buffer cache buffer(s)
 369  *                    before reusing them.
 370  *
 371  * (d) Snapshots can be made instantly but must be flushed and disconnected
 372  *     from their duplicative source before they can be mounted.  This is
 373  *     because while H2's on-media structure supports forks, its in-memory
 374  *     structure only supports very simple forking for background flushing
 375  *     purposes.
 376  *
 377  * TODO: Flush merging.  When fsync() is called on multiple discrete files
 378  *       concurrently there is no reason to stall the second fsync.
 379  *       The final flush that reaches to root can cover both fsync()s.
 380  *
 381  *     The chains typically terminate as they fly onto the disk.  The flush
 382  *     ultimately reaches the volume header.
 383  */
 384 struct hammer2_trans {
 385         TAILQ_ENTRY(hammer2_trans) entry;
 386         struct hammer2_pfsmount *pmp;
 387         hammer2_tid_t           sync_tid;
 388         thread_t                td;             /* pointer */
 389         int                     flags;
 390         int                     blocked;
 391         uint8_t                 inodes_created;
 392         uint8_t                 dummy[7];
 393 };
 394
 395 typedef struct hammer2_trans hammer2_trans_t;
 396
 397 #define HAMMER2_TRANS_ISFLUSH           0x0001  /* formal flush */
 398 #define HAMMER2_TRANS_RESTRICTED        0x0002  /* snapshot flush restrict */
 399 #define HAMMER2_TRANS_BUFCACHE          0x0004  /* from bioq strategy write */
 400
 401 #define HAMMER2_FREEMAP_HEUR_NRADIX     4       /* pwr 2 PBUFRADIX-MINIORADIX */
 402 #define HAMMER2_FREEMAP_HEUR_TYPES      8
 403 #define HAMMER2_FREEMAP_HEUR            (HAMMER2_FREEMAP_HEUR_NRADIX * \
 404                                          HAMMER2_FREEMAP_HEUR_TYPES)
 405
 406 /*
 407  * Global (per device) mount structure for device (aka vp->v_mount->hmp)
 408  */
 409 TAILQ_HEAD(hammer2_trans_queue, hammer2_trans);
 410
 411 struct hammer2_mount {
 412         struct vnode    *devvp;         /* device vnode */
 413         int             ronly;          /* read-only mount */
 414         int             pmp_count;      /* PFS mounts backed by us */
 415         TAILQ_ENTRY(hammer2_mount) mntentry; /* hammer2_mntlist */
 416
 417         struct malloc_type *mchain;
 418         int             nipstacks;
 419         int             maxipstacks;
 420         hammer2_chain_t vchain;         /* anchor chain */
 421         hammer2_chain_t fchain;         /* freemap chain special */
 422         hammer2_chain_t *schain;        /* super-root */
 423         hammer2_inode_t *sroot;         /* super-root inode */
 424         struct lock     alloclk;        /* lockmgr lock */
 425         struct lock     voldatalk;      /* lockmgr lock */
 426         struct hammer2_trans_queue transq; /* all in-progress transactions */
 427         hammer2_trans_t *curflush;      /* current flush in progress */
 428         hammer2_tid_t   topo_flush_tid; /* currently synchronizing flush pt */
 429         hammer2_tid_t   free_flush_tid; /* currently synchronizing flush pt */
 430         hammer2_off_t   heur_freemap[HAMMER2_FREEMAP_HEUR];
 431         int             flushcnt;       /* #of flush trans on the list */
 432
 433         int             volhdrno;       /* last volhdrno written */
 434         hammer2_volume_data_t voldata;
 435         hammer2_volume_data_t volsync;  /* synchronized voldata */
 436         struct bio_queue_head wthread_bioq; /* bio queue for write thread */
 437         struct mtx wthread_mtx;     /* mutex for write thread */
 438         int     wthread_destroy;    /* to control the write thread */
 439 };
 440
 441 typedef struct hammer2_mount hammer2_mount_t;
 442
 443 /*
 444  * HAMMER2 cluster - a device/root associated with a PFS.
 445  *
 446  * A PFS may have several hammer2_cluster's associated with it.
 447  */
 448 struct hammer2_cluster {
 449         struct hammer2_mount    *hmp;           /* device global mount */
 450         hammer2_chain_t         *rchain;        /* PFS root chain */
 451 };
 452
 453 typedef struct hammer2_cluster hammer2_cluster_t;
 454
 455 /*
 456  * HAMMER2 PFS mount point structure (aka vp->v_mount->mnt_data).
 457  *
 458  * This structure represents a cluster mount and not necessarily a
 459  * PFS under a specific device mount (HMP).  The distinction is important
 460  * because the elements backing a cluster mount can change on the fly.
