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;     /* chain's which own this core */
 141         struct h2_layer_list layerq;
 142         u_int           chain_count;    /* total chains in layers */
 143         u_int           sharecnt;
 144         u_int           flags;
 145         u_int           live_count;     /* live (not deleted) chains in tree */
 146 };
 147
 148 typedef struct hammer2_chain_core hammer2_chain_core_t;
 149
 150 #define HAMMER2_CORE_INDIRECT           0x0001
 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         int             live_zero;              /* blockref array opt */
 173         hammer2_media_data_t *data;             /* data pointer shortcut */
 174         TAILQ_ENTRY(hammer2_chain) flush_node;  /* flush deferral list */
 175 };
 176
 177 typedef struct hammer2_chain hammer2_chain_t;
 178
 179 int hammer2_chain_cmp(hammer2_chain_t *chain1, hammer2_chain_t *chain2);
 180 RB_PROTOTYPE(hammer2_chain_tree, hammer2_chain, rbnode, hammer2_chain_cmp);
 181
 182 /*
 183  * Special notes on flags:
 184  *
 185  * INITIAL - This flag allows a chain to be created and for storage to
 186  *           be allocated without having to immediately instantiate the
 187  *           related buffer.  The data is assumed to be all-zeros.  It
 188  *           is primarily used for indirect blocks.
 189  *
 190  * MOVED   - A modified chain becomes MOVED after it flushes.  A chain
 191  *           can also become MOVED if it is moved within the topology
 192  *           (even if not modified).
 193  */
 194 #define HAMMER2_CHAIN_MODIFIED          0x00000001      /* dirty chain data */
 195 #define HAMMER2_CHAIN_ALLOCATED         0x00000002      /* kmalloc'd chain */
 196 #define HAMMER2_CHAIN_DIRTYBP           0x00000004      /* dirty on unlock */
 197 #define HAMMER2_CHAIN_SUBMODIFIED       0x00000008      /* recursive flush */
 198 #define HAMMER2_CHAIN_DELETED           0x00000010      /* deleted chain */
 199 #define HAMMER2_CHAIN_INITIAL           0x00000020      /* initial create */
 200 #define HAMMER2_CHAIN_FLUSHED           0x00000040      /* flush on unlock */
 201 #define HAMMER2_CHAIN_MOVED             0x00000080      /* bref changed */
 202 #define HAMMER2_CHAIN_IOFLUSH           0x00000100      /* bawrite on put */
 203 #define HAMMER2_CHAIN_DEFERRED          0x00000200      /* on a deferral list */
 204 #define HAMMER2_CHAIN_DESTROYED         0x00000400      /* destroying inode */
 205 #define HAMMER2_CHAIN_VOLUMESYNC        0x00000800      /* needs volume sync */
 206 #define HAMMER2_CHAIN_RECYCLE           0x00001000      /* force recycle */
 207 #define HAMMER2_CHAIN_MOUNTED           0x00002000      /* PFS is mounted */
 208 #define HAMMER2_CHAIN_ONRBTREE          0x00004000      /* on parent RB tree */
 209 #define HAMMER2_CHAIN_SNAPSHOT          0x00008000      /* snapshot special */
 210 #define HAMMER2_CHAIN_EMBEDDED          0x00010000      /* embedded data */
 211 #define HAMMER2_CHAIN_HARDLINK          0x00020000      /* converted to hlink */
 212 #define HAMMER2_CHAIN_REPLACE           0x00040000      /* replace bref */
 213 #define HAMMER2_CHAIN_COUNTEDBREFS      0x00080000      /* counted brefs */
 214
 215 /*
 216  * Flags passed to hammer2_chain_lookup() and hammer2_chain_next()
 217  *
 218  * NOTE: MATCHIND allows an indirect block / freemap node to be returned
 219  *       when the passed key range matches the radix.  Remember that key_end
 220  *       is inclusive (e.g. {0x000,0xFFF}, not {0x000,0x1000}).
