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  * $DragonFly: src/sys/vfs/hammer/hammer.h,v 1.114 2008/07/12 23:04:50 dillon Exp $
  35  */
  36 /*
  37  * This header file contains structures used internally by the HAMMERFS
  38  * implementation.  See hammer_disk.h for on-disk structures.
  39  */
  40
  41 #include <sys/param.h>
  42 #include <sys/types.h>
  43 #include <sys/kernel.h>
  44 #include <sys/conf.h>
  45 #include <sys/systm.h>
  46 #include <sys/tree.h>
  47 #include <sys/malloc.h>
  48 #include <sys/mount.h>
  49 #include <sys/mountctl.h>
  50 #include <sys/vnode.h>
  51 #include <sys/proc.h>
  52 #include <sys/stat.h>
  53 #include <sys/globaldata.h>
  54 #include <sys/lockf.h>
  55 #include <sys/buf.h>
  56 #include <sys/queue.h>
  57 #include <sys/ktr.h>
  58 #include <sys/globaldata.h>
  59
  60 #include <sys/buf2.h>
  61 #include <sys/signal2.h>
  62 #include "hammer_disk.h"
  63 #include "hammer_mount.h"
  64 #include "hammer_ioctl.h"
  65
  66 #if defined(_KERNEL) || defined(_KERNEL_STRUCTURES)
  67
  68 MALLOC_DECLARE(M_HAMMER);
  69
  70 /*
  71  * Kernel trace
  72  */
  73 #if !defined(KTR_HAMMER)
  74 #define KTR_HAMMER      KTR_ALL
  75 #endif
  76 KTR_INFO_MASTER_EXTERN(hammer);
  77
  78 /*
  79  * Misc structures
  80  */
  81 struct hammer_mount;
  82
  83 /*
  84  * Key structure used for custom RB tree inode lookups.  This prototypes
  85  * the function hammer_ino_rb_tree_RB_LOOKUP_INFO(root, info).
  86  */
  87 typedef struct hammer_inode_info {
  88         int64_t         obj_id;         /* (key) object identifier */
  89         hammer_tid_t    obj_asof;       /* (key) snapshot transid or 0 */
  90         u_int32_t       obj_localization; /* (key) pseudo-fs */
  91         union {
  92                 struct hammer_btree_leaf_elm *leaf;
  93         } u;
  94 } *hammer_inode_info_t;
  95
  96 typedef enum hammer_transaction_type {
  97         HAMMER_TRANS_RO,
  98         HAMMER_TRANS_STD,
  99         HAMMER_TRANS_FLS
 100 } hammer_transaction_type_t;
 101
 102 /*
 103  * HAMMER Transaction tracking
 104  */
 105 struct hammer_transaction {
 106         hammer_transaction_type_t type;
 107         struct hammer_mount *hmp;
 108         hammer_tid_t    tid;
 109         u_int64_t       time;
 110         u_int32_t       time32;
 111         int             sync_lock_refs;
 112         struct hammer_volume *rootvol;
 113 };
 114
 115 typedef struct hammer_transaction *hammer_transaction_t;
 116
 117 /*
 118  * HAMMER locks
 119  */
 120 struct hammer_lock {
 121         int     refs;           /* active references delay writes */
 122         int     lockcount;      /* lock count for exclusive/shared access */
 123         int     wanted;
 124         int     exwanted;       /* number of threads waiting for ex lock */
 125         struct thread *locktd;
 126 };
 127
 128 static __inline int
 129 hammer_islocked(struct hammer_lock *lock)
 130 {
 131         return(lock->lockcount != 0);
 132 }
 133
 134 static __inline int
 135 hammer_isactive(struct hammer_lock *lock)
 136 {
 137         return(lock->refs != 0);
 138 }
 139
 140 static __inline int
 141 hammer_islastref(struct hammer_lock *lock)
 142 {
 143         return(lock->refs == 1);
 144 }
 145
 146 /*
 147  * Return if we specifically own the lock exclusively.
 148  */
 149 static __inline int
 150 hammer_lock_excl_owned(struct hammer_lock *lock, thread_t td)
 151 {
 152         if (lock->lockcount > 0 && lock->locktd == td)
 153                 return(1);
 154         return(0);
 155 }
 156
 157 /*
 158  * Flush state, used by various structures
 159  */
 160 typedef enum hammer_inode_state {
 161         HAMMER_FST_IDLE,
 162         HAMMER_FST_SETUP,
 163         HAMMER_FST_FLUSH
 164 } hammer_inode_state_t;
 165
 166 TAILQ_HEAD(hammer_record_list, hammer_record);
 167
 168 /*
 169  * Pseudo-filesystem extended data tracking
 170  */
 171 struct hammer_pfs_rb_tree;
 172 struct hammer_pseudofs_inmem;
 173 RB_HEAD(hammer_pfs_rb_tree, hammer_pseudofs_inmem);
 174 RB_PROTOTYPE2(hammer_pfs_rb_tree, hammer_pseudofs_inmem, rb_node,
 175               hammer_pfs_rb_compare, u_int32_t);
 176
 177 struct hammer_pseudofs_inmem {
 178         RB_ENTRY(hammer_pseudofs_inmem) rb_node;
 179         struct hammer_lock      lock;
 180         u_int32_t               localization;
 181         hammer_tid_t            create_tid;
 182         int                     flags;
 183         udev_t                  fsid_udev;
 184         struct hammer_pseudofs_data pfsd;
 185 };
 186
 187 typedef struct hammer_pseudofs_inmem *hammer_pseudofs_inmem_t;
 188
 189 #define HAMMER_PFSM_DELETED     0x0001
 190
 191 /*
 192  * Cache object ids.  A fixed number of objid cache structures are
 193  * created to reserve object id's for newly created files in multiples
 194  * of 100,000, localized to a particular directory, and recycled as
 195  * needed.  This allows parallel create operations in different
 196  * directories to retain fairly localized object ids which in turn
 197  * improves reblocking performance and layout.
 198  */
 199 #define OBJID_CACHE_SIZE        1024
 200 #define OBJID_CACHE_BULK        100000
 201
 202 typedef struct hammer_objid_cache {
 203         TAILQ_ENTRY(hammer_objid_cache) entry;
 204         struct hammer_inode             *dip;
 205         hammer_tid_t                    next_tid;
 206         int                             count;
 207 } *hammer_objid_cache_t;
 208
 209 /*
 210  * Associate an inode with a B-Tree node to cache search start positions
 211  */
 212 typedef struct hammer_node_cache {
 213         TAILQ_ENTRY(hammer_node_cache)  entry;
 214         struct hammer_node              *node;
 215         struct hammer_inode             *ip;
 216 } *hammer_node_cache_t;
 217
 218 TAILQ_HEAD(hammer_node_cache_list, hammer_node_cache);
 219
 220 /*
 221  * Structure used to organize flush groups.  Flush groups must be
 222  * organized into chunks in order to avoid blowing out the UNDO FIFO.
 223  * Without this a 'sync' could end up flushing 50,000 inodes in a single
 224  * transaction.
 225  */
 226 struct hammer_flush_group {
 227         TAILQ_ENTRY(hammer_flush_group) flush_entry;
 228         TAILQ_HEAD(, hammer_inode)      flush_list;
 229         int                             inode_count;    /* inode load */
 230         int                             total_count;    /* record load */
 231         int                             running;        /* group is running */
 232         int                             closed;
 233         int                             refs;
 234 };
 235
 236 typedef struct hammer_flush_group *hammer_flush_group_t;
 237
 238 TAILQ_HEAD(hammer_flush_group_list, hammer_flush_group);
 239
 240 /*
 241  * Structure used to represent an inode in-memory.