 461  */
 462 struct hammer2_pfsmount {
 463         struct mount            *mp;            /* kernel mount */
 464         hammer2_cluster_t       *mount_cluster;
 465         hammer2_cluster_t       *cluster;
 466         hammer2_inode_t         *iroot;         /* PFS root inode */
 467         hammer2_off_t           inode_count;    /* copy of inode_count */
 468         ccms_domain_t           ccms_dom;
 469         struct netexport        export;         /* nfs export */
 470         int                     ronly;          /* read-only mount */
 471         struct malloc_type      *minode;
 472         struct malloc_type      *mmsg;
 473         kdmsg_iocom_t           iocom;
 474         struct spinlock         inum_spin;      /* inumber lookup */
 475         struct hammer2_inode_tree inum_tree;
 476         long                    inmem_inodes;
 477         long                    inmem_chains;
 478         int                     inmem_waiting;
 479 };
 480
 481 typedef struct hammer2_pfsmount hammer2_pfsmount_t;
 482
 483 struct hammer2_cbinfo {
 484         hammer2_chain_t *chain;
 485         void (*func)(hammer2_chain_t *, struct buf *, char *, void *);
 486         void *arg;
 487         size_t boff;
 488 };
 489
 490 typedef struct hammer2_cbinfo hammer2_cbinfo_t;
 491
 492 #if defined(_KERNEL)
 493
 494 MALLOC_DECLARE(M_HAMMER2);
 495
 496 #define VTOI(vp)        ((hammer2_inode_t *)(vp)->v_data)
 497 #define ITOV(ip)        ((ip)->vp)
 498
 499 /*
 500  * Currently locked chains retain the locked buffer cache buffer for
 501  * indirect blocks, and indirect blocks can be one of two sizes.  The
 502  * device buffer has to match the case to avoid deadlocking recursive
 503  * chains that might otherwise try to access different offsets within
 504  * the same device buffer.
 505  */
 506 static __inline
 507 int
 508 hammer2_devblkradix(int radix)
 509 {
 510 #if 1
 511         if (radix <= HAMMER2_LBUFRADIX) {
 512                 return (HAMMER2_LBUFRADIX);
 513         } else {
 514                 return (HAMMER2_PBUFRADIX);
 515         }
 516 #else
 517         return (HAMMER2_PBUFRADIX);
 518 #endif
 519 }
 520
 521 static __inline
 522 size_t
 523 hammer2_devblksize(size_t bytes)
 524 {
 525 #if 1
 526         if (bytes <= HAMMER2_LBUFSIZE) {
 527                 return(HAMMER2_LBUFSIZE);
 528         } else {
 529                 KKASSERT(bytes <= HAMMER2_PBUFSIZE &&
 530                          (bytes ^ (bytes - 1)) == ((bytes << 1) - 1));
 531                 return (HAMMER2_PBUFSIZE);
 532         }
 533 #else
 534         KKASSERT(bytes <= HAMMER2_PBUFSIZE &&
 535                  (bytes ^ (bytes - 1)) == ((bytes << 1) - 1));
 536         return(HAMMER2_PBUFSIZE);
 537 #endif
 538 }
 539
 540
 541 static __inline
 542 hammer2_pfsmount_t *
 543 MPTOPMP(struct mount *mp)
 544 {
 545         return ((hammer2_pfsmount_t *)mp->mnt_data);
 546 }
 547
 548 static __inline
 549 hammer2_mount_t *
 550 MPTOHMP(struct mount *mp)
 551 {
 552         return (((hammer2_pfsmount_t *)mp->mnt_data)->cluster->hmp);
 553 }
 554
 555 static __inline
 556 int
 557 hammer2_chain_refactor_test(hammer2_chain_t *chain, int traverse_hlink)
 558 {
 559         hammer2_chain_t *next;
 560
 561         next = TAILQ_NEXT(chain, core_entry);
 562
 563         if ((chain->flags & HAMMER2_CHAIN_DELETED) &&
 564             next &&
 565             (next->flags & HAMMER2_CHAIN_SNAPSHOT) == 0) {
 566                 return (1);
 567         }
 568         if (traverse_hlink &&
 569             chain->bref.type == HAMMER2_BREF_TYPE_INODE &&
 570             (chain->flags & HAMMER2_CHAIN_HARDLINK) &&
 571             next &&
 572             (next->flags & HAMMER2_CHAIN_SNAPSHOT) == 0) {
 573                 return(1);
 574         }
 575
 576         return (0);
 577 }
 578
