sys/vfs/hammer/hammer.h

   1 /*
   2  * Copyright (c) 2007-2008 The DragonFly Project.  All rights reserved.
   3  *
   4  * This code is derived from software contributed to The DragonFly Project
   5  * by Matthew Dillon <dillon@backplane.com>
   6  *
   7  * Redistribution and use in source and binary forms, with or without
   8  * modification, are permitted provided that the following conditions
   9  * are met:
  10  *
  11  * 1. Redistributions of source code must retain the above copyright
  12  *    notice, this list of conditions and the following disclaimer.
  13  * 2. Redistributions in binary form must reproduce the above copyright
  14  *    notice, this list of conditions and the following disclaimer in
  15  *    the documentation and/or other materials provided with the
  16  *    distribution.
  17  * 3. Neither the name of The DragonFly Project nor the names of its
  18  *    contributors may be used to endorse or promote products derived
  19  *    from this software without specific, prior written permission.
  20  *
  21  * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
  22  * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
  23  * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS
  24  * FOR A PARTICULAR PURPOSE ARE DISCLAIMED.  IN NO EVENT SHALL THE
  25  * COPYRIGHT HOLDERS OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT,
  26  * INCIDENTAL, SPECIAL, EXEMPLARY OR CONSEQUENTIAL DAMAGES (INCLUDING,
  27  * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
  28  * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED
  29  * AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
  30  * OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT
  31  * OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
  32  * SUCH DAMAGE.
  33  */
  34
  35 #ifndef VFS_HAMMER_HAMMER_H_
  36 #define VFS_HAMMER_HAMMER_H_
  37
  38 /*
  39  * This header file contains structures used internally by the HAMMERFS
  40  * implementation.  See hammer_disk.h for on-disk structures.
  41  */
  42
  43 #include <sys/param.h>
  44 #ifdef _KERNEL
  45 #include <sys/kernel.h>
  46 #include <sys/systm.h>
  47 #endif
  48 #include <sys/conf.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/priv.h>
  55 #include <sys/dirent.h>
  56 #include <sys/stat.h>
  57 #include <sys/fcntl.h>
  58 #include <sys/lockf.h>
  59 #include <sys/file.h>
  60 #include <sys/event.h>
  61 #include <sys/buf.h>
  62 #include <sys/queue.h>
  63 #include <sys/ktr.h>
  64 #include <sys/limits.h>
  65 #include <sys/sysctl.h>
  66 #include <vm/swap_pager.h>
  67 #include <vm/vm_extern.h>
  68
  69 #include "hammer_disk.h"
  70 #include "hammer_mount.h"
  71 #include "hammer_ioctl.h"
  72 #include "hammer_crc.h"
  73
  74 #if defined(_KERNEL) || defined(_KERNEL_STRUCTURES)
  75
  76 MALLOC_DECLARE(M_HAMMER);
  77
  78 /*
  79  * Kernel trace
  80  */
  81 #if !defined(KTR_HAMMER)
  82 #define KTR_HAMMER      KTR_ALL
  83 #endif
  84 /* KTR_INFO_MASTER_EXTERN(hammer); */
  85
  86 /*
  87  * Misc structures
  88  */
  89 struct hammer_mount;
  90 struct hammer_inode;
  91 struct hammer_volume;
  92 struct hammer_buffer;
  93 struct hammer_node;
  94 struct hammer_undo;
  95 struct hammer_reserve;
  96 struct hammer_io;
  97
  98 /*
  99  * Key structure used for custom RB tree inode lookups.  This prototypes
 100  * the function hammer_ino_rb_tree_RB_LOOKUP_INFO(root, info).
 101  */
 102 typedef struct hammer_inode_info {
 103         int64_t         obj_id;         /* (key) object identifier */
 104         hammer_tid_t    obj_asof;       /* (key) snapshot transid or 0 */
 105         uint32_t        obj_localization; /* (key) pseudo-fs id for upper 16 bits */
 106         union {
 107                 hammer_btree_leaf_elm_t leaf;
 108         } u;
 109 } *hammer_inode_info_t;
 110
 111 typedef enum hammer_transaction_type {
 112         HAMMER_TRANS_RO,
 113         HAMMER_TRANS_STD,
 114         HAMMER_TRANS_FLS
 115 } hammer_transaction_type_t;
 116
 117 /*
 118  * HAMMER Transaction tracking
 119  */
 120 typedef struct hammer_transaction {
 121         hammer_transaction_type_t type;
 122         struct hammer_mount *hmp;
 123         hammer_tid_t    tid;
 124         uint64_t        time;
 125         uint32_t        time32;
 126         int             sync_lock_refs;
 127         int             flags;
 128         struct hammer_volume *rootvol;
 129 } *hammer_transaction_t;
 130
 131 #define HAMMER_TRANSF_NEWINODE  0x0001
 132 #define HAMMER_TRANSF_CRCDOM    0x0004  /* EDOM on CRC error, less critical */
 133
 134 /*
 135  * HAMMER locks
 136  */
 137 struct hammer_lock {
 138         volatile u_int  refs;           /* active references */
 139         volatile u_int  lockval;        /* lock count and control bits */
 140         struct thread   *lowner;        /* owner if exclusively held */
 141         struct thread   *rowner;        /* owner if exclusively held */
 142 };
 143
 144 #define HAMMER_REFS_LOCKED      0x40000000      /* transition check */
 145 #define HAMMER_REFS_WANTED      0x20000000      /* transition check */
 146 #define HAMMER_REFS_CHECK       0x10000000      /* transition check */
 147
 148 #define HAMMER_REFS_FLAGS       (HAMMER_REFS_LOCKED | \
 149                                  HAMMER_REFS_WANTED | \
 150                                  HAMMER_REFS_CHECK)
 151
 152 #define HAMMER_LOCKF_EXCLUSIVE  0x40000000
 153 #define HAMMER_LOCKF_WANTED     0x20000000
 154
 155 #define HAMMER_LIMIT_RECLAIMS   16384           /* maximum reclaims in-prog */
 156
 157 static __inline int
 158 hammer_notlocked(struct hammer_lock *lock)
 159 {
 160         return(lock->lockval == 0);
 161 }
 162
 163 static __inline int
 164 hammer_islocked(struct hammer_lock *lock)
 165 {
 166         return(lock->lockval != 0);
 167 }
 168
 169 /*
 170  * Returns the number of refs on the object.
 171  */
 172 static __inline int
 173 hammer_isactive(struct hammer_lock *lock)
 174 {
 175         return(lock->refs & ~HAMMER_REFS_FLAGS);
 176 }
 177
 178 static __inline int
 179 hammer_oneref(struct hammer_lock *lock)
 180 {
 181         return((lock->refs & ~HAMMER_REFS_FLAGS) == 1);
 182 }
 183
 184 static __inline int
 185 hammer_norefs(struct hammer_lock *lock)
 186 {
 187         return((lock->refs & ~HAMMER_REFS_FLAGS) == 0);
 188 }
 189
 190 static __inline int
 191 hammer_norefsorlock(struct hammer_lock *lock)
 192 {
 193         return(lock->refs == 0);
 194 }
 195
 196 static __inline int
 197 hammer_refsorlock(struct hammer_lock *lock)
 198 {
 199         return(lock->refs != 0);
 200 }
 201
 202 /*
 203  * Return if we specifically own the lock exclusively.
 204  */
 205 static __inline int
 206 hammer_lock_excl_owned(struct hammer_lock *lock, thread_t td)
 207 {
 208         if ((lock->lockval & HAMMER_LOCKF_EXCLUSIVE) &&
 209             lock->lowner == td) {
 210                 return(1);
 211         }
 212         return(0);
 213 }
 214
 215 /*
 216  * Flush state, used by various structures
 217  */
 218 typedef enum hammer_inode_state {
 219         HAMMER_FST_IDLE,
 220         HAMMER_FST_SETUP,
 221         HAMMER_FST_FLUSH
 222 } hammer_inode_state_t;
 223
 224 /*
 225  * Pseudo-filesystem extended data tracking
 226  */
 227 struct hammer_pseudofs_inmem;
 228 RB_HEAD(hammer_pfs_rb_tree, hammer_pseudofs_inmem);
 229 RB_PROTOTYPE2(hammer_pfs_rb_tree, hammer_pseudofs_inmem, rb_node,
 230               hammer_pfs_rb_compare, uint32_t);
 231
 232 typedef struct hammer_pseudofs_inmem {
 233         RB_ENTRY(hammer_pseudofs_inmem) rb_node;
 234         struct hammer_lock      lock;
 235         uint32_t                localization;
 236         hammer_tid_t            create_tid;
 237         int                     flags;
 238         udev_t                  fsid_udev;
 239         struct hammer_pseudofs_data pfsd;
 240 } *hammer_pseudofs_inmem_t;
 241
 242 /*
 243  * Cache object ids.  A fixed number of objid cache structures are
 244  * created to reserve object id's for newly created files in multiples
 245  * of 100,000, localized to a particular directory, and recycled as
 246  * needed.  This allows parallel create operations in different
 247  * directories to retain fairly localized object ids which in turn
 248  * improves reblocking performance and layout.
 249  */
 250 #define OBJID_CACHE_SIZE        2048
 251 #define OBJID_CACHE_BULK_BITS   10              /* 10 bits (1024)       */
 252 #define OBJID_CACHE_BULK        (32 * 32)       /* two level (1024)     */
 253 #define OBJID_CACHE_BULK_MASK   (OBJID_CACHE_BULK - 1)
 254 #define OBJID_CACHE_BULK_MASK64 ((uint64_t)(OBJID_CACHE_BULK - 1))
 255
 256 typedef struct hammer_objid_cache {
 257         TAILQ_ENTRY(hammer_objid_cache) entry;
 258         struct hammer_inode             *dip;
 259         hammer_tid_t                    base_tid;
 260         int                             count;
 261         uint32_t                        bm0;
 262         uint32_t                        bm1[32];
 263 } *hammer_objid_cache_t;
 264
 265 /*
 266  * Associate an inode with a B-Tree node to cache search start positions
 267  */
 268 typedef struct hammer_node_cache {
 269         TAILQ_ENTRY(hammer_node_cache)  entry;
 270         struct hammer_node              *node;
 271         struct hammer_inode             *ip;
 272 } *hammer_node_cache_t;
 273
 274 TAILQ_HEAD(hammer_node_cache_list, hammer_node_cache);
 275
 276 /*
 277  * Live dedup cache
 278  */
 279 struct hammer_dedup_cache;
 280 RB_HEAD(hammer_dedup_crc_rb_tree, hammer_dedup_cache);
 281 RB_PROTOTYPE2(hammer_dedup_crc_rb_tree, hammer_dedup_cache, crc_entry,
 282                 hammer_dedup_crc_rb_compare, hammer_crc_t);
 283
 284 RB_HEAD(hammer_dedup_off_rb_tree, hammer_dedup_cache);
 285 RB_PROTOTYPE2(hammer_dedup_off_rb_tree, hammer_dedup_cache, off_entry,
 286                 hammer_dedup_off_rb_compare, hammer_off_t);
 287
 288 typedef struct hammer_dedup_cache {
 289         RB_ENTRY(hammer_dedup_cache) crc_entry;
 290         RB_ENTRY(hammer_dedup_cache) off_entry;
 291         TAILQ_ENTRY(hammer_dedup_cache) lru_entry;
 292         struct hammer_mount *hmp;
 293         int64_t obj_id;
 294         uint32_t localization;
 295         off_t file_offset;
 296         int bytes;
 297         hammer_off_t data_offset;
 298         hammer_crc_t crc;
 299 } *hammer_dedup_cache_t;
 300
 301 /*
 302  * Structure used to organize flush groups.  Flush groups must be
 303  * organized into chunks in order to avoid blowing out the UNDO FIFO.