 242  *
 243  * The record and data associated with an inode may be out of sync with
 244  * the disk (xDIRTY flags), or not even on the disk at all (ONDISK flag
 245  * clear).
 246  *
 247  * An inode may also hold a cache of unsynchronized records, used for
 248  * database and directories only.  Unsynchronized regular file data is
 249  * stored in the buffer cache.
 250  *
 251  * NOTE: A file which is created and destroyed within the initial
 252  * synchronization period can wind up not doing any disk I/O at all.
 253  *
 254  * Finally, an inode may cache numerous disk-referencing B-Tree cursors.
 255  */
 256 struct hammer_ino_rb_tree;
 257 struct hammer_inode;
 258 RB_HEAD(hammer_ino_rb_tree, hammer_inode);
 259 RB_PROTOTYPEX(hammer_ino_rb_tree, INFO, hammer_inode, rb_node,
 260               hammer_ino_rb_compare, hammer_inode_info_t);
 261
 262 struct hammer_rec_rb_tree;
 263 struct hammer_record;
 264 RB_HEAD(hammer_rec_rb_tree, hammer_record);
 265 RB_PROTOTYPEX(hammer_rec_rb_tree, INFO, hammer_record, rb_node,
 266               hammer_rec_rb_compare, hammer_btree_leaf_elm_t);
 267
 268 TAILQ_HEAD(hammer_node_list, hammer_node);
 269
 270 struct hammer_inode {
 271         RB_ENTRY(hammer_inode)  rb_node;
 272         hammer_inode_state_t    flush_state;
 273         hammer_flush_group_t    flush_group;
 274         TAILQ_ENTRY(hammer_inode) flush_entry;
 275         struct hammer_record_list target_list;  /* target of dependant recs */
 276         int64_t                 obj_id;         /* (key) object identifier */
 277         hammer_tid_t            obj_asof;       /* (key) snapshot or 0 */
 278         u_int32_t               obj_localization; /* (key) pseudo-fs */
 279         struct hammer_mount     *hmp;
 280         hammer_objid_cache_t    objid_cache;
 281         int                     flags;
 282         int                     error;          /* flush error */
 283         int                     cursor_ip_refs; /* sanity */
 284         int                     rsv_recs;
 285         struct vnode            *vp;
 286         hammer_pseudofs_inmem_t pfsm;
 287         struct lockf            advlock;
 288         struct hammer_lock      lock;           /* sync copy interlock */
 289         off_t                   trunc_off;
 290         struct hammer_btree_leaf_elm ino_leaf;  /* in-memory cache */
 291         struct hammer_inode_data ino_data;      /* in-memory cache */
 292         struct hammer_rec_rb_tree rec_tree;     /* in-memory cache */
 293         struct hammer_node_cache cache[2];      /* search initiate cache */
 294
 295         /*
 296          * When a demark is created to synchronize an inode to
 297          * disk, certain fields are copied so the front-end VOPs
 298          * can continue to run in parallel with the synchronization
 299          * occuring in the background.
 300          */
 301         int             sync_flags;             /* to-sync flags cache */
 302         off_t           sync_trunc_off;         /* to-sync truncation */
 303         off_t           save_trunc_off;         /* write optimization */
 304         struct hammer_btree_leaf_elm sync_ino_leaf; /* to-sync cache */
 305         struct hammer_inode_data sync_ino_data; /* to-sync cache */
 306 };
 307
 308 typedef struct hammer_inode *hammer_inode_t;
 309
 310 #define VTOI(vp)        ((struct hammer_inode *)(vp)->v_data)
 311
 312 #define HAMMER_INODE_DDIRTY     0x0001  /* in-memory ino_data is dirty */
 313                                         /* (not including atime/mtime) */
 314 #define HAMMER_INODE_RSV_INODES 0x0002  /* hmp->rsv_inodes bumped */
 315 #define HAMMER_INODE_UNUSED0004 0x0004
 316 #define HAMMER_INODE_XDIRTY     0x0008  /* in-memory records */
 317 #define HAMMER_INODE_ONDISK     0x0010  /* inode is on-disk (else not yet) */
 318 #define HAMMER_INODE_FLUSH      0x0020  /* flush on last ref */
 319 #define HAMMER_INODE_DELETED    0x0080  /* inode delete (backend) */
 320 #define HAMMER_INODE_DELONDISK  0x0100  /* delete synchronized to disk */
 321 #define HAMMER_INODE_RO         0x0200  /* read-only (because of as-of) */
 322 #define HAMMER_INODE_VHELD      0x0400  /* vnode held on sync */
 323 #define HAMMER_INODE_DONDISK    0x0800  /* data records may be on disk */
 324 #define HAMMER_INODE_BUFS       0x1000  /* dirty high level bps present */
 325 #define HAMMER_INODE_REFLUSH    0x2000  /* pipelined flush during flush */
 326 #define HAMMER_INODE_RECLAIM    0x4000  /* trying to reclaim */
 327 #define HAMMER_INODE_FLUSHW     0x8000  /* Someone waiting for flush */
 328
 329 #define HAMMER_INODE_TRUNCATED  0x00010000
 330 #define HAMMER_INODE_DELETING   0x00020000 /* inode delete request (frontend)*/
 331 #define HAMMER_INODE_RESIGNAL   0x00040000 /* re-signal on re-flush */
 332 #define HAMMER_INODE_ATIME      0x00100000 /* in-memory atime modified */
 333 #define HAMMER_INODE_MTIME      0x00200000 /* in-memory mtime modified */
 334 #define HAMMER_INODE_WOULDBLOCK 0x00400000 /* re-issue to new flush group */
 335
 336 #define HAMMER_INODE_MODMASK    (HAMMER_INODE_DDIRTY|                       \
 337                                  HAMMER_INODE_XDIRTY|HAMMER_INODE_BUFS|     \
 338                                  HAMMER_INODE_ATIME|HAMMER_INODE_MTIME|     \
 339                                  HAMMER_INODE_TRUNCATED|HAMMER_INODE_DELETING)
 340
 341 #define HAMMER_INODE_MODMASK_NOXDIRTY \
 342                                 (HAMMER_INODE_MODMASK & ~HAMMER_INODE_XDIRTY)
 343
 344 #define HAMMER_FLUSH_GROUP_SIZE 64
 345
 346 #define HAMMER_FLUSH_SIGNAL     0x0001
 347 #define HAMMER_FLUSH_RECURSION  0x0002
 348
 349 /*
 350  * Used by the inode reclaim code to pipeline reclaims and avoid
 351  * blowing out kernel memory or letting the flusher get too far
 352  * behind.
 353  */
 354 struct hammer_reclaim {
 355         TAILQ_ENTRY(hammer_reclaim) entry;
 356         int     okydoky;
 357 };
 358
 359 #define HAMMER_RECLAIM_FLUSH    2000
 360 #define HAMMER_RECLAIM_WAIT     4000
 361
 362 /*
 363  * Structure used to represent an unsynchronized record in-memory.  These
 364  * records typically represent directory entries.  Only non-historical
 365  * records are kept in-memory.
 366  *
 367  * Records are organized as a per-inode RB-Tree.  If the inode is not
 368  * on disk then neither are any records and the in-memory record tree
 369  * represents the entire contents of the inode.  If the inode is on disk
 370  * then the on-disk B-Tree is scanned in parallel with the in-memory
 371  * RB-Tree to synthesize the current state of the file.
 372  *
 373  * Records are also used to enforce the ordering of directory create/delete
 374  * operations.  A new inode will not be flushed to disk unless its related
 375  * directory entry is also being flushed at the same time.  A directory entry
 376  * will not be removed unless its related inode is also being removed at the
 377  * same time.