 221  */
 222 #define HAMMER2_LOOKUP_NOLOCK           0x00000001      /* ref only */
 223 #define HAMMER2_LOOKUP_NODATA           0x00000002      /* data left NULL */
 224 #define HAMMER2_LOOKUP_SHARED           0x00000100
 225 #define HAMMER2_LOOKUP_MATCHIND         0x00000200
 226 #define HAMMER2_LOOKUP_FREEMAP          0x00000400      /* freemap base */
 227 #define HAMMER2_LOOKUP_ALWAYS           0x00000800      /* resolve data */
 228
 229 /*
 230  * Flags passed to hammer2_chain_modify() and hammer2_chain_resize()
 231  *
 232  * NOTE: OPTDATA allows us to avoid instantiating buffers for INDIRECT
 233  *       blocks in the INITIAL-create state.
 234  */
 235 #define HAMMER2_MODIFY_OPTDATA          0x00000002      /* data can be NULL */
 236 #define HAMMER2_MODIFY_NO_MODIFY_TID    0x00000004
 237 #define HAMMER2_MODIFY_ASSERTNOCOPY     0x00000008
 238 #define HAMMER2_MODIFY_NOREALLOC        0x00000010
 239
 240 /*
 241  * Flags passed to hammer2_chain_lock()
 242  */
 243 #define HAMMER2_RESOLVE_NEVER           1
 244 #define HAMMER2_RESOLVE_MAYBE           2
 245 #define HAMMER2_RESOLVE_ALWAYS          3
 246 #define HAMMER2_RESOLVE_MASK            0x0F
 247
 248 #define HAMMER2_RESOLVE_SHARED          0x10    /* request shared lock */
 249 #define HAMMER2_RESOLVE_NOREF           0x20    /* already ref'd on lock */
 250
 251 /*
 252  * Flags passed to hammer2_chain_delete()
 253  */
 254 #define HAMMER2_DELETE_WILLDUP          0x0001  /* no blk free, will be dup */
 255
 256 /*
 257  * Flags passed to hammer2_chain_delete_duplicate()
 258  */
 259 #define HAMMER2_DELDUP_RECORE           0x0001
 260
 261 /*
 262  * Cluster different types of storage together for allocations
 263  */
 264 #define HAMMER2_FREECACHE_INODE         0
 265 #define HAMMER2_FREECACHE_INDIR         1
 266 #define HAMMER2_FREECACHE_DATA          2
 267 #define HAMMER2_FREECACHE_UNUSED3       3
 268 #define HAMMER2_FREECACHE_TYPES         4
 269
 270 /*
 271  * hammer2_freemap_alloc() block preference
 272  */
 273 #define HAMMER2_OFF_NOPREF              ((hammer2_off_t)-1)
 274
 275 /*
 276  * BMAP read-ahead maximum parameters
 277  */
 278 #define HAMMER2_BMAP_COUNT              16      /* max bmap read-ahead */
 279 #define HAMMER2_BMAP_BYTES              (HAMMER2_PBUFSIZE * HAMMER2_BMAP_COUNT)
 280
 281 /*
 282  * Misc
 283  */
 284 #define HAMMER2_FLUSH_DEPTH_LIMIT       40      /* stack recursion limit */
 285
 286 /*
 287  * HAMMER2 IN-MEMORY CACHE OF MEDIA STRUCTURES
 288  *
 289  * There is an in-memory representation of all on-media data structure.
 290  *
 291  * When accessed read-only the data will be mapped to the related buffer
 292  * cache buffer.
 293  *
 294  * When accessed read-write (marked modified) a kmalloc()'d copy of the
 295  * is created which can then be modified.  The copy is destroyed when a
 296  * filesystem block is allocated to replace it.
 297  *
 298  * Active inodes (those with vnodes attached) will maintain the kmalloc()'d
 299  * copy for both the read-only and the read-write case.  The combination of
 300  * (bp) and (data) determines whether (data) was allocated or not.
 301  *
 302  * The in-memory representation may remain cached (for example in order to
 303  * placemark clustering locks) even after the related data has been
 304  * detached.
 305  */
 306
 307 RB_HEAD(hammer2_inode_tree, hammer2_inode);
 308
 309 /*
 310  * A hammer2 inode.
 311  *
 312  * NOTE: The inode's attribute CST which is also used to lock the inode
 313  *       is embedded in the chain (chain.cst) and aliased w/ attr_cst.