 304  * Without this a 'sync' could end up flushing 50,000 inodes in a single
 305  * transaction.
 306  */
 307 RB_HEAD(hammer_fls_rb_tree, hammer_inode);
 308 RB_PROTOTYPE(hammer_fls_rb_tree, hammer_inode, rb_flsnode,
 309               hammer_ino_rb_compare);
 310
 311 typedef struct hammer_flush_group {
 312         TAILQ_ENTRY(hammer_flush_group) flush_entry;
 313         struct hammer_fls_rb_tree       flush_tree;
 314         int                             seq;            /* our seq no */
 315         int                             total_count;    /* record load */
 316         int                             running;        /* group is running */
 317         int                             closed;
 318         int                             refs;
 319 } *hammer_flush_group_t;
 320
 321 TAILQ_HEAD(hammer_flush_group_list, hammer_flush_group);
 322
 323 /*
 324  * Structure used to represent an inode in-memory.
 325  *
 326  * The record and data associated with an inode may be out of sync with
 327  * the disk (xDIRTY flags), or not even on the disk at all (ONDISK flag
 328  * clear).
 329  *
 330  * An inode may also hold a cache of unsynchronized records, used for
 331  * database and directories only.  Unsynchronized regular file data is
 332  * stored in the buffer cache.
 333  *
 334  * NOTE: A file which is created and destroyed within the initial
 335  * synchronization period can wind up not doing any disk I/O at all.
 336  *
 337  * Finally, an inode may cache numerous disk-referencing B-Tree cursors.
 338  */
 339 RB_HEAD(hammer_ino_rb_tree, hammer_inode);
 340 RB_PROTOTYPEX(hammer_ino_rb_tree, INFO, hammer_inode, rb_node,
 341               hammer_ino_rb_compare, hammer_inode_info_t);
 342
 343 RB_HEAD(hammer_redo_rb_tree, hammer_inode);
 344 RB_PROTOTYPE2(hammer_redo_rb_tree, hammer_inode, rb_redonode,
 345               hammer_redo_rb_compare, hammer_off_t);
 346
 347 struct hammer_record;
 348 RB_HEAD(hammer_rec_rb_tree, hammer_record);
 349 RB_PROTOTYPEX(hammer_rec_rb_tree, INFO, hammer_record, rb_node,
 350               hammer_rec_rb_compare, hammer_btree_leaf_elm_t);
 351
 352 TAILQ_HEAD(hammer_record_list, hammer_record);
 353 TAILQ_HEAD(hammer_node_list, hammer_node);
 354
 355 typedef struct hammer_inode {
 356         RB_ENTRY(hammer_inode)  rb_node;
 357         hammer_inode_state_t    flush_state;
 358         hammer_flush_group_t    flush_group;
 359         RB_ENTRY(hammer_inode)  rb_flsnode;     /* when on flush list */
 360         RB_ENTRY(hammer_inode)  rb_redonode;    /* when INODE_RDIRTY is set */
 361         struct hammer_record_list target_list;  /* target of dependant recs */
 362         int64_t                 obj_id;         /* (key) object identifier */
 363         hammer_tid_t            obj_asof;       /* (key) snapshot or 0 */
 364         uint32_t                obj_localization; /* (key) pseudo-fs id for upper 16 bits */
 365         struct hammer_mount     *hmp;
 366         hammer_objid_cache_t    objid_cache;
 367         int                     flags;
 368         int                     error;          /* flush error */
 369         int                     cursor_ip_refs; /* sanity */
 370 #if 0
 371         int                     cursor_exclreq_count;
 372 #endif
 373         int                     rsv_recs;
 374         struct vnode            *vp;
 375         hammer_pseudofs_inmem_t pfsm;
 376         struct lockf            advlock;
 377         struct hammer_lock      lock;           /* sync copy interlock */
 378         off_t                   trunc_off;
 379         struct hammer_btree_leaf_elm ino_leaf;  /* in-memory cache */
 380         struct hammer_inode_data ino_data;      /* in-memory cache */
 381         struct hammer_rec_rb_tree rec_tree;     /* in-memory cache */
 382         int                     rec_generation;
 383
 384         /*
 385          * search initiate cache
 386          * cache[0] - this inode
 387          * cache[1] - related data, the content depends on situations
 388          * cache[2] - for dip to cache ip to shortcut B-Tree search
 389          * cache[3] - related data copied from dip to a new ip's cache[1]
 390          */
 391         struct hammer_node_cache cache[4];
 392
 393         /*
 394          * When a demark is created to synchronize an inode to
 395          * disk, certain fields are copied so the front-end VOPs
 396          * can continue to run in parallel with the synchronization
 397          * occuring in the background.
 398          */
 399         int             sync_flags;             /* to-sync flags cache */
 400         off_t           sync_trunc_off;         /* to-sync truncation */
 401         off_t           save_trunc_off;         /* write optimization */
 402         struct hammer_btree_leaf_elm sync_ino_leaf; /* to-sync cache */
 403         struct hammer_inode_data sync_ino_data; /* to-sync cache */
 404         size_t          redo_count;
 405
 406         /*
 407          * Track the earliest offset in the UNDO/REDO FIFO containing
 408          * REDO records.  This is staged to the backend during flush
 409          * sequences.  While the inode is staged redo_fifo_next is used
 410          * to track the earliest offset for rotation into redo_fifo_start
 411          * on completion of the flush.
 412          */
 413         hammer_off_t    redo_fifo_start;
 414         hammer_off_t    redo_fifo_next;
 415 } *hammer_inode_t;
 416
 417 #define VTOI(vp)        ((hammer_inode_t)(vp)->v_data)
 418
 419 /*
 420  * NOTE: DDIRTY does not include atime or mtime and does not include
 421  *       write-append size changes.  SDIRTY handles write-append size
 422  *       changes.
 423  *
 424  *       REDO indicates that REDO logging is active, creating a definitive
 425  *       stream of REDO records in the UNDO/REDO log for writes and
 426  *       truncations, including boundary records when/if REDO is turned off.
 427  *       REDO is typically enabled by fsync() and turned off if excessive
 428  *       writes without an fsync() occurs.
 429  *
 430  *       RDIRTY indicates that REDO records were laid down in the UNDO/REDO
 431  *       FIFO (even if REDO is turned off some might still be active) and
 432  *       still being tracked for this inode.  See hammer_redo.c
 433  */
 434 #define HAMMER_INODE_DDIRTY     0x0001  /* in-memory ino_data is dirty */
 435                                         /* (not including atime/mtime) */
 436 #define HAMMER_INODE_RSV_INODES 0x0002  /* hmp->rsv_inodes bumped */
 437 #define HAMMER_INODE_CONN_DOWN  0x0004  /* include in downward recursion */
 438 #define HAMMER_INODE_XDIRTY     0x0008  /* in-memory records */
 439 #define HAMMER_INODE_ONDISK     0x0010  /* inode is on-disk (else not yet) */
 440 #define HAMMER_INODE_FLUSH      0x0020  /* flush on last ref */
 441 #define HAMMER_INODE_DELETED    0x0080  /* inode delete (backend) */
 442 #define HAMMER_INODE_DELONDISK  0x0100  /* delete synchronized to disk */
 443 #define HAMMER_INODE_RO         0x0200  /* read-only (because of as-of) */
 444 #define HAMMER_INODE_RECSW      0x0400  /* waiting on data record flush */
 445 #define HAMMER_INODE_DONDISK    0x0800  /* data records may be on disk */
 446 #define HAMMER_INODE_BUFS       0x1000  /* dirty high level bps present */
 447 #define HAMMER_INODE_REFLUSH    0x2000  /* flush on dependancy / reflush */
 448 #define HAMMER_INODE_RECLAIM    0x4000  /* trying to reclaim */
 449 #define HAMMER_INODE_FLUSHW     0x8000  /* someone waiting for flush */
 450
 451 #define HAMMER_INODE_TRUNCATED  0x00010000
 452 #define HAMMER_INODE_DELETING   0x00020000 /* inode delete request (frontend)*/
 453 #define HAMMER_INODE_RESIGNAL   0x00040000 /* re-signal on re-flush */
 454 #define HAMMER_INODE_ATIME      0x00100000 /* in-memory atime modified */
 455 #define HAMMER_INODE_MTIME      0x00200000 /* in-memory mtime modified */
 456 #define HAMMER_INODE_WOULDBLOCK 0x00400000 /* re-issue to new flush group */
 457 #define HAMMER_INODE_DUMMY      0x00800000 /* dummy inode covering bad file */
 458 #define HAMMER_INODE_SDIRTY     0x01000000 /* in-memory ino_data.size is dirty*/
 459 #define HAMMER_INODE_REDO       0x02000000 /* REDO logging active */
 460 #define HAMMER_INODE_RDIRTY     0x04000000 /* REDO records active in fifo */
 461 #define HAMMER_INODE_SLAVEFLUSH 0x08000000 /* being flushed by slave */
 462
 463 #define HAMMER_INODE_MODMASK    (HAMMER_INODE_DDIRTY|HAMMER_INODE_SDIRTY|   \
 464                                  HAMMER_INODE_XDIRTY|HAMMER_INODE_BUFS|     \
 465                                  HAMMER_INODE_ATIME|HAMMER_INODE_MTIME|     \
 466                                  HAMMER_INODE_TRUNCATED|HAMMER_INODE_DELETING)
 467
 468 #define HAMMER_INODE_MODMASK_NOXDIRTY   \
 469                                 (HAMMER_INODE_MODMASK & ~HAMMER_INODE_XDIRTY)
 470
 471 #define HAMMER_INODE_MODMASK_NOREDO     \
 472                                 (HAMMER_INODE_DDIRTY|                       \
 473                                  HAMMER_INODE_XDIRTY|                       \
 474                                  HAMMER_INODE_TRUNCATED|HAMMER_INODE_DELETING)
 475
 476 #define HAMMER_FLUSH_SIGNAL     0x0001
 477 #define HAMMER_FLUSH_RECURSION  0x0002
 478
 479 /*
 480  * Used by the inode reclaim code to pipeline reclaims and avoid
 481  * blowing out kernel memory or letting the flusher get too far
 482  * behind.  The reclaim wakes up when count reaches 0 or the
 483  * timer expires.