 378  */
 379 typedef enum hammer_record_type {
 380         HAMMER_MEM_RECORD_GENERAL,      /* misc record */
 381         HAMMER_MEM_RECORD_INODE,        /* inode record */
 382         HAMMER_MEM_RECORD_ADD,          /* positive memory cache record */
 383         HAMMER_MEM_RECORD_DEL,          /* negative delete-on-disk record */
 384         HAMMER_MEM_RECORD_DATA          /* bulk-data record w/on-disk ref */
 385 } hammer_record_type_t;
 386
 387 struct hammer_record {
 388         RB_ENTRY(hammer_record)         rb_node;
 389         TAILQ_ENTRY(hammer_record)      target_entry;
 390         hammer_inode_state_t            flush_state;
 391         hammer_flush_group_t            flush_group;
 392         hammer_record_type_t            type;
 393         struct hammer_lock              lock;
 394         struct hammer_reserve           *resv;
 395         struct hammer_inode             *ip;
 396         struct hammer_inode             *target_ip;
 397         struct hammer_btree_leaf_elm    leaf;
 398         union hammer_data_ondisk        *data;
 399         int                             flags;
 400 };
 401
 402 typedef struct hammer_record *hammer_record_t;
 403
 404 /*
 405  * Record flags.  Note that FE can only be set by the frontend if the
 406  * record has not been interlocked by the backend w/ BE.
 407  */
 408 #define HAMMER_RECF_ALLOCDATA           0x0001
 409 #define HAMMER_RECF_ONRBTREE            0x0002
 410 #define HAMMER_RECF_DELETED_FE          0x0004  /* deleted (frontend) */
 411 #define HAMMER_RECF_DELETED_BE          0x0008  /* deleted (backend) */
 412 #define HAMMER_RECF_UNUSED0010          0x0010
 413 #define HAMMER_RECF_INTERLOCK_BE        0x0020  /* backend interlock */
 414 #define HAMMER_RECF_WANTED              0x0040  /* wanted by the frontend */
 415 #define HAMMER_RECF_CONVERT_DELETE      0x0100 /* special case */
 416
 417 /*
 418  * hammer_delete_at_cursor() flags
 419  */
 420 #define HAMMER_DELETE_ADJUST            0x0001
 421 #define HAMMER_DELETE_DESTROY           0x0002
 422
 423 /*
 424  * In-memory structures representing on-disk structures.
 425  */
 426 struct hammer_volume;
 427 struct hammer_buffer;
 428 struct hammer_node;
 429 struct hammer_undo;
 430 struct hammer_reserve;
 431
 432 RB_HEAD(hammer_vol_rb_tree, hammer_volume);
 433 RB_HEAD(hammer_buf_rb_tree, hammer_buffer);
 434 RB_HEAD(hammer_nod_rb_tree, hammer_node);
 435 RB_HEAD(hammer_und_rb_tree, hammer_undo);
 436 RB_HEAD(hammer_res_rb_tree, hammer_reserve);
 437
 438 RB_PROTOTYPE2(hammer_vol_rb_tree, hammer_volume, rb_node,
 439               hammer_vol_rb_compare, int32_t);
 440 RB_PROTOTYPE2(hammer_buf_rb_tree, hammer_buffer, rb_node,
 441               hammer_buf_rb_compare, hammer_off_t);
 442 RB_PROTOTYPE2(hammer_nod_rb_tree, hammer_node, rb_node,
 443               hammer_nod_rb_compare, hammer_off_t);
 444 RB_PROTOTYPE2(hammer_und_rb_tree, hammer_undo, rb_node,
 445               hammer_und_rb_compare, hammer_off_t);
 446 RB_PROTOTYPE2(hammer_res_rb_tree, hammer_reserve, rb_node,
 447               hammer_res_rb_compare, hammer_off_t);
 448
 449 /*
 450  * IO management - embedded at the head of various in-memory structures
 451  *
 452  * VOLUME       - hammer_volume containing meta-data
 453  * META_BUFFER  - hammer_buffer containing meta-data
 454  * DATA_BUFFER  - hammer_buffer containing pure-data
 455  *
 456  * Dirty volume headers and dirty meta-data buffers are locked until the
 457  * flusher can sequence them out.  Dirty pure-data buffers can be written.
 458  * Clean buffers can be passively released.
 459  */
 460 typedef enum hammer_io_type {
 461         HAMMER_STRUCTURE_VOLUME,
 462         HAMMER_STRUCTURE_META_BUFFER,
 463         HAMMER_STRUCTURE_UNDO_BUFFER,
 464         HAMMER_STRUCTURE_DATA_BUFFER
 465 } hammer_io_type_t;
 466
 467 union hammer_io_structure;
 468 struct hammer_io;
 469
 470 struct worklist {
 471         LIST_ENTRY(worklist) node;
 472 };
 473
 474 TAILQ_HEAD(hammer_io_list, hammer_io);
 475 typedef struct hammer_io_list *hammer_io_list_t;
 476
 477 struct hammer_io {
 478         struct worklist         worklist;
 479         struct hammer_lock      lock;
 480         enum hammer_io_type     type;
 481         struct hammer_mount     *hmp;
 482         TAILQ_ENTRY(hammer_io)  mod_entry; /* list entry if modified */
 483         hammer_io_list_t        mod_list;
 484         struct buf              *bp;
 485         int64_t                 offset;    /* zone-2 offset */
 486         int                     bytes;     /* buffer cache buffer size */
 487         int                     loading;   /* loading/unloading interlock */
 488         int                     modify_refs;
 489
 490         u_int           modified : 1;   /* bp's data was modified */
 491         u_int           released : 1;   /* bp released (w/ B_LOCKED set) */
 492         u_int           running : 1;    /* bp write IO in progress */
 493         u_int           waiting : 1;    /* someone is waiting on us */
 494         u_int           validated : 1;  /* ondisk has been validated */
 495         u_int           waitdep : 1;    /* flush waits for dependancies */
 496         u_int           recovered : 1;  /* has recovery ref */
 497         u_int           waitmod : 1;    /* waiting for modify_refs */
 498         u_int           reclaim : 1;    /* reclaim requested */
 499         u_int           gencrc : 1;     /* crc needs to be generated */
 500 };
 501
 502 typedef struct hammer_io *hammer_io_t;
 503
 504 #define HAMMER_CLUSTER_SIZE     (64 * 1024)
 505 #if HAMMER_CLUSTER_SIZE > MAXBSIZE
 506 #undef  HAMMER_CLUSTER_SIZE
 507 #define HAMMER_CLUSTER_SIZE     MAXBSIZE
 508 #endif
 509 #define HAMMER_CLUSTER_BUFS     (HAMMER_CLUSTER_SIZE / HAMMER_BUFSIZE)
 510
 511 /*
 512  * In-memory volume representing on-disk buffer
 513  */
 514 struct hammer_volume {
 515         struct hammer_io io;
 516         RB_ENTRY(hammer_volume) rb_node;
 517         struct hammer_volume_ondisk *ondisk;
 518         int32_t vol_no;
 519         int64_t nblocks;        /* note: special calculation for statfs */
 520         int64_t buffer_base;    /* base offset of buffer 0 */
 521         hammer_off_t maxbuf_off; /* Maximum buffer offset (zone-2) */
 522         hammer_off_t maxraw_off; /* Maximum raw offset for device */
 523         char    *vol_name;
 524         struct vnode *devvp;
 525         int     vol_flags;
 526 };
 527
 528 typedef struct hammer_volume *hammer_volume_t;
 529
 530 /*
 531  * In-memory buffer (other then volume, super-cluster, or cluster),
 532  * representing an on-disk buffer.