 314  */
 315 struct hammer2_inode {
 316         RB_ENTRY(hammer2_inode) rbnode;         /* inumber lookup (HL) */
 317         ccms_cst_t              topo_cst;       /* directory topology cst */
 318         struct hammer2_pfsmount *pmp;           /* PFS mount */
 319         struct hammer2_inode    *pip;           /* parent inode */
 320         struct vnode            *vp;
 321         hammer2_chain_t         *chain;         /* NOTE: rehomed on rename */
 322         struct lockf            advlock;
 323         hammer2_tid_t           inum;
 324         u_int                   flags;
 325         u_int                   refs;           /* +vpref, +flushref */
 326         uint8_t                 comp_heuristic;
 327         hammer2_off_t           size;
 328         uint64_t                mtime;
 329 };
 330
 331 typedef struct hammer2_inode hammer2_inode_t;
 332
 333 #define HAMMER2_INODE_MODIFIED          0x0001
 334 #define HAMMER2_INODE_SROOT             0x0002  /* kmalloc special case */
 335 #define HAMMER2_INODE_RENAME_INPROG     0x0004
 336 #define HAMMER2_INODE_ONRBTREE          0x0008
 337 #define HAMMER2_INODE_RESIZED           0x0010
 338 #define HAMMER2_INODE_MTIME             0x0020
 339
 340 int hammer2_inode_cmp(hammer2_inode_t *ip1, hammer2_inode_t *ip2);
 341 RB_PROTOTYPE2(hammer2_inode_tree, hammer2_inode, rbnode, hammer2_inode_cmp,
 342                 hammer2_tid_t);
 343
 344 /*
 345  * A hammer2 transaction and flush sequencing structure.
 346  *
 347  * This global structure is tied into hammer2_mount and is used
 348  * to sequence modifying operations and flushes.
 349  *
 350  * (a) Any modifying operations with sync_tid >= flush_tid will stall until
 351  *     all modifying operating with sync_tid < flush_tid complete.
 352  *
 353  *     The flush related to flush_tid stalls until all modifying operations
 354  *     with sync_tid < flush_tid complete.
 355  *
 356  * (b) Once unstalled, modifying operations with sync_tid > flush_tid are
 357  *     allowed to run.  All modifications cause modify/duplicate operations
 358  *     to occur on the related chains.  Note that most INDIRECT blocks will
 359  *     be unaffected because the modifications just overload the RBTREE
 360  *     structurally instead of actually modifying the indirect blocks.
 361  *
 362  * (c) The actual flush unstalls and RUNS CONCURRENTLY with (b), but only
 363  *     utilizes the chain structures with sync_tid <= flush_tid.  The
 364  *     flush will modify related indirect blocks and inodes in-place
 365  *     (rather than duplicate) since the adjustments are compatible with
 366  *     (b)'s RBTREE overloading
 367  *
 368  *     SPECIAL NOTE:  Inode modifications have to also propagate along any
 369  *                    modify/duplicate chains.  File writes detect the flush
 370  *                    and force out the conflicting buffer cache buffer(s)
 371  *                    before reusing them.
 372  *
 373  * (d) Snapshots can be made instantly but must be flushed and disconnected
 374  *     from their duplicative source before they can be mounted.  This is
 375  *     because while H2's on-media structure supports forks, its in-memory
 376  *     structure only supports very simple forking for background flushing
 377  *     purposes.
 378  *
 379  * TODO: Flush merging.  When fsync() is called on multiple discrete files
 380  *       concurrently there is no reason to stall the second fsync.
 381  *       The final flush that reaches to root can cover both fsync()s.
 382  *
 383  *     The chains typically terminate as they fly onto the disk.  The flush
 384  *     ultimately reaches the volume header.