 484  */
 485 struct hammer_reclaim {
 486         TAILQ_ENTRY(hammer_reclaim) entry;
 487         int     count;
 488 };
 489
 490 /*
 491  * Track who is creating the greatest burden on the
 492  * inode cache.
 493  */
 494 struct hammer_inostats {
 495         pid_t           pid;    /* track user process */
 496         int             ltick;  /* last tick */
 497         int             count;  /* count (degenerates) */
 498 };
 499
 500 #define HAMMER_INOSTATS_HSIZE   32
 501 #define HAMMER_INOSTATS_HMASK   (HAMMER_INOSTATS_HSIZE - 1)
 502
 503 /*
 504  * Structure used to represent an unsynchronized record in-memory.  These
 505  * records typically represent directory entries.  Only non-historical
 506  * records are kept in-memory.
 507  *
 508  * Records are organized as a per-inode RB-Tree.  If the inode is not
 509  * on disk then neither are any records and the in-memory record tree
 510  * represents the entire contents of the inode.  If the inode is on disk
 511  * then the on-disk B-Tree is scanned in parallel with the in-memory
 512  * RB-Tree to synthesize the current state of the file.
 513  *
 514  * Records are also used to enforce the ordering of directory create/delete
 515  * operations.  A new inode will not be flushed to disk unless its related
 516  * directory entry is also being flushed at the same time.  A directory entry
 517  * will not be removed unless its related inode is also being removed at the
 518  * same time.
 519  */
 520 typedef enum hammer_record_type {
 521         HAMMER_MEM_RECORD_GENERAL,      /* misc record */
 522         HAMMER_MEM_RECORD_INODE,        /* inode record */
 523         HAMMER_MEM_RECORD_ADD,          /* positive memory cache record */
 524         HAMMER_MEM_RECORD_DEL,          /* negative delete-on-disk record */
 525         HAMMER_MEM_RECORD_DATA          /* bulk-data record w/on-disk ref */
 526 } hammer_record_type_t;
 527
 528 typedef struct hammer_record {
 529         RB_ENTRY(hammer_record)         rb_node;
 530         TAILQ_ENTRY(hammer_record)      target_entry;
 531         hammer_inode_state_t            flush_state;
 532         hammer_flush_group_t            flush_group;
 533         hammer_record_type_t            type;
 534         struct hammer_lock              lock;
 535         struct hammer_reserve           *resv;
 536         hammer_inode_t                  ip;
 537         hammer_inode_t                  target_ip;
 538         struct hammer_btree_leaf_elm    leaf;
 539         hammer_data_ondisk_t            data;
 540         int                             flags;
 541         int                             gflags;
 542         hammer_off_t                    zone2_offset;   /* direct-write only */
 543 } *hammer_record_t;
 544
 545 /*
 546  * Record flags.  Note that FE can only be set by the frontend if the
 547  * record has not been interlocked by the backend w/ BE.
 548  */
 549 #define HAMMER_RECF_ALLOCDATA           0x0001
 550 #define HAMMER_RECF_ONRBTREE            0x0002
 551 #define HAMMER_RECF_DELETED_FE          0x0004  /* deleted (frontend) */
 552 #define HAMMER_RECF_DELETED_BE          0x0008  /* deleted (backend) */
 553 #define HAMMER_RECF_COMMITTED           0x0010  /* committed to the B-Tree */
 554 #define HAMMER_RECF_INTERLOCK_BE        0x0020  /* backend interlock */
 555 #define HAMMER_RECF_WANTED              0x0040  /* wanted by the frontend */
 556 #define HAMMER_RECF_DEDUPED             0x0080  /* will be live-dedup'ed */
 557 #define HAMMER_RECF_CONVERT_DELETE      0x0100  /* special case */
 558 #define HAMMER_RECF_REDO                0x1000  /* REDO was laid down */
 559
 560 /*
 561  * These flags must be separate to deal with SMP races
 562  */
 563 #define HAMMER_RECG_DIRECT_IO           0x0001  /* related direct I/O running*/
 564 #define HAMMER_RECG_DIRECT_WAIT         0x0002  /* related direct I/O running*/
 565 #define HAMMER_RECG_DIRECT_INVAL        0x0004  /* buffer alias invalidation */
 566 /*
 567  * hammer_create_at_cursor() and hammer_delete_at_cursor() flags.
 568  */
 569 #define HAMMER_CREATE_MODE_UMIRROR      0x0001
 570 #define HAMMER_CREATE_MODE_SYS          0x0002
 571
 572 #define HAMMER_DELETE_ADJUST            0x0001
 573 #define HAMMER_DELETE_DESTROY           0x0002
 574
 575 /*
 576  * In-memory structures representing on-disk structures.
 577  */
 578 RB_HEAD(hammer_vol_rb_tree, hammer_volume);
 579 RB_HEAD(hammer_buf_rb_tree, hammer_buffer);
 580 RB_HEAD(hammer_nod_rb_tree, hammer_node);
 581 RB_HEAD(hammer_und_rb_tree, hammer_undo);
 582 RB_HEAD(hammer_res_rb_tree, hammer_reserve);
 583 RB_HEAD(hammer_mod_rb_tree, hammer_io);
 584
 585 RB_PROTOTYPE2(hammer_vol_rb_tree, hammer_volume, rb_node,
 586               hammer_vol_rb_compare, int32_t);
 587 RB_PROTOTYPE2(hammer_buf_rb_tree, hammer_buffer, rb_node,
 588               hammer_buf_rb_compare, hammer_off_t);
 589 RB_PROTOTYPE2(hammer_nod_rb_tree, hammer_node, rb_node,
 590               hammer_nod_rb_compare, hammer_off_t);
 591 RB_PROTOTYPE2(hammer_und_rb_tree, hammer_undo, rb_node,
 592               hammer_und_rb_compare, hammer_off_t);
 593 RB_PROTOTYPE2(hammer_res_rb_tree, hammer_reserve, rb_node,
 594               hammer_res_rb_compare, hammer_off_t);
 595 RB_PROTOTYPE2(hammer_mod_rb_tree, hammer_io, rb_node,
 596               hammer_mod_rb_compare, hammer_off_t);
 597
 598 /*
 599  * IO management - embedded at the head of various in-memory structures
 600  *
 601  * VOLUME       - hammer_volume containing meta-data
 602  * META_BUFFER  - hammer_buffer containing meta-data
 603  * UNDO_BUFFER  - hammer_buffer containing undo-data
 604  * DATA_BUFFER  - hammer_buffer containing pure-data
 605  * DUMMY        - hammer_buffer not containing valid data
 606  *
 607  * Dirty volume headers and dirty meta-data buffers are locked until the
 608  * flusher can sequence them out.  Dirty pure-data buffers can be written.
 609  * Clean buffers can be passively released.
 610  */
 611 typedef enum hammer_io_type {
 612         HAMMER_IOTYPE_VOLUME,
 613         HAMMER_IOTYPE_META_BUFFER,
 614         HAMMER_IOTYPE_UNDO_BUFFER,
 615         HAMMER_IOTYPE_DATA_BUFFER,
 616         HAMMER_IOTYPE_DUMMY
 617 } hammer_io_type_t;
 618
 619 typedef struct hammer_io {
 620         struct hammer_lock      lock;
 621         hammer_io_type_t        type;
 622         struct hammer_mount     *hmp;
 623         struct hammer_volume    *volume;
 624         RB_ENTRY(hammer_io)     rb_node;     /* if modified */
 625         TAILQ_ENTRY(hammer_io)  iorun_entry; /* iorun_list */
 626         struct hammer_mod_rb_tree *mod_root;
 627         struct buf              *bp;
 628         int64_t                 offset;    /* volume offset */
 629         int                     bytes;     /* buffer cache buffer size */
 630         int                     modify_refs;
 631
 632         /*
 633          * These can be modified at any time by the backend while holding
 634          * io_token, due to bio_done and hammer_io_complete() callbacks.
 635          */
 636         u_int           running : 1;    /* bp write IO in progress */
 637         u_int           waiting : 1;    /* someone is waiting on us */
 638         u_int           ioerror : 1;    /* abort on io-error */
 639         u_int           unusedA : 29;
 640
 641         /*
 642          * These can only be modified by the frontend while holding
 643          * fs_token, or by the backend while holding the io interlocked
 644          * with no references (which will block the frontend when it
 645          * tries to reference it).
 646          *
 647          * WARNING! SMP RACES will create havoc if the callbacks ever tried
 648          *          to modify any of these outside the above restrictions.
 649          */
 650         u_int           modified : 1;   /* bp's data was modified */
 651         u_int           released : 1;   /* bp released (w/ B_LOCKED set) */
 652         u_int           waitdep : 1;    /* flush waits for dependancies */
 653         u_int           recovered : 1;  /* has recovery ref */
 654         u_int           waitmod : 1;    /* waiting for modify_refs */
 655         u_int           reclaim : 1;    /* reclaim requested */
 656         u_int           gencrc : 1;     /* crc needs to be generated */
 657         u_int           unusedB : 25;
 658 } *hammer_io_t;
 659
 660 #define HAMMER_CLUSTER_SIZE     (64 * 1024)
 661 #if HAMMER_CLUSTER_SIZE > MAXBSIZE
 662 #undef  HAMMER_CLUSTER_SIZE
 663 #define HAMMER_CLUSTER_SIZE     MAXBSIZE
 664 #endif
 665
 666 /*
 667  * In-memory volume representing on-disk buffer
 668  */
 669 typedef struct hammer_volume {
 670         struct hammer_io io; /* must be at offset 0 */
 671         RB_ENTRY(hammer_volume) rb_node;
 672         hammer_volume_ondisk_t ondisk;
 673         int32_t vol_no;
 674         hammer_off_t maxbuf_off; /* Maximum buffer offset (zone-2) */
 675         char    *vol_name;
 676         struct vnode *devvp;
 677         int     vol_flags;
 678 } *hammer_volume_t;
 679
 680 #define HAMMER_ITOV(iop) ((hammer_volume_t)(iop))
 681
 682 /*
 683  * In-memory buffer representing an on-disk buffer.
 684  */
 685 typedef struct hammer_buffer {
 686         struct hammer_io io; /* must be at offset 0 */
 687         RB_ENTRY(hammer_buffer) rb_node;
 688         void *ondisk;
 689         hammer_off_t zoneX_offset;
 690         hammer_off_t zone2_offset;
 691         struct hammer_reserve *resv;
 692         struct hammer_node_list node_list;
 693 } *hammer_buffer_t;
 694
 695 #define HAMMER_ITOB(iop) ((hammer_buffer_t)(iop))
 696
 697 /*
 698  * In-memory B-Tree node, representing an on-disk B-Tree node.