 579 extern struct vop_ops hammer2_vnode_vops;
 580 extern struct vop_ops hammer2_spec_vops;
 581 extern struct vop_ops hammer2_fifo_vops;
 582
 583 extern int hammer2_debug;
 584 extern int hammer2_cluster_enable;
 585 extern int hammer2_hardlink_enable;
 586 extern long hammer2_iod_file_read;
 587 extern long hammer2_iod_meta_read;
 588 extern long hammer2_iod_indr_read;
 589 extern long hammer2_iod_fmap_read;
 590 extern long hammer2_iod_volu_read;
 591 extern long hammer2_iod_file_write;
 592 extern long hammer2_iod_meta_write;
 593 extern long hammer2_iod_indr_write;
 594 extern long hammer2_iod_fmap_write;
 595 extern long hammer2_iod_volu_write;
 596 extern long hammer2_ioa_file_read;
 597 extern long hammer2_ioa_meta_read;
 598 extern long hammer2_ioa_indr_read;
 599 extern long hammer2_ioa_fmap_read;
 600 extern long hammer2_ioa_volu_read;
 601 extern long hammer2_ioa_file_write;
 602 extern long hammer2_ioa_meta_write;
 603 extern long hammer2_ioa_indr_write;
 604 extern long hammer2_ioa_fmap_write;
 605 extern long hammer2_ioa_volu_write;
 606
 607 extern struct objcache *cache_buffer_read;
 608 extern struct objcache *cache_buffer_write;
 609
 610 extern int destroy;
 611 extern int write_thread_wakeup;
 612
 613 extern mtx_t thread_protect;
 614
 615 /*
 616  * hammer2_subr.c
 617  */
 618 #define hammer2_icrc32(buf, size)       iscsi_crc32((buf), (size))
 619 #define hammer2_icrc32c(buf, size, crc) iscsi_crc32_ext((buf), (size), (crc))
 620
 621 hammer2_chain_t *hammer2_inode_lock_ex(hammer2_inode_t *ip);
 622 hammer2_chain_t *hammer2_inode_lock_sh(hammer2_inode_t *ip);
 623 void hammer2_inode_unlock_ex(hammer2_inode_t *ip, hammer2_chain_t *chain);
 624 void hammer2_inode_unlock_sh(hammer2_inode_t *ip, hammer2_chain_t *chain);
 625 void hammer2_voldata_lock(hammer2_mount_t *hmp);
 626 void hammer2_voldata_unlock(hammer2_mount_t *hmp, int modify);
 627 ccms_state_t hammer2_inode_lock_temp_release(hammer2_inode_t *ip);
 628 void hammer2_inode_lock_temp_restore(hammer2_inode_t *ip, ccms_state_t ostate);
 629 ccms_state_t hammer2_inode_lock_upgrade(hammer2_inode_t *ip);
 630 void hammer2_inode_lock_downgrade(hammer2_inode_t *ip, ccms_state_t ostate);
 631
 632 void hammer2_mount_exlock(hammer2_mount_t *hmp);
 633 void hammer2_mount_shlock(hammer2_mount_t *hmp);
 634 void hammer2_mount_unlock(hammer2_mount_t *hmp);
 635
 636 int hammer2_get_dtype(hammer2_chain_t *chain);
 637 int hammer2_get_vtype(hammer2_chain_t *chain);
 638 u_int8_t hammer2_get_obj_type(enum vtype vtype);
 639 void hammer2_time_to_timespec(u_int64_t xtime, struct timespec *ts);
 640 u_int64_t hammer2_timespec_to_time(struct timespec *ts);
 641 u_int32_t hammer2_to_unix_xid(uuid_t *uuid);
 642 void hammer2_guid_to_uuid(uuid_t *uuid, u_int32_t guid);
 643
 644 hammer2_key_t hammer2_dirhash(const unsigned char *name, size_t len);
 645 int hammer2_getradix(size_t bytes);
 646
 647 int hammer2_calc_logical(hammer2_inode_t *ip, hammer2_off_t uoff,
 648                         hammer2_key_t *lbasep, hammer2_key_t *leofp);
 649 int hammer2_calc_physical(hammer2_inode_t *ip, hammer2_key_t lbase);
 650 void hammer2_update_time(uint64_t *timep);
 651
 652 /*
 653  * hammer2_inode.c
 654  */
 655 struct vnode *hammer2_igetv(hammer2_inode_t *ip, int *errorp);
 656
 657 void hammer2_inode_lock_nlinks(hammer2_inode_t *ip);
 658 void hammer2_inode_unlock_nlinks(hammer2_inode_t *ip);
 659 hammer2_inode_t *hammer2_inode_lookup(hammer2_pfsmount_t *pmp,
 660                         hammer2_tid_t inum);
 661 hammer2_inode_t *hammer2_inode_get(hammer2_pfsmount_t *pmp,
 662                         hammer2_inode_t *dip, hammer2_chain_t *chain);
 663 void hammer2_inode_free(hammer2_inode_t *ip);
 664 void hammer2_inode_ref(hammer2_inode_t *ip);
 665 void hammer2_inode_drop(hammer2_inode_t *ip);
 666 void hammer2_inode_repoint(hammer2_inode_t *ip, hammer2_inode_t *pip,