 533  */
 534 struct hammer_buffer {
 535         struct hammer_io io;
 536         RB_ENTRY(hammer_buffer) rb_node;
 537         void *ondisk;
 538         struct hammer_volume *volume;
 539         hammer_off_t zoneX_offset;
 540         hammer_off_t zone2_offset;
 541         struct hammer_reserve *resv;
 542         struct hammer_node_list clist;
 543 };
 544
 545 typedef struct hammer_buffer *hammer_buffer_t;
 546
 547 /*
 548  * In-memory B-Tree node, representing an on-disk B-Tree node.
 549  *
 550  * This is a hang-on structure which is backed by a hammer_buffer,
 551  * indexed by a hammer_cluster, and used for fine-grained locking of
 552  * B-Tree nodes in order to properly control lock ordering.  A hammer_buffer
 553  * can contain multiple nodes representing wildly disassociated portions
 554  * of the B-Tree so locking cannot be done on a buffer-by-buffer basis.
 555  *
 556  * This structure uses a cluster-relative index to reduce the number
 557  * of layers required to access it, and also because all on-disk B-Tree
 558  * references are cluster-relative offsets.
 559  */
 560 struct hammer_node {
 561         struct hammer_lock      lock;           /* node-by-node lock */
 562         TAILQ_ENTRY(hammer_node) entry;         /* per-buffer linkage */
 563         RB_ENTRY(hammer_node)   rb_node;        /* per-cluster linkage */
 564         hammer_off_t            node_offset;    /* full offset spec */
 565         struct hammer_mount     *hmp;
 566         struct hammer_buffer    *buffer;        /* backing buffer */
 567         hammer_node_ondisk_t    ondisk;         /* ptr to on-disk structure */
 568         TAILQ_HEAD(, hammer_cursor) cursor_list;  /* deadlock recovery */
 569         struct hammer_node_cache_list cache_list; /* passive caches */
 570         int                     flags;
 571         int                     loading;        /* load interlock */
 572 };
 573
 574 #define HAMMER_NODE_DELETED     0x0001
 575 #define HAMMER_NODE_FLUSH       0x0002
 576 #define HAMMER_NODE_CRCGOOD     0x0004
 577 #define HAMMER_NODE_NEEDSCRC    0x0008
 578 #define HAMMER_NODE_NEEDSMIRROR 0x0010
 579
 580 typedef struct hammer_node      *hammer_node_t;
 581
 582 /*
 583  * List of locked nodes.
 584  */
 585 struct hammer_node_locklist {
 586         struct hammer_node_locklist *next;
 587         hammer_node_t   node;
 588 };
 589
 590 typedef struct hammer_node_locklist *hammer_node_locklist_t;
 591
 592
 593 /*
 594  * Common I/O management structure - embedded in in-memory structures
 595  * which are backed by filesystem buffers.
 596  */
 597 union hammer_io_structure {
 598         struct hammer_io        io;
 599         struct hammer_volume    volume;
 600         struct hammer_buffer    buffer;
 601 };
 602
 603 typedef union hammer_io_structure *hammer_io_structure_t;
 604
 605 /*
 606  * The reserve structure prevents the blockmap from allocating
 607  * out of a reserved bigblock.  Such reservations are used by
 608  * the direct-write mechanism.
 609  *
 610  * The structure is also used to hold off on reallocations of
 611  * big blocks from the freemap until flush dependancies have
 612  * been dealt with.
 613  */
 614 struct hammer_reserve {
 615         RB_ENTRY(hammer_reserve) rb_node;
 616         TAILQ_ENTRY(hammer_reserve) delay_entry;
 617         int             flush_group;
 618         int             flags;
 619         int             refs;
 620         int             zone;
 621         int             append_off;
 622         hammer_off_t    zone_offset;
 623 };
 624
 625 typedef struct hammer_reserve *hammer_reserve_t;
 626
 627 #define HAMMER_RESF_ONDELAY     0x0001
 628
 629 #include "hammer_cursor.h"
 630
 631 /*
 632  * The undo structure tracks recent undos to avoid laying down duplicate
 633  * undos within a flush group, saving us a significant amount of overhead.
 634  *
 635  * This is strictly a heuristic.
 636  */
 637 #define HAMMER_MAX_UNDOS                1024
 638 #define HAMMER_MAX_FLUSHERS             4
 639
 640 struct hammer_undo {
 641         RB_ENTRY(hammer_undo)   rb_node;
 642         TAILQ_ENTRY(hammer_undo) lru_entry;
 643         hammer_off_t            offset;
 644         int                     bytes;
 645 };
 646
 647 typedef struct hammer_undo *hammer_undo_t;
 648
 649 struct hammer_flusher_info;
 650 TAILQ_HEAD(hammer_flusher_info_list, hammer_flusher_info);
 651
 652 struct hammer_flusher {
 653         int             signal;         /* flusher thread sequencer */
 654         int             act;            /* currently active flush group */
 655         int             done;           /* set to act when complete */
 656         int             next;           /* next flush group */
 657         int             group_lock;     /* lock sequencing of the next flush */
 658         int             exiting;        /* request master exit */
 659         thread_t        td;             /* master flusher thread */
 660         hammer_tid_t    tid;            /* last flushed transaction id */
 661         int             finalize_want;          /* serialize finalization */
 662         struct hammer_lock finalize_lock;       /* serialize finalization */
 663         struct hammer_transaction trans;        /* shared transaction */
 664         struct hammer_flusher_info_list run_list;
 665         struct hammer_flusher_info_list ready_list;
 666 };
 667
 668 /*
 669  * Internal hammer mount data structure
 670  */
 671 struct hammer_mount {
 672         struct mount *mp;
 673         /*struct vnode *rootvp;*/
 674         struct hammer_ino_rb_tree rb_inos_root;
 675         struct hammer_vol_rb_tree rb_vols_root;
 676         struct hammer_nod_rb_tree rb_nods_root;
 677         struct hammer_und_rb_tree rb_undo_root;
 678         struct hammer_res_rb_tree rb_resv_root;
 679         struct hammer_buf_rb_tree rb_bufs_root;
 680         struct hammer_pfs_rb_tree rb_pfsm_root;
 681         struct hammer_volume *rootvol;
 682         struct hammer_base_elm root_btree_beg;
 683         struct hammer_base_elm root_btree_end;
 684         int     flags;
 685         int     hflags;
 686         int     ronly;
 687         int     nvolumes;
 688         int     volume_iterator;
 689         int     masterid;       /* -1 or 0-15 - clustering and mirroring */
 690         int     rsv_inodes;     /* reserved space due to dirty inodes */
 691         int64_t rsv_databytes;  /* reserved space due to record data */
 692         int     rsv_recs;       /* reserved space due to dirty records */
 693         int     rsv_fromdelay;  /* bigblocks reserved due to flush delay */
 694         int     undo_rec_limit; /* based on size of undo area */
 695         int     last_newrecords;
 696         int     count_newrecords;
 697
 698         int     inode_reclaims; /* inodes pending reclaim by flusher */
 699         int     count_inodes;   /* total number of inodes */
 700         int     count_iqueued;  /* inodes queued to flusher */
 701
 702         struct hammer_flusher flusher;
 703
 704         u_int   check_interrupt;
 705         uuid_t  fsid;
 706         struct hammer_io_list volu_list;        /* dirty undo buffers */
 707         struct hammer_io_list undo_list;        /* dirty undo buffers */
 708         struct hammer_io_list data_list;        /* dirty data buffers */
 709         struct hammer_io_list alt_data_list;    /* dirty data buffers */
 710         struct hammer_io_list meta_list;        /* dirty meta bufs    */
 711         struct hammer_io_list lose_list;        /* loose buffers      */
 712         int     locked_dirty_space;             /* meta/volu count    */
 713         int     io_running_space;
 714         int     objid_cache_count;
 715         hammer_tid_t    asof;                   /* snapshot mount */
 716         hammer_off_t    next_tid;
 717         int64_t copy_stat_freebigblocks;        /* number of free bigblocks */
 718
 719         u_int32_t namekey_iterator;
 720         struct netexport export;
 721         struct hammer_lock sync_lock;
 722         struct hammer_lock free_lock;
 723         struct hammer_lock undo_lock;
 724         struct hammer_lock blkmap_lock;
 725         struct hammer_blockmap  blockmap[HAMMER_MAX_ZONES];
 726         struct hammer_undo      undos[HAMMER_MAX_UNDOS];
 727         int                     undo_alloc;
 728         TAILQ_HEAD(, hammer_undo)  undo_lru_list;
 729         TAILQ_HEAD(, hammer_reserve) delay_list;
 730         struct hammer_flush_group_list  flush_group_list;
 731         TAILQ_HEAD(, hammer_objid_cache) objid_cache_list;
 732         TAILQ_HEAD(, hammer_reclaim) reclaim_list;
 733 };
 734
 735 typedef struct hammer_mount     *hammer_mount_t;
 736
 737 #define HAMMER_MOUNT_UNUSED0001 0x0001
 738
 739 struct hammer_sync_info {
 740         int error;
 741         int waitfor;
 742 };
 743
 744 #endif
 745
 746 /*
 747  * checkspace slop (8MB chunks), higher numbers are more conservative.