 699  *
 700  * This is a hang-on structure which is backed by a hammer_buffer,
 701  * and used for fine-grained locking of B-Tree nodes in order to
 702  * properly control lock ordering.
 703  */
 704 typedef struct hammer_node {
 705         struct hammer_lock      lock;           /* node-by-node lock */
 706         TAILQ_ENTRY(hammer_node) entry;         /* per-buffer linkage */
 707         RB_ENTRY(hammer_node)   rb_node;        /* per-mount linkage */
 708         hammer_off_t            node_offset;    /* full offset spec */
 709         struct hammer_mount     *hmp;
 710         hammer_buffer_t         buffer;         /* backing buffer */
 711         hammer_node_ondisk_t    ondisk;         /* ptr to on-disk structure */
 712         TAILQ_HEAD(, hammer_cursor) cursor_list;  /* deadlock recovery */
 713         struct hammer_node_cache_list cache_list; /* passive caches */
 714         int                     flags;
 715 #if 0
 716         int                     cursor_exclreq_count;
 717 #endif
 718 } *hammer_node_t;
 719
 720 #define HAMMER_NODE_DELETED     0x0001
 721 #define HAMMER_NODE_FLUSH       0x0002
 722 #define HAMMER_NODE_CRCGOOD     0x0004
 723 #define HAMMER_NODE_NEEDSCRC    0x0008
 724 #define HAMMER_NODE_NEEDSMIRROR 0x0010
 725 #define HAMMER_NODE_CRCBAD      0x0020
 726 #define HAMMER_NODE_NONLINEAR   0x0040          /* linear heuristic */
 727
 728 #define HAMMER_NODE_CRCANY      (HAMMER_NODE_CRCGOOD | HAMMER_NODE_CRCBAD)
 729
 730 /*
 731  * List of locked nodes.  This structure is used to lock potentially large
 732  * numbers of nodes as an aid for complex B-Tree operations.
 733  */
 734 struct hammer_node_lock;
 735 TAILQ_HEAD(hammer_node_lock_list, hammer_node_lock);
 736
 737 typedef struct hammer_node_lock {
 738         TAILQ_ENTRY(hammer_node_lock) entry;
 739         struct hammer_node_lock_list  list;
 740         struct hammer_node_lock       *parent;
 741         hammer_node_t   node;
 742         hammer_node_ondisk_t copy;      /* copy of on-disk data */
 743         int             index;          /* index of this node in parent */
 744         int             count;          /* count children */
 745         int             flags;
 746 } *hammer_node_lock_t;
 747
 748 #define HAMMER_NODE_LOCK_UPDATED        0x0001
 749 #define HAMMER_NODE_LOCK_LCACHE         0x0002
 750
 751 /*
 752  * The reserve structure prevents the blockmap from allocating
 753  * out of a reserved big-block.  Such reservations are used by
 754  * the direct-write mechanism.
 755  *
 756  * The structure is also used to hold off on reallocations of
 757  * big-blocks from the freemap until flush dependancies have
 758  * been dealt with.
 759  */
 760 typedef struct hammer_reserve {
 761         RB_ENTRY(hammer_reserve) rb_node;
 762         TAILQ_ENTRY(hammer_reserve) delay_entry;
 763         int             flg_no;
 764         int             flags;
 765         int             refs;
 766         int             zone;
 767         int             append_off;
 768         int32_t         bytes_free;
 769         hammer_off_t    zone_offset;
 770 } *hammer_reserve_t;
 771
 772 #define HAMMER_RESF_ONDELAY     0x0001
 773 #define HAMMER_RESF_LAYER2FREE  0x0002
 774
 775 #include "hammer_cursor.h"
 776
 777 /*
 778  * The undo structure tracks recent undos to avoid laying down duplicate
 779  * undos within a flush group, saving us a significant amount of overhead.
 780  *
 781  * This is strictly a heuristic.
 782  */
 783 #define HAMMER_MAX_UNDOS                1024
 784 #define HAMMER_MAX_FLUSHERS             4
 785
 786 typedef struct hammer_undo {
 787         RB_ENTRY(hammer_undo)   rb_node;
 788         TAILQ_ENTRY(hammer_undo) lru_entry;
 789         hammer_off_t            offset;
 790         int                     bytes;
 791 } *hammer_undo_t;
 792
 793 struct hammer_flusher_info;
 794 TAILQ_HEAD(hammer_flusher_info_list, hammer_flusher_info);
 795
 796 struct hammer_flusher {
 797         int             signal;         /* flusher thread sequencer */
 798         int             done;           /* last completed flush group */
 799         int             next;           /* next unallocated flg seqno */
 800         int             group_lock;     /* lock sequencing of the next flush */
 801         int             exiting;        /* request master exit */
 802         thread_t        td;             /* master flusher thread */
 803         hammer_tid_t    tid;            /* last flushed transaction id */
 804         int             finalize_want;          /* serialize finalization */
 805         struct hammer_lock finalize_lock;       /* serialize finalization */
 806         struct hammer_transaction trans;        /* shared transaction */
 807         struct hammer_flusher_info_list run_list;
 808         struct hammer_flusher_info_list ready_list;
 809 };
 810
 811 #define HAMMER_FLUSH_UNDOS_RELAXED      0
 812 #define HAMMER_FLUSH_UNDOS_FORCED       1
 813 #define HAMMER_FLUSH_UNDOS_AUTO         2
 814 /*
 815  * Internal hammer mount data structure
 816  */
 817 typedef struct hammer_mount {
 818         struct mount *mp;
 819         struct hammer_ino_rb_tree rb_inos_root;
 820         struct hammer_redo_rb_tree rb_redo_root;
 821         struct hammer_vol_rb_tree rb_vols_root;
 822         struct hammer_nod_rb_tree rb_nods_root;
 823         struct hammer_und_rb_tree rb_undo_root;
 824         struct hammer_res_rb_tree rb_resv_root;
 825         struct hammer_buf_rb_tree rb_bufs_root;
 826         struct hammer_pfs_rb_tree rb_pfsm_root;
 827
 828         struct hammer_dedup_crc_rb_tree rb_dedup_crc_root;
 829         struct hammer_dedup_off_rb_tree rb_dedup_off_root;
 830
 831         hammer_volume_t rootvol;
 832         struct hammer_base_elm root_btree_beg;
 833         struct hammer_base_elm root_btree_end;
 834
 835         struct malloc_type      *m_misc;
 836         struct malloc_type      *m_inodes;
 837
 838         int     flags;          /* HAMMER_MOUNT_xxx flags */
 839         int     hflags;
 840         int     ronly;
 841         int     nvolumes;
 842         int     master_id;      /* default 0, no-mirror -1, otherwise 1-15 */
 843         int     version;        /* hammer filesystem version to use */
 844         int     rsv_inodes;     /* reserved space due to dirty inodes */
 845         int64_t rsv_databytes;  /* reserved space due to record data */
 846         int     rsv_recs;       /* reserved space due to dirty records */
 847         int     rsv_fromdelay;  /* big-blocks reserved due to flush delay */
 848         int     undo_rec_limit; /* based on size of undo area */
 849
 850         int     volume_to_remove; /* volume that is currently being removed */
 851
 852         int     count_inodes;   /* total number of inodes */
 853         int     count_iqueued;  /* inodes queued to flusher */
 854         int     count_reclaims; /* inodes pending reclaim by flusher */
 855
 856         struct hammer_flusher flusher;
 857
 858         u_int   check_interrupt;
 859         u_int   check_yield;
 860         uuid_t  fsid;
 861         struct hammer_mod_rb_tree volu_root;    /* dirty undo buffers */
 862         struct hammer_mod_rb_tree undo_root;    /* dirty undo buffers */
 863         struct hammer_mod_rb_tree data_root;    /* dirty data buffers */
 864         struct hammer_mod_rb_tree meta_root;    /* dirty meta bufs    */
 865         struct hammer_mod_rb_tree lose_root;    /* loose buffers      */
 866         long    locked_dirty_space;             /* meta/volu count    */
 867         long    io_running_space;               /* io_token */
 868         int     objid_cache_count;
 869         int     dedup_cache_count;
 870         int     error;                          /* critical I/O error */
 871         struct krate    krate;                  /* rate limited kprintf */
 872         struct krate    kdiag;                  /* rate limited kprintf */
 873         hammer_tid_t    asof;                   /* snapshot mount */
 874         hammer_tid_t    next_tid;
 875         hammer_tid_t    flush_tid1;             /* flusher tid sequencing */
 876         hammer_tid_t    flush_tid2;             /* flusher tid sequencing */
 877         int64_t copy_stat_freebigblocks;        /* number of free big-blocks */
 878         uint32_t        undo_seqno;             /* UNDO/REDO FIFO seqno */
 879         uint32_t        recover_stage2_seqno;   /* REDO recovery seqno */
 880         hammer_off_t    recover_stage2_offset;  /* REDO recovery offset */
 881
 882         struct netexport export;
 883         struct hammer_lock sync_lock;
 884         struct hammer_lock undo_lock;
 885         struct hammer_lock blkmap_lock;
 886         struct hammer_lock snapshot_lock;
 887         struct hammer_lock volume_lock;
 888         struct hammer_blockmap  blockmap[HAMMER_MAX_ZONES];
 889         struct hammer_undo      undos[HAMMER_MAX_UNDOS];
 890         int                     undo_alloc;
 891         TAILQ_HEAD(, hammer_undo)  undo_lru_list;
 892         TAILQ_HEAD(, hammer_reserve) delay_list;
 893         struct hammer_flush_group_list  flush_group_list;
 894         hammer_flush_group_t    fill_flush_group;
 895         hammer_flush_group_t    next_flush_group;
 896         TAILQ_HEAD(, hammer_objid_cache) objid_cache_list;
 897         TAILQ_HEAD(, hammer_dedup_cache) dedup_lru_list;
 898         hammer_dedup_cache_t    dedup_free_cache;
 899         TAILQ_HEAD(, hammer_reclaim) reclaim_list;
 900         TAILQ_HEAD(, hammer_io) iorun_list;
 901
 902         struct lwkt_token       fs_token;       /* high level */
 903         struct lwkt_token       io_token;       /* low level (IO callback) */
 904
 905         struct hammer_inostats  inostats[HAMMER_INOSTATS_HSIZE];
 906         uint64_t volume_map[4];  /* 256 bits bitfield */
 907 } *hammer_mount_t;
 908
 909 #define HAMMER_MOUNT_CRITICAL_ERROR     0x0001
 910 #define HAMMER_MOUNT_FLUSH_RECOVERY     0x0002
 911 #define HAMMER_MOUNT_REDO_SYNC          0x0004
 912 #define HAMMER_MOUNT_REDO_RECOVERY_REQ  0x0008
 913 #define HAMMER_MOUNT_REDO_RECOVERY_RUN  0x0010
 914
 915 #define HAMMER_VOLUME_NUMBER_FOREACH(hmp, n)            \
 916         for (n = 0; n < HAMMER_MAX_VOLUMES; n++)        \
 917                 if (hammer_volume_number_test(hmp, n))
 918
 919 /*
 920  * Minium buffer cache bufs required to rebalance the B-Tree.