 385  */
 386 struct hammer2_trans {
 387         TAILQ_ENTRY(hammer2_trans) entry;
 388         struct hammer2_pfsmount *pmp;
 389         hammer2_tid_t           sync_tid;
 390         thread_t                td;             /* pointer */
 391         int                     flags;
 392         int                     blocked;
 393         uint8_t                 inodes_created;
 394         uint8_t                 dummy[7];
 395 };
 396
 397 typedef struct hammer2_trans hammer2_trans_t;
 398
 399 #define HAMMER2_TRANS_ISFLUSH           0x0001  /* formal flush */
 400 #define HAMMER2_TRANS_RESTRICTED        0x0002  /* snapshot flush restrict */
 401 #define HAMMER2_TRANS_BUFCACHE          0x0004  /* from bioq strategy write */
 402
 403 #define HAMMER2_FREEMAP_HEUR_NRADIX     4       /* pwr 2 PBUFRADIX-MINIORADIX */
 404 #define HAMMER2_FREEMAP_HEUR_TYPES      8
 405 #define HAMMER2_FREEMAP_HEUR            (HAMMER2_FREEMAP_HEUR_NRADIX * \
 406                                          HAMMER2_FREEMAP_HEUR_TYPES)
 407
 408 /*
 409  * Global (per device) mount structure for device (aka vp->v_mount->hmp)
 410  */
 411 TAILQ_HEAD(hammer2_trans_queue, hammer2_trans);
 412
 413 struct hammer2_mount {
 414         struct vnode    *devvp;         /* device vnode */
 415         int             ronly;          /* read-only mount */
 416         int             pmp_count;      /* PFS mounts backed by us */
 417         TAILQ_ENTRY(hammer2_mount) mntentry; /* hammer2_mntlist */
 418
 419         struct malloc_type *mchain;
 420         int             nipstacks;
 421         int             maxipstacks;
 422         hammer2_chain_t vchain;         /* anchor chain */
 423         hammer2_chain_t fchain;         /* freemap chain special */
 424         hammer2_chain_t *schain;        /* super-root */
 425         hammer2_inode_t *sroot;         /* super-root inode */
 426         struct lock     alloclk;        /* lockmgr lock */
 427         struct lock     voldatalk;      /* lockmgr lock */
 428         struct hammer2_trans_queue transq; /* all in-progress transactions */
 429         hammer2_trans_t *curflush;      /* current flush in progress */
 430         hammer2_tid_t   topo_flush_tid; /* currently synchronizing flush pt */
 431         hammer2_tid_t   free_flush_tid; /* currently synchronizing flush pt */
 432         hammer2_off_t   heur_freemap[HAMMER2_FREEMAP_HEUR];
 433         int             flushcnt;       /* #of flush trans on the list */
 434
 435         int             volhdrno;       /* last volhdrno written */
 436         hammer2_volume_data_t voldata;
 437         hammer2_volume_data_t volsync;  /* synchronized voldata */
 438         struct bio_queue_head wthread_bioq; /* bio queue for write thread */
 439         struct mtx wthread_mtx;     /* mutex for write thread */
 440         int     wthread_destroy;    /* to control the write thread */
 441 };
 442
 443 typedef struct hammer2_mount hammer2_mount_t;
 444
 445 /*
 446  * HAMMER2 cluster - a device/root associated with a PFS.
 447  *
 448  * A PFS may have several hammer2_cluster's associated with it.
 449  */
 450 struct hammer2_cluster {
 451         struct hammer2_mount    *hmp;           /* device global mount */
 452         hammer2_chain_t         *rchain;        /* PFS root chain */
 453 };
 454
 455 typedef struct hammer2_cluster hammer2_cluster_t;
 456
 457 /*
 458  * HAMMER2 PFS mount point structure (aka vp->v_mount->mnt_data).
 459  *
 460  * This structure represents a cluster mount and not necessarily a
 461  * PFS under a specific device mount (HMP).  The distinction is important
 462  * because the elements backing a cluster mount can change on the fly.
 463  */
 464 struct hammer2_pfsmount {
 465         struct mount            *mp;            /* kernel mount */
 466         hammer2_cluster_t       *mount_cluster;
 467         hammer2_cluster_t       *cluster;
 468         hammer2_inode_t         *iroot;         /* PFS root inode */
 469         hammer2_off_t           inode_count;    /* copy of inode_count */
 470         ccms_domain_t           ccms_dom;
 471         struct netexport        export;         /* nfs export */
 472         int                     ronly;          /* read-only mount */
 473         struct malloc_type      *minode;
 474         struct malloc_type      *mmsg;
 475         kdmsg_iocom_t           iocom;
 476         struct spinlock         inum_spin;      /* inumber lookup */
 477         struct hammer2_inode_tree inum_tree;
 478         long                    inmem_inodes;
 479         long                    inmem_chains;
 480         int                     inmem_waiting;
 481 };
 482
 483 typedef struct hammer2_pfsmount hammer2_pfsmount_t;
 484
 485 struct hammer2_cbinfo {
 486         hammer2_chain_t *chain;
 487         void (*func)(hammer2_chain_t *, struct buf *, char *, void *);
 488         void *arg;
 489         size_t boff;
 490 };
 491
 492 typedef struct hammer2_cbinfo hammer2_cbinfo_t;
 493
 494 #if defined(_KERNEL)
 495
 496 MALLOC_DECLARE(M_HAMMER2);
 497
 498 #define VTOI(vp)        ((hammer2_inode_t *)(vp)->v_data)
 499 #define ITOV(ip)        ((ip)->vp)
 500
 501 /*
 502  * Currently locked chains retain the locked buffer cache buffer for
 503  * indirect blocks, and indirect blocks can be one of two sizes.  The
 504  * device buffer has to match the case to avoid deadlocking recursive
 505  * chains that might otherwise try to access different offsets within
 506  * the same device buffer.