 748  */
 749 #define HAMMER_CHKSPC_REBLOCK   25
 750 #define HAMMER_CHKSPC_MIRROR    20
 751 #define HAMMER_CHKSPC_WRITE     20
 752 #define HAMMER_CHKSPC_CREATE    20
 753 #define HAMMER_CHKSPC_REMOVE    10
 754
 755 #if defined(_KERNEL)
 756
 757 extern struct vop_ops hammer_vnode_vops;
 758 extern struct vop_ops hammer_spec_vops;
 759 extern struct vop_ops hammer_fifo_vops;
 760 extern struct bio_ops hammer_bioops;
 761
 762 extern int hammer_debug_io;
 763 extern int hammer_debug_general;
 764 extern int hammer_debug_debug;
 765 extern int hammer_debug_inode;
 766 extern int hammer_debug_locks;
 767 extern int hammer_debug_btree;
 768 extern int hammer_debug_tid;
 769 extern int hammer_debug_recover;
 770 extern int hammer_debug_recover_faults;
 771 extern int hammer_debug_cluster_enable;
 772 extern int hammer_count_fsyncs;
 773 extern int hammer_count_inodes;
 774 extern int hammer_count_iqueued;
 775 extern int hammer_count_reclaiming;
 776 extern int hammer_count_records;
 777 extern int hammer_count_record_datas;
 778 extern int hammer_count_volumes;
 779 extern int hammer_count_buffers;
 780 extern int hammer_count_nodes;
 781 extern int64_t hammer_count_extra_space_used;
 782 extern int64_t hammer_stats_btree_lookups;
 783 extern int64_t hammer_stats_btree_searches;
 784 extern int64_t hammer_stats_btree_inserts;
 785 extern int64_t hammer_stats_btree_deletes;
 786 extern int64_t hammer_stats_btree_elements;
 787 extern int64_t hammer_stats_btree_splits;
 788 extern int64_t hammer_stats_btree_iterations;
 789 extern int64_t hammer_stats_record_iterations;
 790 extern int hammer_count_dirtybufspace;
 791 extern int hammer_count_refedbufs;
 792 extern int hammer_count_reservations;
 793 extern int hammer_count_io_running_read;
 794 extern int hammer_count_io_running_write;
 795 extern int hammer_count_io_locked;
 796 extern int hammer_limit_dirtybufspace;
 797 extern int hammer_limit_iqueued;
 798 extern int hammer_limit_recs;
 799 extern int hammer_bio_count;
 800 extern int hammer_verify_zone;
 801 extern int hammer_write_mode;
 802 extern int64_t hammer_contention_count;
 803
 804 int     hammer_vop_inactive(struct vop_inactive_args *);
 805 int     hammer_vop_reclaim(struct vop_reclaim_args *);
 806 int     hammer_get_vnode(struct hammer_inode *ip, struct vnode **vpp);
 807 struct hammer_inode *hammer_get_inode(hammer_transaction_t trans,
 808                         hammer_inode_t dip, int64_t obj_id,
 809                         hammer_tid_t asof, u_int32_t localization,
 810                         int flags, int *errorp);
 811 void    hammer_scan_inode_snapshots(hammer_mount_t hmp,
 812                         hammer_inode_info_t iinfo,
 813                         int (*callback)(hammer_inode_t ip, void *data),
 814                         void *data);
 815 void    hammer_put_inode(struct hammer_inode *ip);
 816 void    hammer_put_inode_ref(struct hammer_inode *ip);
 817 void    hammer_inode_waitreclaims(hammer_mount_t hmp);
 818
 819 int     hammer_unload_volume(hammer_volume_t volume, void *data __unused);
 820 int     hammer_adjust_volume_mode(hammer_volume_t volume, void *data __unused);
 821
 822 int     hammer_unload_buffer(hammer_buffer_t buffer, void *data __unused);
 823 int     hammer_install_volume(hammer_mount_t hmp, const char *volname);
 824
 825 int     hammer_ip_lookup(hammer_cursor_t cursor);
 826 int     hammer_ip_first(hammer_cursor_t cursor);
 827 int     hammer_ip_next(hammer_cursor_t cursor);
 828 int     hammer_ip_resolve_data(hammer_cursor_t cursor);
 829 int     hammer_ip_delete_record(hammer_cursor_t cursor, hammer_inode_t ip,
 830                         hammer_tid_t tid);
 831 int     hammer_delete_at_cursor(hammer_cursor_t cursor, int delete_flags,
 832                         hammer_tid_t delete_tid, u_int32_t delete_ts,
 833                         int track, int64_t *stat_bytes);
 834 int     hammer_ip_check_directory_empty(hammer_transaction_t trans,
 835                         hammer_inode_t ip);
 836 int     hammer_sync_hmp(hammer_mount_t hmp, int waitfor);
 837 int     hammer_queue_inodes_flusher(hammer_mount_t hmp, int waitfor);
 838
 839 hammer_record_t
 840         hammer_alloc_mem_record(hammer_inode_t ip, int data_len);
 841 void    hammer_flush_record_done(hammer_record_t record, int error);
 842 void    hammer_wait_mem_record_ident(hammer_record_t record, const char *ident);
 843 void    hammer_rel_mem_record(hammer_record_t record);
 844
 845 int     hammer_cursor_up(hammer_cursor_t cursor);
 846 int     hammer_cursor_up_locked(hammer_cursor_t cursor);
 847 int     hammer_cursor_down(hammer_cursor_t cursor);
 848 int     hammer_cursor_upgrade(hammer_cursor_t cursor);
 849 int     hammer_cursor_upgrade_node(hammer_cursor_t cursor);
 850 void    hammer_cursor_downgrade(hammer_cursor_t cursor);
 851 int     hammer_cursor_seek(hammer_cursor_t cursor, hammer_node_t node,
 852                         int index);
 853 void    hammer_lock_ex_ident(struct hammer_lock *lock, const char *ident);
 854 int     hammer_lock_ex_try(struct hammer_lock *lock);
 855 void    hammer_lock_sh(struct hammer_lock *lock);
 856 int     hammer_lock_sh_try(struct hammer_lock *lock);
 857 int     hammer_lock_upgrade(struct hammer_lock *lock);
 858 void    hammer_lock_downgrade(struct hammer_lock *lock);
 859 int     hammer_lock_status(struct hammer_lock *lock);
 860 void    hammer_unlock(struct hammer_lock *lock);
 861 void    hammer_ref(struct hammer_lock *lock);
 862 void    hammer_unref(struct hammer_lock *lock);
 863
 864 void    hammer_sync_lock_ex(hammer_transaction_t trans);
 865 void    hammer_sync_lock_sh(hammer_transaction_t trans);
 866 int     hammer_sync_lock_sh_try(hammer_transaction_t trans);
 867 void    hammer_sync_unlock(hammer_transaction_t trans);
 868
 869 u_int32_t hammer_to_unix_xid(uuid_t *uuid);
 870 void hammer_guid_to_uuid(uuid_t *uuid, u_int32_t guid);
 871 void    hammer_time_to_timespec(u_int64_t xtime, struct timespec *ts);
 872 u_int64_t hammer_timespec_to_time(struct timespec *ts);
 873 hammer_tid_t hammer_str_to_tid(const char *str, int *ispfs,
 874                         u_int32_t *localizationp);
 875 hammer_tid_t hammer_alloc_objid(hammer_mount_t hmp, hammer_inode_t dip);