 921  * This is because we must hold the children and the children's children
 922  * locked.  Even this might not be enough if things are horribly out
 923  * of balance.
 924  */
 925 #define HAMMER_REBALANCE_MIN_BUFS       \
 926         (HAMMER_BTREE_LEAF_ELMS * HAMMER_BTREE_LEAF_ELMS)
 927
 928 #endif  /* _KERNEL || _KERNEL_STRUCTURES */
 929
 930 #if defined(_KERNEL)
 931 /*
 932  * checkspace slop (8MB chunks), higher numbers are more conservative.
 933  */
 934 #define HAMMER_CHKSPC_REBLOCK   25
 935 #define HAMMER_CHKSPC_MIRROR    20
 936 #define HAMMER_CHKSPC_WRITE     20
 937 #define HAMMER_CHKSPC_CREATE    20
 938 #define HAMMER_CHKSPC_REMOVE    10
 939 #define HAMMER_CHKSPC_EMERGENCY 0
 940
 941 extern struct vop_ops hammer_vnode_vops;
 942 extern struct vop_ops hammer_spec_vops;
 943 extern struct vop_ops hammer_fifo_vops;
 944
 945 extern int hammer_debug_io;
 946 extern int hammer_debug_general;
 947 extern int hammer_debug_inode;
 948 extern int hammer_debug_locks;
 949 extern int hammer_debug_btree;
 950 extern int hammer_debug_tid;
 951 extern int hammer_debug_recover;
 952 extern int hammer_debug_critical;
 953 extern int hammer_cluster_enable;
 954 extern int hammer_live_dedup;
 955 extern int hammer_tdmux_ticks;
 956 extern int hammer_count_fsyncs;
 957 extern int hammer_count_inodes;
 958 extern int hammer_count_iqueued;
 959 extern int hammer_count_reclaims;
 960 extern int hammer_count_records;
 961 extern int hammer_count_record_datas;
 962 extern int hammer_count_volumes;
 963 extern int hammer_count_buffers;
 964 extern int hammer_count_nodes;
 965 extern int64_t hammer_stats_btree_lookups;
 966 extern int64_t hammer_stats_btree_searches;
 967 extern int64_t hammer_stats_btree_inserts;
 968 extern int64_t hammer_stats_btree_deletes;
 969 extern int64_t hammer_stats_btree_elements;
 970 extern int64_t hammer_stats_btree_splits;
 971 extern int64_t hammer_stats_btree_iterations;
 972 extern int64_t hammer_stats_btree_root_iterations;
 973 extern int64_t hammer_stats_record_iterations;
 974 extern int64_t hammer_stats_file_read;
 975 extern int64_t hammer_stats_file_write;
 976 extern int64_t hammer_stats_disk_read;
 977 extern int64_t hammer_stats_disk_write;
 978 extern int64_t hammer_stats_inode_flushes;
 979 extern int64_t hammer_stats_commits;
 980 extern int64_t hammer_stats_undo;
 981 extern int64_t hammer_stats_redo;
 982 extern long hammer_count_dirtybufspace;
 983 extern int hammer_count_refedbufs;
 984 extern int hammer_count_reservations;
 985 extern long hammer_count_io_running_read;
 986 extern long hammer_count_io_running_write;
 987 extern int hammer_count_io_locked;
 988 extern long hammer_limit_dirtybufspace;
 989 extern int hammer_limit_recs;
 990 extern int hammer_limit_inode_recs;
 991 extern int hammer_limit_reclaims;
 992 extern int hammer_live_dedup_cache_size;
 993 extern int hammer_limit_redo;
 994 extern int hammer_verify_zone;
 995 extern int hammer_verify_data;
 996 extern int hammer_double_buffer;
 997 extern int hammer_btree_full_undo;
 998 extern int hammer_yield_check;
 999 extern int hammer_fsync_mode;
1000 extern int hammer_autoflush;
1001 extern int64_t hammer_contention_count;
1002
1003 extern int64_t hammer_live_dedup_vnode_bcmps;
1004 extern int64_t hammer_live_dedup_device_bcmps;
1005 extern int64_t hammer_live_dedup_findblk_failures;
1006 extern int64_t hammer_live_dedup_bmap_saves;
1007
1008 void    hammer_critical_error(hammer_mount_t hmp, hammer_inode_t ip,
1009                         int error, const char *msg);
1010 int     hammer_vop_inactive(struct vop_inactive_args *);
1011 int     hammer_vop_reclaim(struct vop_reclaim_args *);
1012 int     hammer_get_vnode(hammer_inode_t ip, struct vnode **vpp);
1013 hammer_inode_t hammer_get_inode(hammer_transaction_t trans,
1014                         hammer_inode_t dip, int64_t obj_id,
1015                         hammer_tid_t asof, uint32_t localization,
1016                         int flags, int *errorp);
1017 hammer_inode_t hammer_get_dummy_inode(hammer_transaction_t trans,
1018                         hammer_inode_t dip, int64_t obj_id,
1019                         hammer_tid_t asof, uint32_t localization,
1020                         int flags, int *errorp);
1021 hammer_inode_t hammer_find_inode(hammer_transaction_t trans,
1022                         int64_t obj_id, hammer_tid_t asof,
1023                         uint32_t localization);
1024 void    hammer_scan_inode_snapshots(hammer_mount_t hmp,
1025                         hammer_inode_info_t iinfo,
1026                         int (*callback)(hammer_inode_t ip, void *data),
1027                         void *data);
1028 void    hammer_put_inode(hammer_inode_t ip);
1029 void    hammer_put_inode_ref(hammer_inode_t ip);
1030 void    hammer_inode_waitreclaims(hammer_transaction_t trans);
1031 void    hammer_inode_dirty(hammer_inode_t ip);
1032
1033 int     hammer_unload_volume(hammer_volume_t volume, void *data);
1034 int     hammer_adjust_volume_mode(hammer_volume_t volume, void *data __unused);
1035
1036 int     hammer_unload_buffer(hammer_buffer_t buffer, void *data);
1037 int     hammer_install_volume(hammer_mount_t hmp, const char *volname,
1038                         struct vnode *devvp, void *data);
1039 int     hammer_mountcheck_volumes(hammer_mount_t hmp);
1040 int     hammer_get_installed_volumes(hammer_mount_t hmp);
1041
1042 int     hammer_mem_add(hammer_record_t record);
1043 int     hammer_ip_lookup(hammer_cursor_t cursor);
1044 int     hammer_ip_first(hammer_cursor_t cursor);
1045 int     hammer_ip_next(hammer_cursor_t cursor);
1046 int     hammer_ip_resolve_data(hammer_cursor_t cursor);
1047 int     hammer_ip_delete_record(hammer_cursor_t cursor, hammer_inode_t ip,
1048                         hammer_tid_t tid);
1049 int     hammer_create_at_cursor(hammer_cursor_t cursor,
1050                         hammer_btree_leaf_elm_t leaf, void *udata, int mode);
1051 int     hammer_delete_at_cursor(hammer_cursor_t cursor, int delete_flags,
1052                         hammer_tid_t delete_tid, uint32_t delete_ts,
1053                         int track, int64_t *stat_bytes);
1054 int     hammer_ip_check_directory_empty(hammer_transaction_t trans,
1055                         hammer_inode_t ip);
1056 int     hammer_sync_hmp(hammer_mount_t hmp, int waitfor);
1057 int     hammer_queue_inodes_flusher(hammer_mount_t hmp, int waitfor);
1058
1059 hammer_record_t
1060         hammer_alloc_mem_record(hammer_inode_t ip, int data_len);
1061 void    hammer_flush_record_done(hammer_record_t record, int error);
1062 void    hammer_wait_mem_record_ident(hammer_record_t record, const char *ident);
1063 void    hammer_rel_mem_record(hammer_record_t record);
1064
1065 int     hammer_cursor_up(hammer_cursor_t cursor);
1066 int     hammer_cursor_up_locked(hammer_cursor_t cursor);
1067 int     hammer_cursor_down(hammer_cursor_t cursor);
1068 int     hammer_cursor_upgrade(hammer_cursor_t cursor);
1069 int     hammer_cursor_upgrade_node(hammer_cursor_t cursor);
1070 void    hammer_cursor_downgrade(hammer_cursor_t cursor);
1071 int     hammer_cursor_upgrade2(hammer_cursor_t c1, hammer_cursor_t c2);
1072 void    hammer_cursor_downgrade2(hammer_cursor_t c1, hammer_cursor_t c2);
1073 int     hammer_cursor_seek(hammer_cursor_t cursor, hammer_node_t node,
1074                         int index);
1075 void    hammer_lock_ex_ident(struct hammer_lock *lock, const char *ident);
1076 int     hammer_lock_ex_try(struct hammer_lock *lock);
1077 void    hammer_lock_sh(struct hammer_lock *lock);
1078 int     hammer_lock_sh_try(struct hammer_lock *lock);
1079 int     hammer_lock_upgrade(struct hammer_lock *lock, int shcount);
1080 void    hammer_lock_downgrade(struct hammer_lock *lock, int shcount);
1081 int     hammer_lock_status(struct hammer_lock *lock);
1082 void    hammer_unlock(struct hammer_lock *lock);
1083 void    hammer_ref(struct hammer_lock *lock);
1084 int     hammer_ref_interlock(struct hammer_lock *lock);
1085 int     hammer_ref_interlock_true(struct hammer_lock *lock);
1086 void    hammer_ref_interlock_done(struct hammer_lock *lock);
1087 void    hammer_rel(struct hammer_lock *lock);
1088 int     hammer_rel_interlock(struct hammer_lock *lock, int locked);
1089 void    hammer_rel_interlock_done(struct hammer_lock *lock, int orig_locked);
1090 int     hammer_get_interlock(struct hammer_lock *lock);
1091 int     hammer_try_interlock_norefs(struct hammer_lock *lock);
1092 void    hammer_put_interlock(struct hammer_lock *lock, int error);
1093
1094 void    hammer_sync_lock_ex(hammer_transaction_t trans);
1095 void    hammer_sync_lock_sh(hammer_transaction_t trans);
1096 int     hammer_sync_lock_sh_try(hammer_transaction_t trans);
1097 void    hammer_sync_unlock(hammer_transaction_t trans);
1098
1099 uint32_t hammer_to_unix_xid(uuid_t *uuid);
1100 void hammer_guid_to_uuid(uuid_t *uuid, uint32_t guid);
1101 void    hammer_time_to_timespec(uint64_t xtime, struct timespec *ts);
1102 uint64_t hammer_timespec_to_time(struct timespec *ts);
1103 int     hammer_str_to_tid(const char *str, int *ispfsp,
1104                         hammer_tid_t *tidp, uint32_t *localizationp);
1105 hammer_tid_t hammer_alloc_objid(hammer_mount_t hmp, hammer_inode_t dip,
1106                         int64_t namekey);
1107 void hammer_clear_objid(hammer_inode_t dip);
1108 void hammer_destroy_objid_cache(hammer_mount_t hmp);
1109
1110 int hammer_dedup_crc_rb_compare(hammer_dedup_cache_t dc1,
1111                         hammer_dedup_cache_t dc2);
1112 int hammer_dedup_off_rb_compare(hammer_dedup_cache_t dc1,
1113                         hammer_dedup_cache_t dc2);
1114 hammer_dedup_cache_t hammer_dedup_cache_add(hammer_inode_t ip,
1115                         hammer_btree_leaf_elm_t leaf);
1116 hammer_dedup_cache_t hammer_dedup_cache_lookup(hammer_mount_t hmp,
1117                         hammer_crc_t crc);
1118 void hammer_dedup_cache_inval(hammer_mount_t hmp, hammer_off_t base_offset);
1119 void hammer_destroy_dedup_cache(hammer_mount_t hmp);
1120 void hammer_dump_dedup_cache(hammer_mount_t hmp);