 667                         hammer2_chain_t *chain);
 668
 669 hammer2_inode_t *hammer2_inode_create(hammer2_trans_t *trans,
 670                         hammer2_inode_t *dip,
 671                         struct vattr *vap, struct ucred *cred,
 672                         const uint8_t *name, size_t name_len,
 673                         hammer2_chain_t **chainp, int *errorp);
 674 int hammer2_inode_connect(hammer2_trans_t *trans, int hlink,
 675                         hammer2_inode_t *dip, hammer2_chain_t **chainp,
 676                         const uint8_t *name, size_t name_len);
 677 hammer2_inode_t *hammer2_inode_common_parent(hammer2_inode_t *fdip,
 678                         hammer2_inode_t *tdip);
 679 void hammer2_inode_fsync(hammer2_trans_t *trans, hammer2_inode_t *ip,
 680                         hammer2_chain_t **parentp);
 681 int hammer2_unlink_file(hammer2_trans_t *trans, hammer2_inode_t *dip,
 682                         const uint8_t *name, size_t name_len, int isdir,
 683                         int *hlinkp);
 684 int hammer2_hardlink_consolidate(hammer2_trans_t *trans, hammer2_inode_t *ip,
 685                         hammer2_chain_t **chainp,
 686                         hammer2_inode_t *tdip, int linkcnt);
 687 int hammer2_hardlink_deconsolidate(hammer2_trans_t *trans, hammer2_inode_t *dip,
 688                         hammer2_chain_t **chainp, hammer2_chain_t **ochainp);
 689 int hammer2_hardlink_find(hammer2_inode_t *dip,
 690                         hammer2_chain_t **chainp, hammer2_chain_t **ochainp);
 691
 692 /*
 693  * hammer2_chain.c
 694  */
 695 void hammer2_modify_volume(hammer2_mount_t *hmp);
 696 hammer2_chain_t *hammer2_chain_alloc(hammer2_mount_t *hmp, hammer2_pfsmount_t *pmp,
 697                                 hammer2_trans_t *trans, hammer2_blockref_t *bref);
 698 void hammer2_chain_core_alloc(hammer2_trans_t *trans, hammer2_chain_t *nchain,
 699                                 hammer2_chain_t *ochain);
 700 void hammer2_chain_ref(hammer2_chain_t *chain);
 701 void hammer2_chain_drop(hammer2_chain_t *chain);
 702 int hammer2_chain_lock(hammer2_chain_t *chain, int how);
 703 void hammer2_chain_load_async(hammer2_chain_t *chain,
 704                                 void (*func)(hammer2_chain_t *, struct buf *,
 705                                              char *, void *),
 706                                 void *arg);
 707 void hammer2_chain_moved(hammer2_chain_t *chain);
 708 void hammer2_chain_modify(hammer2_trans_t *trans,
 709                                 hammer2_chain_t **chainp, int flags);
 710 hammer2_inode_data_t *hammer2_chain_modify_ip(hammer2_trans_t *trans,
 711                                 hammer2_inode_t *ip, hammer2_chain_t **chainp,
 712                                 int flags);
 713 void hammer2_chain_resize(hammer2_trans_t *trans, hammer2_inode_t *ip,
 714                                 hammer2_chain_t *parent,
 715                                 hammer2_chain_t **chainp,
 716                                 int nradix, int flags);
 717 void hammer2_chain_unlock(hammer2_chain_t *chain);
 718 void hammer2_chain_wait(hammer2_chain_t *chain);
 719 hammer2_chain_t *hammer2_chain_get(hammer2_chain_t *parent,
 720                                 hammer2_blockref_t *bref);
 721 hammer2_chain_t *hammer2_chain_lookup_init(hammer2_chain_t *parent, int flags);
 722 void hammer2_chain_lookup_done(hammer2_chain_t *parent);
 723 hammer2_chain_t *hammer2_chain_lookup(hammer2_chain_t **parentp,
 724                                 hammer2_key_t *key_nextp,
 725                                 hammer2_key_t key_beg, hammer2_key_t key_end,
 726                                 int *cache_indexp, int flags);
 727 hammer2_chain_t *hammer2_chain_next(hammer2_chain_t **parentp,
 728                                 hammer2_chain_t *chain,