 507  */
 508 static __inline
 509 int
 510 hammer2_devblkradix(int radix)
 511 {
 512 #if 1
 513         if (radix <= HAMMER2_LBUFRADIX) {
 514                 return (HAMMER2_LBUFRADIX);
 515         } else {
 516                 return (HAMMER2_PBUFRADIX);
 517         }
 518 #else
 519         return (HAMMER2_PBUFRADIX);
 520 #endif
 521 }
 522
 523 static __inline
 524 size_t
 525 hammer2_devblksize(size_t bytes)
 526 {
 527 #if 1
 528         if (bytes <= HAMMER2_LBUFSIZE) {
 529                 return(HAMMER2_LBUFSIZE);
 530         } else {
 531                 KKASSERT(bytes <= HAMMER2_PBUFSIZE &&
 532                          (bytes ^ (bytes - 1)) == ((bytes << 1) - 1));
 533                 return (HAMMER2_PBUFSIZE);
 534         }
 535 #else
 536         KKASSERT(bytes <= HAMMER2_PBUFSIZE &&
 537                  (bytes ^ (bytes - 1)) == ((bytes << 1) - 1));
 538         return(HAMMER2_PBUFSIZE);
 539 #endif
 540 }
 541
 542
 543 static __inline
 544 hammer2_pfsmount_t *
 545 MPTOPMP(struct mount *mp)
 546 {
 547         return ((hammer2_pfsmount_t *)mp->mnt_data);
 548 }
 549
 550 static __inline
 551 hammer2_mount_t *
 552 MPTOHMP(struct mount *mp)
 553 {
 554         return (((hammer2_pfsmount_t *)mp->mnt_data)->cluster->hmp);
 555 }
 556
 557 static __inline
 558 int
 559 hammer2_chain_refactor_test(hammer2_chain_t *chain, int traverse_hlink)
 560 {
 561         hammer2_chain_t *next;
 562
 563         next = TAILQ_NEXT(chain, core_entry);
 564
 565         if ((chain->flags & HAMMER2_CHAIN_DELETED) &&
 566             next &&
 567             (next->flags & HAMMER2_CHAIN_SNAPSHOT) == 0) {
 568                 return (1);
 569         }
 570         if (traverse_hlink &&
 571             chain->bref.type == HAMMER2_BREF_TYPE_INODE &&
 572             (chain->flags & HAMMER2_CHAIN_HARDLINK) &&
 573             next &&
 574             (next->flags & HAMMER2_CHAIN_SNAPSHOT) == 0) {
 575                 return(1);
 576         }
 577
 578         return (0);
 579 }
 580
 581 extern struct vop_ops hammer2_vnode_vops;
 582 extern struct vop_ops hammer2_spec_vops;
 583 extern struct vop_ops hammer2_fifo_vops;
 584
 585 extern int hammer2_debug;
 586 extern int hammer2_cluster_enable;
 587 extern int hammer2_hardlink_enable;
 588 extern long hammer2_iod_file_read;
 589 extern long hammer2_iod_meta_read;
 590 extern long hammer2_iod_indr_read;
 591 extern long hammer2_iod_fmap_read;
 592 extern long hammer2_iod_volu_read;
 593 extern long hammer2_iod_file_write;
 594 extern long hammer2_iod_meta_write;
 595 extern long hammer2_iod_indr_write;
 596 extern long hammer2_iod_fmap_write;
 597 extern long hammer2_iod_volu_write;
 598 extern long hammer2_ioa_file_read;
 599 extern long hammer2_ioa_meta_read;
 600 extern long hammer2_ioa_indr_read;
 601 extern long hammer2_ioa_fmap_read;
 602 extern long hammer2_ioa_volu_read;
 603 extern long hammer2_ioa_file_write;
 604 extern long hammer2_ioa_meta_write;
 605 extern long hammer2_ioa_indr_write;
 606 extern long hammer2_ioa_fmap_write;
 607 extern long hammer2_ioa_volu_write;
 608
 609 extern struct objcache *cache_buffer_read;
 610 extern struct objcache *cache_buffer_write;
 611
 612 extern int destroy;
 613 extern int write_thread_wakeup;
 614
 615 extern mtx_t thread_protect;
 616
 617 /*
 618  * hammer2_subr.c
 619  */
 620 #define hammer2_icrc32(buf, size)       iscsi_crc32((buf), (size))