1121 int hammer_dedup_validate(hammer_dedup_cache_t dcp, int zone, int bytes,
1122                         void *data);
1123
1124 int hammer_enter_undo_history(hammer_mount_t hmp, hammer_off_t offset,
1125                         int bytes);
1126 void hammer_clear_undo_history(hammer_mount_t hmp);
1127 enum vtype hammer_get_vnode_type(uint8_t obj_type);
1128 int hammer_get_dtype(uint8_t obj_type);
1129 uint8_t hammer_get_obj_type(enum vtype vtype);
1130 int64_t hammer_direntry_namekey(hammer_inode_t dip, const void *name, int len,
1131                         uint32_t *max_iterationsp);
1132 int     hammer_nohistory(hammer_inode_t ip);
1133
1134 int     hammer_init_cursor(hammer_transaction_t trans, hammer_cursor_t cursor,
1135                         hammer_node_cache_t cache, hammer_inode_t ip);
1136 void    hammer_normalize_cursor(hammer_cursor_t cursor);
1137 void    hammer_done_cursor(hammer_cursor_t cursor);
1138 int     hammer_recover_cursor(hammer_cursor_t cursor);
1139 void    hammer_unlock_cursor(hammer_cursor_t cursor);
1140 int     hammer_lock_cursor(hammer_cursor_t cursor);
1141 hammer_cursor_t hammer_push_cursor(hammer_cursor_t ocursor);
1142 void    hammer_pop_cursor(hammer_cursor_t ocursor, hammer_cursor_t ncursor);
1143
1144 void    hammer_cursor_replaced_node(hammer_node_t onode, hammer_node_t nnode);
1145 void    hammer_cursor_removed_node(hammer_node_t onode, hammer_node_t parent,
1146                         int index);
1147 void    hammer_cursor_split_node(hammer_node_t onode, hammer_node_t nnode,
1148                         int index);
1149 void    hammer_cursor_moved_element(hammer_node_t oparent, int pindex,
1150                         hammer_node_t onode, int oindex,
1151                         hammer_node_t nnode, int nindex);
1152 void    hammer_cursor_parent_changed(hammer_node_t node, hammer_node_t oparent,
1153                         hammer_node_t nparent, int nindex);
1154 void    hammer_cursor_inserted_element(hammer_node_t node, int index);
1155 void    hammer_cursor_deleted_element(hammer_node_t node, int index);
1156 void    hammer_cursor_invalidate_cache(hammer_cursor_t cursor);
1157
1158 int     hammer_btree_lookup(hammer_cursor_t cursor);
1159 int     hammer_btree_first(hammer_cursor_t cursor);
1160 int     hammer_btree_last(hammer_cursor_t cursor);
1161 int     hammer_btree_extract(hammer_cursor_t cursor, int flags);
1162 int     hammer_btree_iterate(hammer_cursor_t cursor);
1163 int     hammer_btree_iterate_reverse(hammer_cursor_t cursor);
1164 int     hammer_btree_insert(hammer_cursor_t cursor,
1165                             hammer_btree_leaf_elm_t elm, int *doprop);
1166 int     hammer_btree_delete(hammer_cursor_t cursor, int *ndelete);
1167 void    hammer_btree_do_propagation(hammer_cursor_t cursor,
1168                             hammer_btree_leaf_elm_t leaf);
1169 int     hammer_btree_cmp(hammer_base_elm_t key1, hammer_base_elm_t key2);
1170 int     hammer_btree_chkts(hammer_tid_t ts, hammer_base_elm_t key);
1171 int     hammer_btree_correct_rhb(hammer_cursor_t cursor, hammer_tid_t tid);
1172 int     hammer_btree_correct_lhb(hammer_cursor_t cursor, hammer_tid_t tid);
1173
1174 int     btree_set_parent_of_child(hammer_transaction_t trans,
1175                         hammer_node_t node,
1176                         hammer_btree_elm_t elm);
1177 void    hammer_node_lock_init(hammer_node_lock_t parent, hammer_node_t node);
1178 void    hammer_btree_lcache_init(hammer_mount_t hmp, hammer_node_lock_t lcache,
1179                         int depth);
1180 void    hammer_btree_lcache_free(hammer_mount_t hmp, hammer_node_lock_t lcache);
1181 int     hammer_btree_lock_children(hammer_cursor_t cursor, int depth,
1182                         hammer_node_lock_t parent,
1183                         hammer_node_lock_t lcache);
1184 void    hammer_btree_lock_copy(hammer_cursor_t cursor,
1185                         hammer_node_lock_t parent);
1186 int     hammer_btree_sync_copy(hammer_cursor_t cursor,
1187                         hammer_node_lock_t parent);
1188 void    hammer_btree_unlock_children(hammer_mount_t hmp,
1189                         hammer_node_lock_t parent,
1190                         hammer_node_lock_t lcache);
1191 int     hammer_btree_search_node(hammer_base_elm_t elm, hammer_node_ondisk_t node);
1192 hammer_node_t hammer_btree_get_parent(hammer_transaction_t trans,
1193                         hammer_node_t node, int *parent_indexp,
1194                         int *errorp, int try_exclusive);
1195
1196 void    hammer_print_btree_node(hammer_node_ondisk_t ondisk);
1197 void    hammer_print_btree_elm(hammer_btree_elm_t elm);
1198
1199 void    *hammer_bread(hammer_mount_t hmp, hammer_off_t off,
1200                         int *errorp, hammer_buffer_t *bufferp);
1201 void    *hammer_bnew(hammer_mount_t hmp, hammer_off_t off,
1202                         int *errorp, hammer_buffer_t *bufferp);
1203 void    *hammer_bread_ext(hammer_mount_t hmp, hammer_off_t off, int bytes,
1204                         int *errorp, hammer_buffer_t *bufferp);
1205 void    *hammer_bnew_ext(hammer_mount_t hmp, hammer_off_t off, int bytes,
1206                         int *errorp, hammer_buffer_t *bufferp);
1207
1208 hammer_volume_t hammer_get_root_volume(hammer_mount_t hmp, int *errorp);
1209
1210 hammer_volume_t hammer_get_volume(hammer_mount_t hmp,
1211                         int32_t vol_no, int *errorp);
1212 hammer_buffer_t hammer_get_buffer(hammer_mount_t hmp, hammer_off_t buf_offset,
1213                         int bytes, int isnew, int *errorp);
1214 void            hammer_sync_buffers(hammer_mount_t hmp,
1215                         hammer_off_t base_offset, int bytes);
1216 int             hammer_del_buffers(hammer_mount_t hmp,
1217                         hammer_off_t base_offset,
1218                         hammer_off_t zone2_offset, int bytes,
1219                         int report_conflicts);
1220
1221 int             hammer_ref_volume(hammer_volume_t volume);
1222 int             hammer_ref_buffer(hammer_buffer_t buffer);
1223 void            hammer_flush_buffer_nodes(hammer_buffer_t buffer);
1224
1225 void            hammer_rel_volume(hammer_volume_t volume, int locked);
1226 void            hammer_rel_buffer(hammer_buffer_t buffer, int locked);
1227
1228 int             hammer_vfs_export(struct mount *mp, int op,
1229                         const struct export_args *export);
1230 hammer_node_t   hammer_get_node(hammer_transaction_t trans,
1231                         hammer_off_t node_offset, int isnew, int *errorp);
1232 void            hammer_ref_node(hammer_node_t node);
1233 hammer_node_t   hammer_ref_node_safe(hammer_transaction_t trans,
1234                         hammer_node_cache_t cache, int *errorp);
1235 void            hammer_rel_node(hammer_node_t node);
1236 void            hammer_delete_node(hammer_transaction_t trans,
1237                         hammer_node_t node);
1238 void            hammer_cache_node(hammer_node_cache_t cache,
1239                         hammer_node_t node);
1240 void            hammer_uncache_node(hammer_node_cache_t cache);
1241 void            hammer_flush_node(hammer_node_t node, int locked);
1242
1243 hammer_node_t hammer_alloc_btree(hammer_transaction_t trans,
1244                         hammer_off_t hint, int *errorp);
1245 void *hammer_alloc_data(hammer_transaction_t trans, int32_t data_len,
1246                         uint16_t rec_type, hammer_off_t *data_offsetp,
1247                         hammer_buffer_t *data_bufferp,
1248                         hammer_off_t hint, int *errorp);
1249
1250 int hammer_generate_undo(hammer_transaction_t trans,
1251                         hammer_off_t zone_offset, void *base, int len);
1252 int hammer_generate_redo(hammer_transaction_t trans, hammer_inode_t ip,
1253                         hammer_off_t file_offset, uint32_t flags,
1254                         void *base, int len);
1255 void hammer_generate_redo_sync(hammer_transaction_t trans);
1256 void hammer_redo_fifo_start_flush(hammer_inode_t ip);
1257 void hammer_redo_fifo_end_flush(hammer_inode_t ip);
1258
1259 void hammer_format_undo(void *base, uint32_t seqno);
1260 int hammer_upgrade_undo_4(hammer_transaction_t trans);
1261
1262 hammer_off_t hammer_freemap_alloc(hammer_transaction_t trans,
1263                         hammer_off_t owner, int *errorp);
1264 void hammer_freemap_free(hammer_transaction_t trans, hammer_off_t phys_offset,
1265                         hammer_off_t owner, int *errorp);
1266 int _hammer_checkspace(hammer_mount_t hmp, int slop, int64_t *resp);
1267 hammer_off_t hammer_blockmap_alloc(hammer_transaction_t trans, int zone,
1268                         int bytes, hammer_off_t hint, int *errorp);
1269 hammer_reserve_t hammer_blockmap_reserve(hammer_mount_t hmp, int zone,
1270                         int bytes, hammer_off_t *zone_offp, int *errorp);
1271 hammer_reserve_t hammer_blockmap_reserve_dedup(hammer_mount_t hmp, int zone,
1272                         int bytes, hammer_off_t zone_offset, int *errorp);
1273 void hammer_blockmap_reserve_complete(hammer_mount_t hmp,
1274                         hammer_reserve_t resv);
1275 void hammer_reserve_clrdelay(hammer_mount_t hmp, hammer_reserve_t resv);
1276 void hammer_blockmap_free(hammer_transaction_t trans,
1277                         hammer_off_t zone_offset, int bytes);
1278 int hammer_blockmap_dedup(hammer_transaction_t trans,
1279                         hammer_off_t zone_offset, int bytes);
1280 int hammer_blockmap_finalize(hammer_transaction_t trans,
1281                         hammer_reserve_t resv,
1282                         hammer_off_t zone_offset, int bytes);
1283 int hammer_blockmap_getfree(hammer_mount_t hmp, hammer_off_t zone_offset,
1284                         int *curp, int *errorp);
1285 hammer_off_t hammer_blockmap_lookup_verify(hammer_mount_t hmp,
1286                         hammer_off_t zone_offset, int *errorp);
1287
1288 hammer_off_t hammer_undo_lookup(hammer_mount_t hmp, hammer_off_t zone_offset,
