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 1.130 2008/11/13 02:18:43 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         int             flags;
 113         struct hammer_volume *rootvol;
 114 };
 115
 116 typedef struct hammer_transaction *hammer_transaction_t;
 117
 118 #define HAMMER_TRANSF_NEWINODE  0x0001
 119
 120 /*
 121  * HAMMER locks
 122  */
 123 struct hammer_lock {
 124         int     refs;           /* active references delay writes */
 125         int     lockcount;      /* lock count for exclusive/shared access */
 126         int     wanted;
 127         int     exwanted;       /* number of threads waiting for ex lock */
 128         struct thread *locktd;
 129 };
 130
 131 static __inline int
 132 hammer_islocked(struct hammer_lock *lock)
 133 {
 134         return(lock->lockcount != 0);
 135 }
 136
 137 static __inline int
 138 hammer_isactive(struct hammer_lock *lock)
 139 {
 140         return(lock->refs != 0);
 141 }
 142
 143 static __inline int
 144 hammer_islastref(struct hammer_lock *lock)
 145 {
 146         return(lock->refs == 1);
 147 }
 148
 149 /*
 150  * Return if we specifically own the lock exclusively.
 151  */
 152 static __inline int
 153 hammer_lock_excl_owned(struct hammer_lock *lock, thread_t td)
 154 {
 155         if (lock->lockcount > 0 && lock->locktd == td)
 156                 return(1);
 157         return(0);
 158 }
 159
 160 /*
 161  * Flush state, used by various structures
 162  */
 163 typedef enum hammer_inode_state {
 164         HAMMER_FST_IDLE,
 165         HAMMER_FST_SETUP,
 166         HAMMER_FST_FLUSH
 167 } hammer_inode_state_t;
 168
 169 TAILQ_HEAD(hammer_record_list, hammer_record);
 170
 171 /*
 172  * Pseudo-filesystem extended data tracking
 173  */
 174 struct hammer_pfs_rb_tree;
 175 struct hammer_pseudofs_inmem;
 176 RB_HEAD(hammer_pfs_rb_tree, hammer_pseudofs_inmem);
 177 RB_PROTOTYPE2(hammer_pfs_rb_tree, hammer_pseudofs_inmem, rb_node,
 178               hammer_pfs_rb_compare, u_int32_t);
 179
 180 struct hammer_pseudofs_inmem {
 181         RB_ENTRY(hammer_pseudofs_inmem) rb_node;
 182         struct hammer_lock      lock;
 183         u_int32_t               localization;
 184         hammer_tid_t            create_tid;
 185         int                     flags;
 186         udev_t                  fsid_udev;
 187         struct hammer_pseudofs_data pfsd;
 188 };
 189
 190 typedef struct hammer_pseudofs_inmem *hammer_pseudofs_inmem_t;
 191
 192 #define HAMMER_PFSM_DELETED     0x0001
 193
 194 /*
 195  * Cache object ids.  A fixed number of objid cache structures are
 196  * created to reserve object id's for newly created files in multiples
 197  * of 100,000, localized to a particular directory, and recycled as
 198  * needed.  This allows parallel create operations in different
 199  * directories to retain fairly localized object ids which in turn
 200  * improves reblocking performance and layout.
 201  */
 202 #define OBJID_CACHE_SIZE        1024
 203 #define OBJID_CACHE_BULK        100000
 204
 205 typedef struct hammer_objid_cache {
 206         TAILQ_ENTRY(hammer_objid_cache) entry;
 207         struct hammer_inode             *dip;
 208         hammer_tid_t                    next_tid;
 209         int                             count;
 210 } *hammer_objid_cache_t;
 211
 212 /*
 213  * Associate an inode with a B-Tree node to cache search start positions
 214  */
 215 typedef struct hammer_node_cache {
 216         TAILQ_ENTRY(hammer_node_cache)  entry;
 217         struct hammer_node              *node;
 218         struct hammer_inode             *ip;
 219 } *hammer_node_cache_t;
 220
 221 TAILQ_HEAD(hammer_node_cache_list, hammer_node_cache);
 222
 223 /*
 224  * Structure used to organize flush groups.  Flush groups must be
 225  * organized into chunks in order to avoid blowing out the UNDO FIFO.
 226  * Without this a 'sync' could end up flushing 50,000 inodes in a single
 227  * transaction.
 228  */
 229 struct hammer_flush_group {
 230         TAILQ_ENTRY(hammer_flush_group) flush_entry;
 231         TAILQ_HEAD(, hammer_inode)      flush_list;
 232         int                             inode_count;    /* inode load */
 233         int                             total_count;    /* record load */
 234         int                             running;        /* group is running */
 235         int                             closed;
 236         int                             refs;
 237 };
 238
 239 typedef struct hammer_flush_group *hammer_flush_group_t;
 240
 241 TAILQ_HEAD(hammer_flush_group_list, hammer_flush_group);
 242
 243 /*
 244  * Structure used to represent an inode in-memory.