 621 #define hammer2_icrc32c(buf, size, crc) iscsi_crc32_ext((buf), (size), (crc))
 622
 623 hammer2_chain_t *hammer2_inode_lock_ex(hammer2_inode_t *ip);
 624 hammer2_chain_t *hammer2_inode_lock_sh(hammer2_inode_t *ip);
 625 void hammer2_inode_unlock_ex(hammer2_inode_t *ip, hammer2_chain_t *chain);
 626 void hammer2_inode_unlock_sh(hammer2_inode_t *ip, hammer2_chain_t *chain);
 627 void hammer2_voldata_lock(hammer2_mount_t *hmp);
 628 void hammer2_voldata_unlock(hammer2_mount_t *hmp, int modify);
 629 ccms_state_t hammer2_inode_lock_temp_release(hammer2_inode_t *ip);
 630 void hammer2_inode_lock_temp_restore(hammer2_inode_t *ip, ccms_state_t ostate);
 631 ccms_state_t hammer2_inode_lock_upgrade(hammer2_inode_t *ip);
 632 void hammer2_inode_lock_downgrade(hammer2_inode_t *ip, ccms_state_t ostate);
 633
 634 void hammer2_mount_exlock(hammer2_mount_t *hmp);
 635 void hammer2_mount_shlock(hammer2_mount_t *hmp);
 636 void hammer2_mount_unlock(hammer2_mount_t *hmp);
 637
 638 int hammer2_get_dtype(hammer2_chain_t *chain);
 639 int hammer2_get_vtype(hammer2_chain_t *chain);
 640 u_int8_t hammer2_get_obj_type(enum vtype vtype);
 641 void hammer2_time_to_timespec(u_int64_t xtime, struct timespec *ts);
 642 u_int64_t hammer2_timespec_to_time(struct timespec *ts);
 643 u_int32_t hammer2_to_unix_xid(uuid_t *uuid);
 644 void hammer2_guid_to_uuid(uuid_t *uuid, u_int32_t guid);
 645
 646 hammer2_key_t hammer2_dirhash(const unsigned char *name, size_t len);
 647 int hammer2_getradix(size_t bytes);
 648
 649 int hammer2_calc_logical(hammer2_inode_t *ip, hammer2_off_t uoff,
 650                         hammer2_key_t *lbasep, hammer2_key_t *leofp);
 651 int hammer2_calc_physical(hammer2_inode_t *ip, hammer2_key_t lbase);
 652 void hammer2_update_time(uint64_t *timep);
 653
 654 /*
 655  * hammer2_inode.c
 656  */
 657 struct vnode *hammer2_igetv(hammer2_inode_t *ip, int *errorp);
 658
 659 void hammer2_inode_lock_nlinks(hammer2_inode_t *ip);
 660 void hammer2_inode_unlock_nlinks(hammer2_inode_t *ip);
 661 hammer2_inode_t *hammer2_inode_lookup(hammer2_pfsmount_t *pmp,
 662                         hammer2_tid_t inum);
 663 hammer2_inode_t *hammer2_inode_get(hammer2_pfsmount_t *pmp,
 664                         hammer2_inode_t *dip, hammer2_chain_t *chain);
 665 void hammer2_inode_free(hammer2_inode_t *ip);
 666 void hammer2_inode_ref(hammer2_inode_t *ip);
 667 void hammer2_inode_drop(hammer2_inode_t *ip);
 668 void hammer2_inode_repoint(hammer2_inode_t *ip, hammer2_inode_t *pip,
 669                         hammer2_chain_t *chain);
 670
 671 hammer2_inode_t *hammer2_inode_create(hammer2_trans_t *trans,
 672                         hammer2_inode_t *dip,
 673                         struct vattr *vap, struct ucred *cred,
 674                         const uint8_t *name, size_t name_len,
 675                         hammer2_chain_t **chainp, int *errorp);
 676 int hammer2_inode_connect(hammer2_trans_t *trans, int hlink,
 677                         hammer2_inode_t *dip, hammer2_chain_t **chainp,
 678                         const uint8_t *name, size_t name_len);
 679 hammer2_inode_t *hammer2_inode_common_parent(hammer2_inode_t *fdip,
 680                         hammer2_inode_t *tdip);
 681 void hammer2_inode_fsync(hammer2_trans_t *trans, hammer2_inode_t *ip,
 682                         hammer2_chain_t **parentp);
 683 int hammer2_unlink_file(hammer2_trans_t *trans, hammer2_inode_t *dip,