1289                         int *errorp);
1290 int64_t hammer_undo_used(hammer_transaction_t trans);
1291 int64_t hammer_undo_space(hammer_transaction_t trans);
1292 int64_t hammer_undo_max(hammer_mount_t hmp);
1293 int hammer_undo_reclaim(hammer_io_t io);
1294
1295 void hammer_start_transaction(hammer_transaction_t trans,
1296                               hammer_mount_t hmp);
1297 void hammer_simple_transaction(hammer_transaction_t trans,
1298                               hammer_mount_t hmp);
1299 void hammer_start_transaction_fls(hammer_transaction_t trans,
1300                                   hammer_mount_t hmp);
1301 void hammer_done_transaction(hammer_transaction_t trans);
1302 hammer_tid_t hammer_alloc_tid(hammer_mount_t hmp, int count);
1303
1304 void hammer_modify_inode(hammer_transaction_t trans, hammer_inode_t ip, int flags);
1305 void hammer_flush_inode(hammer_inode_t ip, int flags);
1306 void hammer_flush_inode_done(hammer_inode_t ip, int error);
1307 void hammer_wait_inode(hammer_inode_t ip);
1308
1309 int  hammer_create_inode(hammer_transaction_t trans, struct vattr *vap,
1310                         struct ucred *cred, hammer_inode_t dip,
1311                         const char *name, int namelen,
1312                         hammer_pseudofs_inmem_t pfsm,
1313                         hammer_inode_t *ipp);
1314 void hammer_rel_inode(hammer_inode_t ip, int flush);
1315 int hammer_reload_inode(hammer_inode_t ip, void *arg __unused);
1316 int hammer_ino_rb_compare(hammer_inode_t ip1, hammer_inode_t ip2);
1317 int hammer_redo_rb_compare(hammer_inode_t ip1, hammer_inode_t ip2);
1318 int hammer_destroy_inode_callback(hammer_inode_t ip, void *data __unused);
1319
1320 int hammer_sync_inode(hammer_transaction_t trans, hammer_inode_t ip);
1321 void hammer_test_inode(hammer_inode_t dip);
1322 void hammer_inode_unloadable_check(hammer_inode_t ip, int getvp);
1323 int hammer_update_atime_quick(hammer_inode_t ip);
1324
1325 int  hammer_ip_add_direntry(hammer_transaction_t trans,
1326                         hammer_inode_t dip, const char *name, int bytes,
1327                         hammer_inode_t nip);
1328 int  hammer_ip_del_direntry(hammer_transaction_t trans,
1329                         hammer_cursor_t cursor, hammer_inode_t dip,
1330                         hammer_inode_t ip);
1331 void hammer_ip_replace_bulk(hammer_mount_t hmp, hammer_record_t record);
1332 hammer_record_t hammer_ip_add_bulk(hammer_inode_t ip, off_t file_offset,
1333                         void *data, int bytes, int *errorp);
1334 int  hammer_ip_frontend_trunc(hammer_inode_t ip, off_t file_size);
1335 int  hammer_ip_add_record(hammer_transaction_t trans,
1336                         hammer_record_t record);
1337 int  hammer_ip_delete_range(hammer_cursor_t cursor, hammer_inode_t ip,
1338                         int64_t ran_beg, int64_t ran_end, int truncating);
1339 int  hammer_ip_delete_clean(hammer_cursor_t cursor, hammer_inode_t ip,
1340                         int *countp);
1341 int  hammer_ip_sync_data(hammer_cursor_t cursor, hammer_inode_t ip,
1342                         int64_t offset, void *data, int bytes);
1343 int  hammer_ip_sync_record_cursor(hammer_cursor_t cursor, hammer_record_t rec);
1344 hammer_pseudofs_inmem_t  hammer_load_pseudofs(hammer_transaction_t trans,
1345                         uint32_t localization, int *errorp);
1346 int  hammer_mkroot_pseudofs(hammer_transaction_t trans, struct ucred *cred,
1347                         hammer_pseudofs_inmem_t pfsm, hammer_inode_t dip);
1348 int  hammer_save_pseudofs(hammer_transaction_t trans,
1349                         hammer_pseudofs_inmem_t pfsm);
1350 int  hammer_unload_pseudofs(hammer_transaction_t trans, uint32_t localization);
1351 void hammer_rel_pseudofs(hammer_mount_t hmp, hammer_pseudofs_inmem_t pfsm);
1352 int hammer_ioctl(hammer_inode_t ip, u_long com, caddr_t data, int fflag,
1353                         struct ucred *cred);
1354
1355 void hammer_io_init(hammer_io_t io, hammer_volume_t volume,
1356                         hammer_io_type_t type);
1357 hammer_io_type_t hammer_zone_to_iotype(int zone);
1358 int hammer_io_read(struct vnode *devvp, hammer_io_t io, int limit);
1359 void hammer_io_advance(hammer_io_t io);
1360 int hammer_io_new(struct vnode *devvp, hammer_io_t io);
1361 int hammer_io_inval(hammer_volume_t volume, hammer_off_t zone2_offset);
1362 struct buf *hammer_io_release(hammer_io_t io, int flush);
1363 void hammer_io_flush(hammer_io_t io, int reclaim);
1364 void hammer_io_wait(hammer_io_t io);
1365 void hammer_io_waitdep(hammer_io_t io);
1366 void hammer_io_wait_all(hammer_mount_t hmp, const char *ident, int doflush);
1367 int hammer_io_direct_read(hammer_mount_t hmp, struct bio *bio,
1368                         hammer_btree_leaf_elm_t leaf);
1369 int hammer_io_indirect_read(hammer_mount_t hmp, struct bio *bio,
1370                         hammer_btree_leaf_elm_t leaf);
1371 int hammer_io_direct_write(hammer_mount_t hmp, struct bio *bio,
1372                         hammer_record_t record);
1373 void hammer_io_direct_wait(hammer_record_t record);
1374 void hammer_io_direct_uncache(hammer_mount_t hmp, hammer_btree_leaf_elm_t leaf);
1375 void hammer_io_write_interlock(hammer_io_t io);
1376 void hammer_io_done_interlock(hammer_io_t io);
1377 void hammer_io_clear_modify(hammer_io_t io, int inval);
1378 void hammer_io_clear_modlist(hammer_io_t io);
1379 void hammer_io_flush_sync(hammer_mount_t hmp);
1380 void hammer_io_clear_error(hammer_io_t io);
1381 void hammer_io_clear_error_noassert(hammer_io_t io);
1382 void hammer_io_notmeta(hammer_buffer_t buffer);
1383 void hammer_io_limit_backlog(hammer_mount_t hmp);
1384
1385 void hammer_modify_volume(hammer_transaction_t trans, hammer_volume_t volume,
1386                         void *base, int len);
1387 void hammer_modify_buffer(hammer_transaction_t trans, hammer_buffer_t buffer,
1388                         void *base, int len);
1389 void hammer_modify_volume_done(hammer_volume_t volume);
1390 void hammer_modify_buffer_done(hammer_buffer_t buffer);
1391
1392 int hammer_ioc_reblock(hammer_transaction_t trans, hammer_inode_t ip,
1393                         struct hammer_ioc_reblock *reblock);
1394 int hammer_ioc_rebalance(hammer_transaction_t trans, hammer_inode_t ip,
1395                         struct hammer_ioc_rebalance *rebal);
1396 int hammer_ioc_prune(hammer_transaction_t trans, hammer_inode_t ip,
1397                         struct hammer_ioc_prune *prune);
1398 int hammer_ioc_mirror_read(hammer_transaction_t trans, hammer_inode_t ip,
1399                         struct hammer_ioc_mirror_rw *mirror);
1400 int hammer_ioc_mirror_write(hammer_transaction_t trans, hammer_inode_t ip,
1401                         struct hammer_ioc_mirror_rw *mirror);
1402 int hammer_ioc_set_pseudofs(hammer_transaction_t trans, hammer_inode_t ip,
1403                         struct ucred *cred, struct hammer_ioc_pseudofs_rw *pfs);
1404 int hammer_ioc_get_pseudofs(hammer_transaction_t trans, hammer_inode_t ip,
1405                         struct hammer_ioc_pseudofs_rw *pfs);
1406 int hammer_ioc_destroy_pseudofs(hammer_transaction_t trans, hammer_inode_t ip,
1407                         struct hammer_ioc_pseudofs_rw *pfs);
1408 int hammer_ioc_downgrade_pseudofs(hammer_transaction_t trans, hammer_inode_t ip,
1409                         struct hammer_ioc_pseudofs_rw *pfs);
1410 int hammer_ioc_upgrade_pseudofs(hammer_transaction_t trans, hammer_inode_t ip,
1411                         struct hammer_ioc_pseudofs_rw *pfs);
1412 int hammer_ioc_wait_pseudofs(hammer_transaction_t trans, hammer_inode_t ip,
1413                         struct hammer_ioc_pseudofs_rw *pfs);
1414 int hammer_ioc_scan_pseudofs(hammer_transaction_t trans, hammer_inode_t ip,
1415                         struct hammer_ioc_pseudofs_rw *pfs);
1416 int hammer_ioc_volume_add(hammer_transaction_t trans, hammer_inode_t ip,
1417                         struct hammer_ioc_volume *ioc);
1418 int hammer_ioc_volume_del(hammer_transaction_t trans, hammer_inode_t ip,
1419                         struct hammer_ioc_volume *ioc);
1420 int hammer_ioc_volume_list(hammer_transaction_t trans, hammer_inode_t ip,
1421                         struct hammer_ioc_volume_list *ioc);
1422 int hammer_ioc_dedup(hammer_transaction_t trans, hammer_inode_t ip,
1423                         struct hammer_ioc_dedup *dedup);
1424
1425 int hammer_signal_check(hammer_mount_t hmp);
1426
1427 void hammer_flusher_create(hammer_mount_t hmp);
1428 void hammer_flusher_destroy(hammer_mount_t hmp);
1429 void hammer_flusher_sync(hammer_mount_t hmp);
1430 int  hammer_flusher_async(hammer_mount_t hmp, hammer_flush_group_t flg);
1431 int  hammer_flusher_async_one(hammer_mount_t hmp);
1432 int hammer_flusher_running(hammer_mount_t hmp);
1433 void hammer_flusher_wait(hammer_mount_t hmp, int seq);
1434 void hammer_flusher_wait_next(hammer_mount_t hmp);
1435 int  hammer_flusher_meta_limit(hammer_mount_t hmp);
1436 int  hammer_flusher_meta_halflimit(hammer_mount_t hmp);
1437 int  hammer_flusher_undo_exhausted(hammer_transaction_t trans, int quarter);
1438 void hammer_flusher_clean_loose_ios(hammer_mount_t hmp);
1439 void hammer_flusher_finalize(hammer_transaction_t trans, int final);
1440 int  hammer_flusher_haswork(hammer_mount_t hmp);
1441 int  hammer_flush_dirty(hammer_mount_t hmp, int max_count);
1442 void hammer_flusher_flush_undos(hammer_mount_t hmp, int already_flushed);
1443
1444 int hammer_recover_stage1(hammer_mount_t hmp, hammer_volume_t rootvol);
1445 int hammer_recover_stage2(hammer_mount_t hmp, hammer_volume_t rootvol);
1446 void hammer_recover_flush_buffers(hammer_mount_t hmp,
1447                         hammer_volume_t root_volume, int final);
1448
1449 udev_t hammer_fsid_to_udev(uuid_t *uuid);
1450
1451
1452 int hammer_blocksize(int64_t file_offset);
1453 int hammer_blockoff(int64_t file_offset);
1454 int64_t hammer_blockdemarc(int64_t file_offset1, int64_t file_offset2);
1455
1456 /*
1457  * Shortcut for _hammer_checkspace(), used all over the code.