 729                                 hammer2_key_t *key_nextp,
 730                                 hammer2_key_t key_beg, hammer2_key_t key_end,
 731                                 int *cache_indexp, int flags);
 732
 733 int hammer2_chain_create(hammer2_trans_t *trans,
 734                                 hammer2_chain_t **parentp,
 735                                 hammer2_chain_t **chainp,
 736                                 hammer2_key_t key, int keybits,
 737                                 int type, size_t bytes);
 738 void hammer2_chain_duplicate(hammer2_trans_t *trans, hammer2_chain_t *parent,
 739                                 hammer2_chain_t **chainp,
 740                                 hammer2_blockref_t *bref);
 741 int hammer2_chain_snapshot(hammer2_trans_t *trans, hammer2_inode_t *ip,
 742                                 hammer2_ioc_pfs_t *pfs);
 743 void hammer2_chain_delete(hammer2_trans_t *trans, hammer2_chain_t *chain,
 744                                 int flags);
 745 void hammer2_chain_delete_duplicate(hammer2_trans_t *trans,
 746                                 hammer2_chain_t **chainp, int flags);
 747 void hammer2_chain_flush(hammer2_trans_t *trans, hammer2_chain_t *chain);
 748 void hammer2_chain_commit(hammer2_trans_t *trans, hammer2_chain_t *chain);
 749 void hammer2_chain_setsubmod(hammer2_trans_t *trans, hammer2_chain_t *chain);
 750
 751 void hammer2_chain_memory_wait(hammer2_pfsmount_t *pmp);
 752 void hammer2_chain_memory_wakeup(hammer2_pfsmount_t *pmp);
 753 void hammer2_chain_countbrefs(hammer2_chain_core_t *above,
 754                                 hammer2_blockref_t *base, int count);
 755 void hammer2_chain_layer_check_locked(hammer2_mount_t *hmp,
 756                                 hammer2_chain_core_t *core);
 757
 758 int hammer2_base_find(hammer2_blockref_t *base, int count,
 759                                 hammer2_chain_core_t *above,
 760                                 int *cache_indexp,
 761                                 hammer2_key_t *key_nextp, hammer2_key_t key);
 762 void hammer2_base_delete(hammer2_blockref_t *base, int count,
 763                                 hammer2_chain_core_t *above,
 764                                 int *cache_indexp, hammer2_blockref_t *elm);
 765 void hammer2_base_insert(hammer2_blockref_t *base, int count,
 766                                 hammer2_chain_core_t *above,
 767                                 int *cache_indexp, hammer2_blockref_t *elm,
 768                                 int flags);
 769
 770 /*
 771  * hammer2_trans.c
 772  */
 773 void hammer2_trans_init(hammer2_trans_t *trans,
 774                         hammer2_pfsmount_t *pmp, int flags);
 775 void hammer2_trans_done(hammer2_trans_t *trans);
 776
 777 /*
 778  * hammer2_ioctl.c
 779  */
 780 int hammer2_ioctl(hammer2_inode_t *ip, u_long com, void *data,
 781                                 int fflag, struct ucred *cred);
 782
 783 /*
 784  * hammer2_msgops.c
 785  */
 786 int hammer2_msg_dbg_rcvmsg(kdmsg_msg_t *msg);
 787 int hammer2_msg_adhoc_input(kdmsg_msg_t *msg);
 788
 789 /*
 790  * hammer2_vfsops.c
 791  */
 792 void hammer2_clusterctl_wakeup(kdmsg_iocom_t *iocom);
 793 void hammer2_volconf_update(hammer2_pfsmount_t *pmp, int index);
 794 void hammer2_cluster_reconnect(hammer2_pfsmount_t *pmp, struct file *fp);
 795 void hammer2_dump_chain(hammer2_chain_t *chain, int tab, int *countp);
 796
 797 /*
 798  * hammer2_freemap.c
 799  */
 800 int hammer2_freemap_alloc(hammer2_trans_t *trans, hammer2_mount_t *hmp,
 801                                 hammer2_blockref_t *bref, size_t bytes);
 802 void hammer2_freemap_free(hammer2_trans_t *trans, hammer2_mount_t *hmp,
 803                                 hammer2_blockref_t *bref, int how);
 804
 805
 806 #endif /* !_KERNEL */
 807 #endif /* !_VFS_HAMMER2_HAMMER2_H_ */