 684                         const uint8_t *name, size_t name_len, int isdir,
 685                         int *hlinkp);
 686 int hammer2_hardlink_consolidate(hammer2_trans_t *trans, hammer2_inode_t *ip,
 687                         hammer2_chain_t **chainp,
 688                         hammer2_inode_t *tdip, int linkcnt);
 689 int hammer2_hardlink_deconsolidate(hammer2_trans_t *trans, hammer2_inode_t *dip,
 690                         hammer2_chain_t **chainp, hammer2_chain_t **ochainp);
 691 int hammer2_hardlink_find(hammer2_inode_t *dip,
 692                         hammer2_chain_t **chainp, hammer2_chain_t **ochainp);
 693
 694 /*
 695  * hammer2_chain.c
 696  */
 697 void hammer2_modify_volume(hammer2_mount_t *hmp);
 698 hammer2_chain_t *hammer2_chain_alloc(hammer2_mount_t *hmp, hammer2_pfsmount_t *pmp,
 699                                 hammer2_trans_t *trans, hammer2_blockref_t *bref);
 700 void hammer2_chain_core_alloc(hammer2_trans_t *trans, hammer2_chain_t *nchain,
 701                                 hammer2_chain_t *ochain);
 702 void hammer2_chain_ref(hammer2_chain_t *chain);
 703 void hammer2_chain_drop(hammer2_chain_t *chain);
 704 int hammer2_chain_lock(hammer2_chain_t *chain, int how);
 705 void hammer2_chain_load_async(hammer2_chain_t *chain,
 706                                 void (*func)(hammer2_chain_t *, struct buf *,
 707                                              char *, void *),
 708                                 void *arg);
 709 void hammer2_chain_moved(hammer2_chain_t *chain);
 710 void hammer2_chain_modify(hammer2_trans_t *trans,
 711                                 hammer2_chain_t **chainp, int flags);
 712 hammer2_inode_data_t *hammer2_chain_modify_ip(hammer2_trans_t *trans,
 713                                 hammer2_inode_t *ip, hammer2_chain_t **chainp,
 714                                 int flags);
 715 void hammer2_chain_resize(hammer2_trans_t *trans, hammer2_inode_t *ip,
 716                                 hammer2_chain_t *parent,
 717                                 hammer2_chain_t **chainp,
 718                                 int nradix, int flags);
 719 void hammer2_chain_unlock(hammer2_chain_t *chain);
 720 void hammer2_chain_wait(hammer2_chain_t *chain);
 721 hammer2_chain_t *hammer2_chain_get(hammer2_chain_t *parent,
 722                                 hammer2_blockref_t *bref);
 723 hammer2_chain_t *hammer2_chain_lookup_init(hammer2_chain_t *parent, int flags);
 724 void hammer2_chain_lookup_done(hammer2_chain_t *parent);
 725 hammer2_chain_t *hammer2_chain_lookup(hammer2_chain_t **parentp,
 726                                 hammer2_key_t *key_nextp,
 727                                 hammer2_key_t key_beg, hammer2_key_t key_end,
 728                                 int *cache_indexp, int flags);
 729 hammer2_chain_t *hammer2_chain_next(hammer2_chain_t **parentp,
 730                                 hammer2_chain_t *chain,
 731                                 hammer2_key_t *key_nextp,
 732                                 hammer2_key_t key_beg, hammer2_key_t key_end,
 733                                 int *cache_indexp, int flags);
 734
 735 int hammer2_chain_create(hammer2_trans_t *trans,
 736                                 hammer2_chain_t **parentp,
 737                                 hammer2_chain_t **chainp,
 738                                 hammer2_key_t key, int keybits,