 876 void hammer_clear_objid(hammer_inode_t dip);
 877 void hammer_destroy_objid_cache(hammer_mount_t hmp);
 878
 879 int hammer_enter_undo_history(hammer_mount_t hmp, hammer_off_t offset,
 880                         int bytes);
 881 void hammer_clear_undo_history(hammer_mount_t hmp);
 882 enum vtype hammer_get_vnode_type(u_int8_t obj_type);
 883 int hammer_get_dtype(u_int8_t obj_type);
 884 u_int8_t hammer_get_obj_type(enum vtype vtype);
 885 int64_t hammer_directory_namekey(const void *name, int len);
 886 int     hammer_nohistory(hammer_inode_t ip);
 887
 888 int     hammer_init_cursor(hammer_transaction_t trans, hammer_cursor_t cursor,
 889                         hammer_node_cache_t cache, hammer_inode_t ip);
 890 void    hammer_normalize_cursor(hammer_cursor_t cursor);
 891 void    hammer_done_cursor(hammer_cursor_t cursor);
 892 int     hammer_recover_cursor(hammer_cursor_t cursor);
 893 void    hammer_unlock_cursor(hammer_cursor_t cursor, int also_ip);
 894 int     hammer_lock_cursor(hammer_cursor_t cursor, int also_ip);
 895 hammer_cursor_t hammer_push_cursor(hammer_cursor_t ocursor);
 896 void    hammer_pop_cursor(hammer_cursor_t ocursor, hammer_cursor_t ncursor);
 897
 898 void    hammer_cursor_replaced_node(hammer_node_t onode, hammer_node_t nnode);
 899 void    hammer_cursor_removed_node(hammer_node_t onode, hammer_node_t parent,
 900                         int index);
 901 void    hammer_cursor_split_node(hammer_node_t onode, hammer_node_t nnode,
 902                         int index);
 903 void    hammer_cursor_inserted_element(hammer_node_t node, int index);
 904 void    hammer_cursor_deleted_element(hammer_node_t node, int index);
 905
 906 int     hammer_btree_lookup(hammer_cursor_t cursor);
 907 int     hammer_btree_first(hammer_cursor_t cursor);
 908 int     hammer_btree_last(hammer_cursor_t cursor);
 909 int     hammer_btree_extract(hammer_cursor_t cursor, int flags);
 910 int     hammer_btree_iterate(hammer_cursor_t cursor);
 911 int     hammer_btree_iterate_reverse(hammer_cursor_t cursor);
 912 int     hammer_btree_insert(hammer_cursor_t cursor,
 913                             hammer_btree_leaf_elm_t elm, int *doprop);
 914 int     hammer_btree_delete(hammer_cursor_t cursor);
 915 void    hammer_btree_do_propagation(hammer_cursor_t cursor,
 916                             hammer_pseudofs_inmem_t pfsm,
 917                             hammer_btree_leaf_elm_t leaf);
 918 int     hammer_btree_cmp(hammer_base_elm_t key1, hammer_base_elm_t key2);
 919 int     hammer_btree_chkts(hammer_tid_t ts, hammer_base_elm_t key);
 920 int     hammer_btree_correct_rhb(hammer_cursor_t cursor, hammer_tid_t tid);
 921 int     hammer_btree_correct_lhb(hammer_cursor_t cursor, hammer_tid_t tid);
 922
 923 int     btree_set_parent(hammer_transaction_t trans, hammer_node_t node,
 924                         hammer_btree_elm_t elm);
 925 int     hammer_btree_lock_children(hammer_cursor_t cursor,
 926                         struct hammer_node_locklist **locklistp);
 927 void    hammer_btree_unlock_children(struct hammer_node_locklist **locklistp);
 928 int     hammer_btree_search_node(hammer_base_elm_t elm, hammer_node_ondisk_t node);
 929 hammer_node_t hammer_btree_get_parent(hammer_node_t node, int *parent_indexp,
 930                         int *errorp, int try_exclusive);
 931
 932 void    hammer_print_btree_node(hammer_node_ondisk_t ondisk);
 933 void    hammer_print_btree_elm(hammer_btree_elm_t elm, u_int8_t type, int i);
 934
 935 void    *hammer_bread(struct hammer_mount *hmp, hammer_off_t off,
 936                         int *errorp, struct hammer_buffer **bufferp);
 937 void    *hammer_bnew(struct hammer_mount *hmp, hammer_off_t off,
 938                         int *errorp, struct hammer_buffer **bufferp);
 939 void    *hammer_bread_ext(struct hammer_mount *hmp, hammer_off_t off, int bytes,
 940                         int *errorp, struct hammer_buffer **bufferp);
 941 void    *hammer_bnew_ext(struct hammer_mount *hmp, hammer_off_t off, int bytes,
 942                         int *errorp, struct hammer_buffer **bufferp);
 943
 944 hammer_volume_t hammer_get_root_volume(hammer_mount_t hmp, int *errorp);
 945
 946 hammer_volume_t hammer_get_volume(hammer_mount_t hmp,
 947                         int32_t vol_no, int *errorp);
 948 hammer_buffer_t hammer_get_buffer(hammer_mount_t hmp, hammer_off_t buf_offset,
 949                         int bytes, int isnew, int *errorp);
 950 void            hammer_del_buffers(hammer_mount_t hmp, hammer_off_t base_offset,
 951                         hammer_off_t zone2_offset, int bytes);
 952
 953 int             hammer_ref_volume(hammer_volume_t volume);
 954 int             hammer_ref_buffer(hammer_buffer_t buffer);
 955 void            hammer_flush_buffer_nodes(hammer_buffer_t buffer);
 956
 957 void            hammer_rel_volume(hammer_volume_t volume, int flush);
 958 void            hammer_rel_buffer(hammer_buffer_t buffer, int flush);
 959
 960 int             hammer_vfs_export(struct mount *mp, int op,
 961                         const struct export_args *export);
 962 hammer_node_t   hammer_get_node(hammer_mount_t hmp, hammer_off_t node_offset,
 963                         int isnew, int *errorp);
 964 void            hammer_ref_node(hammer_node_t node);
 965 hammer_node_t   hammer_ref_node_safe(struct hammer_mount *hmp,
 966                         hammer_node_cache_t cache, int *errorp);
 967 void            hammer_rel_node(hammer_node_t node);
 968 void            hammer_delete_node(hammer_transaction_t trans,
 969                         hammer_node_t node);
 970 void            hammer_cache_node(hammer_node_cache_t cache,
 971                         hammer_node_t node);
 972 void            hammer_uncache_node(hammer_node_cache_t cache);
 973 void            hammer_flush_node(hammer_node_t node);
 974
 975 void hammer_dup_buffer(struct hammer_buffer **bufferp,
 976                         struct hammer_buffer *buffer);
 977 hammer_node_t hammer_alloc_btree(hammer_transaction_t trans, int *errorp);
 978 void *hammer_alloc_data(hammer_transaction_t trans, int32_t data_len,
 979                         u_int16_t rec_type, hammer_off_t *data_offsetp,