1458  */
1459 static __inline int
1460 hammer_checkspace(hammer_mount_t hmp, int slop)
1461 {
1462         return(_hammer_checkspace(hmp, slop, NULL));
1463 }
1464
1465 static __inline void
1466 hammer_wait_mem_record(hammer_record_t record)
1467 {
1468         hammer_wait_mem_record_ident(record, "hmmwai");
1469 }
1470
1471 static __inline void
1472 hammer_lock_ex(struct hammer_lock *lock)
1473 {
1474         hammer_lock_ex_ident(lock, "hmrlck");
1475 }
1476
1477 static __inline void
1478 hammer_modify_volume_noundo(hammer_transaction_t trans, hammer_volume_t volume)
1479 {
1480         hammer_modify_volume(trans, volume, NULL, 0);
1481 }
1482
1483 static __inline void
1484 hammer_modify_buffer_noundo(hammer_transaction_t trans, hammer_buffer_t buffer)
1485 {
1486         hammer_modify_buffer(trans, buffer, NULL, 0);
1487 }
1488
1489 /*
1490  * Indicate that a B-Tree node is being modified.
1491  */
1492 static __inline void
1493 hammer_modify_node_noundo(hammer_transaction_t trans, hammer_node_t node)
1494 {
1495         KKASSERT((node->flags & HAMMER_NODE_CRCBAD) == 0);
1496         hammer_modify_buffer(trans, node->buffer, NULL, 0);
1497 }
1498
1499 static __inline void
1500 hammer_modify_node_all(hammer_transaction_t trans, hammer_node_t node)
1501 {
1502         KKASSERT((node->flags & HAMMER_NODE_CRCBAD) == 0);
1503         hammer_modify_buffer(trans, node->buffer,
1504                              node->ondisk, sizeof(*node->ondisk));
1505 }
1506
1507 static __inline void
1508 hammer_modify_node(hammer_transaction_t trans, hammer_node_t node,
1509                    void *base, int len)
1510 {
1511         KKASSERT((char *)base >= (char *)node->ondisk &&
1512                  (char *)base + len <=
1513                     (char *)node->ondisk + sizeof(*node->ondisk));
1514         KKASSERT((node->flags & HAMMER_NODE_CRCBAD) == 0);
1515
1516         if (hammer_btree_full_undo) {
1517                 hammer_modify_node_all(trans, node);
1518         } else {
1519                 hammer_modify_buffer(trans, node->buffer, base, len);
1520                 hammer_modify_buffer(trans, node->buffer, &node->ondisk->crc,
1521                                      sizeof(hammer_crc_t));
1522                 --node->buffer->io.modify_refs; /* only want one ref */
1523         }
1524 }
1525
1526 /*
1527  * Indicate that the specified modifications have been completed.
1528  *
1529  * Do not try to generate the crc here, it's very expensive to do and a
1530  * sequence of insertions or deletions can result in many calls to this
1531  * function on the same node.
1532  */
1533 static __inline void
1534 hammer_modify_node_done(hammer_node_t node)
1535 {
1536         node->flags |= HAMMER_NODE_CRCGOOD;
1537         if ((node->flags & HAMMER_NODE_NEEDSCRC) == 0) {
1538                 node->flags |= HAMMER_NODE_NEEDSCRC;
1539                 node->buffer->io.gencrc = 1;
1540                 hammer_ref_node(node);
1541         }
1542         hammer_modify_buffer_done(node->buffer);
1543 }
1544
1545 static __inline int
1546 hammer_btree_extract_leaf(hammer_cursor_t cursor)
1547 {
1548         return(hammer_btree_extract(cursor, 0));
1549 }
1550
1551 static __inline int
1552 hammer_btree_extract_data(hammer_cursor_t cursor)
1553 {
1554         return(hammer_btree_extract(cursor, HAMMER_CURSOR_GET_DATA));
1555 }
1556
1557 /*
1558  * Lookup a blockmap offset.
1559  */
1560 static __inline hammer_off_t
1561 hammer_blockmap_lookup(hammer_mount_t hmp, hammer_off_t zone_offset,
1562                         int *errorp)
1563 {
1564 #if defined INVARIANTS
1565         KKASSERT(hammer_is_zone_record(zone_offset));
1566 #endif
1567
1568         /*
1569          * We can actually skip blockmap verify by default,
1570          * as normal blockmaps are now direct-mapped onto the freemap
1571          * and so represent zone-2 addresses.
1572          */
1573         if (hammer_verify_zone == 0) {
1574                 *errorp = 0;
1575                 return hammer_xlate_to_zone2(zone_offset);
1576         }
1577
1578         return hammer_blockmap_lookup_verify(hmp, zone_offset, errorp);
1579 }
1580
1581 #define hammer_modify_volume_field(trans, vol, field)           \
1582         hammer_modify_volume(trans, vol, &(vol)->ondisk->field, \
1583                              sizeof((vol)->ondisk->field))
1584
1585 #define hammer_modify_node_field(trans, node, field)            \
1586         hammer_modify_node(trans, node, &(node)->ondisk->field, \
1587                              sizeof((node)->ondisk->field))
1588
1589 /*
1590  * The HAMMER_INODE_CAP_DIR_LOCAL_INO capability is set on newly
1591  * created directories for HAMMER version 2 or greater and causes
1592  * directory entries to be placed the inode localization zone in
1593  * the B-Tree instead of the misc zone.
1594  *
1595  * This greatly improves localization between directory entries and
1596  * inodes
1597  */
1598 static __inline uint32_t
1599 hammer_dir_localization(hammer_inode_t dip)
1600 {
1601         return(HAMMER_DIR_INODE_LOCALIZATION(&dip->ino_data));
1602 }
1603
1604 static __inline
1605 hammer_io_t
1606 hammer_buf_peek_io(struct buf *bp)
1607 {
1608         return((hammer_io_t)bp->b_priv);
1609 }
1610
1611 static __inline
1612 void
1613 hammer_buf_attach_io(struct buf *bp, hammer_io_t io)
1614 {
1615         /* struct buf and struct hammer_io are 1:1 */
1616         KKASSERT(hammer_buf_peek_io(bp) == NULL);
1617         bp->b_priv = io;
1618 }
1619
1620 static __inline int
1621 __hammer_vol_index(int vol_no)
1622 {
1623         return(vol_no >> 6);
1624 }
1625
1626 static __inline uint64_t
1627 __hammer_vol_low(int vol_no)
1628 {
1629         return((uint64_t)1 << (vol_no & ((1 << 6) - 1)));
1630 }
1631
1632 static __inline void
1633 hammer_volume_number_add(hammer_mount_t hmp, hammer_volume_t vol)
1634 {
1635         int i = __hammer_vol_index(vol->vol_no);
1636         hmp->volume_map[i] |= __hammer_vol_low(vol->vol_no);
1637 }
1638
1639 static __inline void
1640 hammer_volume_number_del(hammer_mount_t hmp, hammer_volume_t vol)
1641 {
1642         int i = __hammer_vol_index(vol->vol_no);
1643         hmp->volume_map[i] &= ~__hammer_vol_low(vol->vol_no);
1644 }
1645
1646 static __inline int
1647 hammer_volume_number_test(hammer_mount_t hmp, int n)
1648 {
1649         int i = __hammer_vol_index(n);
1650         return((hmp->volume_map[i] & __hammer_vol_low(n)) != 0);
1651 }
1652
1653 #define hkprintf(format, args...)                       \
1654         kprintf("HAMMER: "format,## args)
1655 #define hvkprintf(vol, format, args...)                 \
1656         kprintf("HAMMER(%s) "format, vol->ondisk->vol_label,## args)
1657 #define hmkprintf(hmp, format, args...)                 \
1658         kprintf("HAMMER(%s) "format, hmp->mp->mnt_stat.f_mntfromname,## args)
1659 #define hdkprintf(format, args...)                      \
1660         kprintf("%s: "format, __func__,## args)
1661
1662 #define hkrateprintf(rate , format, args...)            \
1663         krateprintf(rate, "HAMMER: "format,## args)
1664 #define hvkrateprintf(rate, vol, format, args...)       \
1665         krateprintf(rate, "HAMMER(%s) "format, vol->ondisk->vol_label,## args)
1666 #define hmkrateprintf(rate, hmp, format, args...)       \
1667         krateprintf(rate, "HAMMER(%s) "format, hmp->mp->mnt_stat.f_mntfromname,## args)
1668 #define hdkrateprintf(rate, format, args...)            \
1669         krateprintf(rate, "%s: "format, __func__,## args)
1670
1671 #define hpanic(format, args...)                         \
1672         panic("%s: "format, __func__,## args)
1673 #endif  /* _KERNEL */
1674
1675 #endif /* !VFS_HAMMER_HAMMER_H_ */