 739                                 int type, size_t bytes);
 740 void hammer2_chain_duplicate(hammer2_trans_t *trans, hammer2_chain_t *parent,
 741                                 hammer2_chain_t **chainp,
 742                                 hammer2_blockref_t *bref);
 743 int hammer2_chain_snapshot(hammer2_trans_t *trans, hammer2_inode_t *ip,
 744                                 hammer2_ioc_pfs_t *pfs);
 745 void hammer2_chain_delete(hammer2_trans_t *trans, hammer2_chain_t *chain,
 746                                 int flags);
 747 void hammer2_chain_delete_duplicate(hammer2_trans_t *trans,
 748                                 hammer2_chain_t **chainp, int flags);
 749 void hammer2_chain_flush(hammer2_trans_t *trans, hammer2_chain_t *chain);
 750 void hammer2_chain_commit(hammer2_trans_t *trans, hammer2_chain_t *chain);
 751 void hammer2_chain_setsubmod(hammer2_trans_t *trans, hammer2_chain_t *chain);
 752
 753 void hammer2_chain_memory_wait(hammer2_pfsmount_t *pmp);
 754 void hammer2_chain_memory_wakeup(hammer2_pfsmount_t *pmp);
 755 void hammer2_chain_countbrefs(hammer2_chain_t *chain,
 756                                 hammer2_blockref_t *base, int count);
 757 void hammer2_chain_layer_check_locked(hammer2_mount_t *hmp,
 758                                 hammer2_chain_core_t *core);
 759
 760 int hammer2_base_find(hammer2_chain_t *chain,
 761                                 hammer2_blockref_t *base, int count,
 762                                 int *cache_indexp, hammer2_key_t *key_nextp,
 763                                 hammer2_key_t key_beg, hammer2_key_t key_end);
 764 void hammer2_base_delete(hammer2_chain_t *chain,
 765                                 hammer2_blockref_t *base, int count,
 766                                 int *cache_indexp, hammer2_blockref_t *elm);
 767 void hammer2_base_insert(hammer2_chain_t *chain,
 768                                 hammer2_blockref_t *base, int count,
 769                                 int *cache_indexp, hammer2_blockref_t *elm,
 770                                 int flags);
 771
 772 /*
 773  * hammer2_trans.c
 774  */
 775 void hammer2_trans_init(hammer2_trans_t *trans,
 776                         hammer2_pfsmount_t *pmp, int flags);
 777 void hammer2_trans_done(hammer2_trans_t *trans);
 778
 779 /*
 780  * hammer2_ioctl.c
 781  */
 782 int hammer2_ioctl(hammer2_inode_t *ip, u_long com, void *data,
 783                                 int fflag, struct ucred *cred);
 784
 785 /*
 786  * hammer2_msgops.c
 787  */
 788 int hammer2_msg_dbg_rcvmsg(kdmsg_msg_t *msg);
 789 int hammer2_msg_adhoc_input(kdmsg_msg_t *msg);
 790
 791 /*
 792  * hammer2_vfsops.c
 793  */
 794 void hammer2_clusterctl_wakeup(kdmsg_iocom_t *iocom);
 795 void hammer2_volconf_update(hammer2_pfsmount_t *pmp, int index);
 796 void hammer2_cluster_reconnect(hammer2_pfsmount_t *pmp, struct file *fp);
 797 void hammer2_dump_chain(hammer2_chain_t *chain, int tab, int *countp);
 798
 799 /*
 800  * hammer2_freemap.c
 801  */
 802 int hammer2_freemap_alloc(hammer2_trans_t *trans, hammer2_mount_t *hmp,
 803                                 hammer2_blockref_t *bref, size_t bytes);
 804 void hammer2_freemap_free(hammer2_trans_t *trans, hammer2_mount_t *hmp,
 805                                 hammer2_blockref_t *bref, int how);
 806
 807
 808 #endif /* !_KERNEL */
 809 #endif /* !_VFS_HAMMER2_HAMMER2_H_ */