 980                         struct hammer_buffer **data_bufferp, int *errorp);
 981
 982 int hammer_generate_undo(hammer_transaction_t trans, hammer_io_t io,
 983                         hammer_off_t zone1_offset, void *base, int len);
 984
 985 void hammer_put_volume(struct hammer_volume *volume, int flush);
 986 void hammer_put_buffer(struct hammer_buffer *buffer, int flush);
 987
 988 hammer_off_t hammer_freemap_alloc(hammer_transaction_t trans,
 989                         hammer_off_t owner, int *errorp);
 990 void hammer_freemap_free(hammer_transaction_t trans, hammer_off_t phys_offset,
 991                         hammer_off_t owner, int *errorp);
 992 int hammer_checkspace(hammer_mount_t hmp, int slop);
 993 hammer_off_t hammer_blockmap_alloc(hammer_transaction_t trans, int zone,
 994                         int bytes, int *errorp);
 995 hammer_reserve_t hammer_blockmap_reserve(hammer_mount_t hmp, int zone,
 996                         int bytes, hammer_off_t *zone_offp, int *errorp);
 997 void hammer_blockmap_reserve_complete(hammer_mount_t hmp,
 998                         hammer_reserve_t resv);
 999 void hammer_reserve_clrdelay(hammer_mount_t hmp, hammer_reserve_t resv);
1000 void hammer_blockmap_free(hammer_transaction_t trans,
1001                         hammer_off_t bmap_off, int bytes);
1002 void hammer_blockmap_finalize(hammer_transaction_t trans,
1003                         hammer_off_t bmap_off, int bytes);
1004 int hammer_blockmap_getfree(hammer_mount_t hmp, hammer_off_t bmap_off,
1005                         int *curp, int *errorp);
1006 hammer_off_t hammer_blockmap_lookup(hammer_mount_t hmp, hammer_off_t bmap_off,
1007                         int *errorp);
1008 hammer_off_t hammer_undo_lookup(hammer_mount_t hmp, hammer_off_t bmap_off,
1009                         int *errorp);
1010 int64_t hammer_undo_used(hammer_transaction_t trans);
1011 int64_t hammer_undo_space(hammer_transaction_t trans);
1012 int64_t hammer_undo_max(hammer_mount_t hmp);
1013
1014 void hammer_start_transaction(struct hammer_transaction *trans,
1015                               struct hammer_mount *hmp);
1016 void hammer_simple_transaction(struct hammer_transaction *trans,
1017                               struct hammer_mount *hmp);
1018 void hammer_start_transaction_fls(struct hammer_transaction *trans,
1019                                   struct hammer_mount *hmp);
1020 void hammer_done_transaction(struct hammer_transaction *trans);
1021
1022 void hammer_modify_inode(hammer_inode_t ip, int flags);
1023 void hammer_flush_inode(hammer_inode_t ip, int flags);
1024 void hammer_flush_inode_done(hammer_inode_t ip);
1025 void hammer_wait_inode(hammer_inode_t ip);
1026
1027 int  hammer_create_inode(struct hammer_transaction *trans, struct vattr *vap,
1028                         struct ucred *cred, struct hammer_inode *dip,
1029                         hammer_pseudofs_inmem_t pfsm,
1030                         struct hammer_inode **ipp);
1031 void hammer_rel_inode(hammer_inode_t ip, int flush);
1032 int hammer_reload_inode(hammer_inode_t ip, void *arg __unused);
1033 int hammer_ino_rb_compare(hammer_inode_t ip1, hammer_inode_t ip2);
1034
1035 int hammer_sync_inode(hammer_inode_t ip);
1036 void hammer_test_inode(hammer_inode_t ip);
1037 void hammer_inode_unloadable_check(hammer_inode_t ip, int getvp);
1038
1039 int  hammer_ip_add_directory(struct hammer_transaction *trans,
1040                         hammer_inode_t dip, const char *name, int bytes,
1041                         hammer_inode_t nip);
1042 int  hammer_ip_del_directory(struct hammer_transaction *trans,
1043                         hammer_cursor_t cursor, hammer_inode_t dip,
1044                         hammer_inode_t ip);
1045 hammer_record_t hammer_ip_add_bulk(hammer_inode_t ip, off_t file_offset,
1046                         void *data, int bytes, int *errorp);
1047 int  hammer_ip_frontend_trunc(struct hammer_inode *ip, off_t file_size);
1048 int  hammer_ip_add_record(struct hammer_transaction *trans,
1049                         hammer_record_t record);
1050 int  hammer_ip_delete_range(hammer_cursor_t cursor, hammer_inode_t ip,
1051                         int64_t ran_beg, int64_t ran_end, int truncating);
1052 int  hammer_ip_delete_clean(hammer_cursor_t cursor, hammer_inode_t ip,
1053                         int *countp);
1054 int  hammer_ip_sync_data(hammer_cursor_t cursor, hammer_inode_t ip,
1055                         int64_t offset, void *data, int bytes);
1056 int  hammer_ip_sync_record(hammer_transaction_t trans, hammer_record_t rec);
1057 int  hammer_ip_sync_record_cursor(hammer_cursor_t cursor, hammer_record_t rec);
1058 hammer_pseudofs_inmem_t  hammer_load_pseudofs(hammer_transaction_t trans,
1059                         u_int32_t localization, int *errorp);
1060 int  hammer_mkroot_pseudofs(hammer_transaction_t trans, struct ucred *cred,
1061                         hammer_pseudofs_inmem_t pfsm);
1062 int  hammer_save_pseudofs(hammer_transaction_t trans,
1063                         hammer_pseudofs_inmem_t pfsm);
1064 int  hammer_unload_pseudofs(hammer_transaction_t trans, u_int32_t localization);
1065 void hammer_rel_pseudofs(hammer_mount_t hmp, hammer_pseudofs_inmem_t pfsm);
1066 int hammer_ioctl(hammer_inode_t ip, u_long com, caddr_t data, int fflag,
1067                         struct ucred *cred);
1068
1069 void hammer_io_init(hammer_io_t io, hammer_mount_t hmp,
1070                         enum hammer_io_type type);
1071 int hammer_io_read(struct vnode *devvp, struct hammer_io *io,
1072                         hammer_off_t limit);
1073 int hammer_io_new(struct vnode *devvp, struct hammer_io *io);
1074 void hammer_io_inval(hammer_volume_t volume, hammer_off_t zone2_offset);
1075 struct buf *hammer_io_release(struct hammer_io *io, int flush);
1076 void hammer_io_flush(struct hammer_io *io);
1077 void hammer_io_waitdep(struct hammer_io *io);
1078 void hammer_io_wait_all(hammer_mount_t hmp, const char *ident);
1079 int hammer_io_direct_read(hammer_mount_t hmp, struct bio *bio);
1080 int hammer_io_direct_write(hammer_mount_t hmp, hammer_btree_leaf_elm_t leaf,
1081                           struct bio *bio);
1082 void hammer_io_direct_uncache(hammer_mount_t hmp, hammer_btree_leaf_elm_t leaf);
1083 void hammer_io_write_interlock(hammer_io_t io);
1084 void hammer_io_done_interlock(hammer_io_t io);
1085 void hammer_io_clear_modify(struct hammer_io *io, int inval);
1086 void hammer_io_clear_modlist(struct hammer_io *io);
1087 void hammer_modify_volume(hammer_transaction_t trans, hammer_volume_t volume,
1088                         void *base, int len);
1089 void hammer_modify_buffer(hammer_transaction_t trans, hammer_buffer_t buffer,
1090                         void *base, int len);
1091 void hammer_modify_volume_done(hammer_volume_t volume);
1092 void hammer_modify_buffer_done(hammer_buffer_t buffer);
1093
1094 int hammer_ioc_reblock(hammer_transaction_t trans, hammer_inode_t ip,
1095                         struct hammer_ioc_reblock *reblock);
1096 int hammer_ioc_prune(hammer_transaction_t trans, hammer_inode_t ip,
1097                         struct hammer_ioc_prune *prune);
1098 int hammer_ioc_mirror_read(hammer_transaction_t trans, hammer_inode_t ip,
1099                         struct hammer_ioc_mirror_rw *mirror);
1100 int hammer_ioc_mirror_write(hammer_transaction_t trans, hammer_inode_t ip,
1101                         struct hammer_ioc_mirror_rw *mirror);
1102 int hammer_ioc_set_pseudofs(hammer_transaction_t trans, hammer_inode_t ip,
1103                         struct ucred *cred, struct hammer_ioc_pseudofs_rw *pfs);
1104 int hammer_ioc_get_pseudofs(hammer_transaction_t trans, hammer_inode_t ip,
1105                         struct hammer_ioc_pseudofs_rw *pfs);
1106 int hammer_ioc_destroy_pseudofs(hammer_transaction_t trans, hammer_inode_t ip,
1107                         struct hammer_ioc_pseudofs_rw *pfs);
1108 int hammer_ioc_downgrade_pseudofs(hammer_transaction_t trans, hammer_inode_t ip,
1109                         struct hammer_ioc_pseudofs_rw *pfs);
1110 int hammer_ioc_upgrade_pseudofs(hammer_transaction_t trans, hammer_inode_t ip,
1111                         struct hammer_ioc_pseudofs_rw *pfs);
1112
1113 int hammer_signal_check(hammer_mount_t hmp);
1114
1115 void hammer_flusher_create(hammer_mount_t hmp);
1116 void hammer_flusher_destroy(hammer_mount_t hmp);
1117 void hammer_flusher_sync(hammer_mount_t hmp);
1118 int  hammer_flusher_async(hammer_mount_t hmp, hammer_flush_group_t flg);
1119 void hammer_flusher_wait(hammer_mount_t hmp, int seq);
1120 int  hammer_flusher_meta_limit(hammer_mount_t hmp);
1121 int  hammer_flusher_meta_halflimit(hammer_mount_t hmp);
1122 int  hammer_flusher_undo_exhausted(hammer_transaction_t trans, int quarter);
1123 void hammer_flusher_clean_loose_ios(hammer_mount_t hmp);
1124 void hammer_flusher_finalize(hammer_transaction_t trans, int final);
1125
1126 int hammer_recover(hammer_mount_t hmp, hammer_volume_t rootvol);
1127 void hammer_recover_flush_buffers(hammer_mount_t hmp,
1128                         hammer_volume_t root_volume, int final);
1129
1130 void hammer_crc_set_blockmap(hammer_blockmap_t blockmap);
1131 void hammer_crc_set_volume(hammer_volume_ondisk_t ondisk);
1132 void hammer_crc_set_leaf(void *data, hammer_btree_leaf_elm_t leaf);
1133
1134 int hammer_crc_test_blockmap(hammer_blockmap_t blockmap);
1135 int hammer_crc_test_volume(hammer_volume_ondisk_t ondisk);
1136 int hammer_crc_test_btree(hammer_node_ondisk_t ondisk);
1137 int hammer_crc_test_leaf(void *data, hammer_btree_leaf_elm_t leaf);
1138 void hkprintf(const char *ctl, ...);
1139 udev_t hammer_fsid_to_udev(uuid_t *uuid);
1140
1141
1142 int hammer_blocksize(int64_t file_offset);
1143 int64_t hammer_blockdemarc(int64_t file_offset1, int64_t file_offset2);
1144
1145 #endif
1146
1147 static __inline void
1148 hammer_wait_mem_record(hammer_record_t record)
1149 {
1150         hammer_wait_mem_record_ident(record, "hmmwai");
1151 }
1152
1153 static __inline void
1154 hammer_lock_ex(struct hammer_lock *lock)
1155 {
1156         hammer_lock_ex_ident(lock, "hmrlck");
1157 }
1158
1159 /*
1160  * Indicate that a B-Tree node is being modified.
1161  */
1162 static __inline void
1163 hammer_modify_node_noundo(hammer_transaction_t trans, hammer_node_t node)
1164 {
1165         hammer_modify_buffer(trans, node->buffer, NULL, 0);
1166 }
1167
1168 static __inline void
1169 hammer_modify_node_all(hammer_transaction_t trans, struct hammer_node *node)
1170 {
1171         hammer_modify_buffer(trans, node->buffer,
1172                              node->ondisk, sizeof(*node->ondisk));
1173 }
1174
1175 static __inline void
1176 hammer_modify_node(hammer_transaction_t trans, hammer_node_t node,
1177                    void *base, int len)
1178 {
1179         hammer_crc_t *crcptr;
1180
1181         KKASSERT((char *)base >= (char *)node->ondisk &&
1182                  (char *)base + len <=
1183                     (char *)node->ondisk + sizeof(*node->ondisk));
1184         hammer_modify_buffer(trans, node->buffer, base, len);
1185         crcptr = &node->ondisk->crc;
1186         hammer_modify_buffer(trans, node->buffer, crcptr, sizeof(hammer_crc_t));
1187         --node->buffer->io.modify_refs; /* only want one ref */
1188 }
1189
1190 /*
1191  * Indicate that the specified modifications have been completed.
1192  *
1193  * Do not try to generate the crc here, it's very expensive to do and a
1194  * sequence of insertions or deletions can result in many calls to this
1195  * function on the same node.
1196  */
1197 static __inline void
1198 hammer_modify_node_done(hammer_node_t node)
1199 {
1200         node->flags |= HAMMER_NODE_CRCGOOD;
1201         if ((node->flags & HAMMER_NODE_NEEDSCRC) == 0) {
1202                 node->flags |= HAMMER_NODE_NEEDSCRC;
1203                 node->buffer->io.gencrc = 1;
1204                 hammer_ref_node(node);
1205         }
1206         hammer_modify_buffer_done(node->buffer);
1207 }
1208
1209 #define hammer_modify_volume_field(trans, vol, field)           \
1210         hammer_modify_volume(trans, vol, &(vol)->ondisk->field, \
1211                              sizeof((vol)->ondisk->field))
1212
1213 #define hammer_modify_node_field(trans, node, field)            \
1214         hammer_modify_node(trans, node, &(node)->ondisk->field, \
1215                              sizeof((node)->ondisk->field))
1216