 245  *
 246  * The record and data associated with an inode may be out of sync with
 247  * the disk (xDIRTY flags), or not even on the disk at all (ONDISK flag
 248  * clear).
 249  *
 250  * An inode may also hold a cache of unsynchronized records, used for
 251  * database and directories only.  Unsynchronized regular file data is
 252  * stored in the buffer cache.
 253  *
 254  * NOTE: A file which is created and destroyed within the initial
 255  * synchronization period can wind up not doing any disk I/O at all.
 256  *
 257  * Finally, an inode may cache numerous disk-referencing B-Tree cursors.
 258  */
 259 struct hammer_ino_rb_tree;
 260 struct hammer_inode;
 261 RB_HEAD(hammer_ino_rb_tree, hammer_inode);
 262 RB_PROTOTYPEX(hammer_ino_rb_tree, INFO, hammer_inode, rb_node,
 263               hammer_ino_rb_compare, hammer_inode_info_t);
 264
 265 struct hammer_rec_rb_tree;
 266 struct hammer_record;
 267 RB_HEAD(hammer_rec_rb_tree, hammer_record);
 268 RB_PROTOTYPEX(hammer_rec_rb_tree, INFO, hammer_record, rb_node,
 269               hammer_rec_rb_compare, hammer_btree_leaf_elm_t);
 270
 271 TAILQ_HEAD(hammer_node_list, hammer_node);
 272
 273 struct hammer_inode {
 274         RB_ENTRY(hammer_inode)  rb_node;
 275         hammer_inode_state_t    flush_state;
 276         hammer_flush_group_t    flush_group;
 277         TAILQ_ENTRY(hammer_inode) flush_entry;
 278         struct hammer_record_list target_list;  /* target of dependant recs */
 279         int64_t                 obj_id;         /* (key) object identifier */
 280         hammer_tid_t            obj_asof;       /* (key) snapshot or 0 */
 281         u_int32_t               obj_localization; /* (key) pseudo-fs */
 282         struct hammer_mount     *hmp;
 283         hammer_objid_cache_t    objid_cache;
 284         int                     flags;
 285         int                     error;          /* flush error */
 286         int                     cursor_ip_refs; /* sanity */
 287         int                     rsv_recs;
 288         struct vnode            *vp;
 289         hammer_pseudofs_inmem_t pfsm;
 290         struct lockf            advlock;
 291         struct hammer_lock      lock;           /* sync copy interlock */
 292         off_t                   trunc_off;
 293         struct hammer_btree_leaf_elm ino_leaf;  /* in-memory cache */
 294         struct hammer_inode_data ino_data;      /* in-memory cache */
 295         struct hammer_rec_rb_tree rec_tree;     /* in-memory cache */
 296         struct hammer_node_cache cache[2];      /* search initiate cache */
 297
 298         /*
 299          * When a demark is created to synchronize an inode to
 300          * disk, certain fields are copied so the front-end VOPs
 301          * can continue to run in parallel with the synchronization
 302          * occuring in the background.
 303          */
 304         int             sync_flags;             /* to-sync flags cache */
 305         off_t           sync_trunc_off;         /* to-sync truncation */
 306         off_t           save_trunc_off;         /* write optimization */
 307         struct hammer_btree_leaf_elm sync_ino_leaf; /* to-sync cache */
 308         struct hammer_inode_data sync_ino_data; /* to-sync cache */
 309 };
 310
 311 typedef struct hammer_inode *hammer_inode_t;
 312
 313 #define VTOI(vp)        ((struct hammer_inode *)(vp)->v_data)
 314
 315 #define HAMMER_INODE_DDIRTY     0x0001  /* in-memory ino_data is dirty */
 316                                         /* (not including atime/mtime) */
 317 #define HAMMER_INODE_RSV_INODES 0x0002  /* hmp->rsv_inodes bumped */
 318 #define HAMMER_INODE_CONN_DOWN  0x0004  /* include in downward recursion */
 319 #define HAMMER_INODE_XDIRTY     0x0008  /* in-memory records */
 320 #define HAMMER_INODE_ONDISK     0x0010  /* inode is on-disk (else not yet) */
 321 #define HAMMER_INODE_FLUSH      0x0020  /* flush on last ref */
 322 #define HAMMER_INODE_DELETED    0x0080  /* inode delete (backend) */
 323 #define HAMMER_INODE_DELONDISK  0x0100  /* delete synchronized to disk */
 324 #define HAMMER_INODE_RO         0x0200  /* read-only (because of as-of) */
 325 #define HAMMER_INODE_VHELD      0x0400  /* vnode held on sync */
 326 #define HAMMER_INODE_DONDISK    0x0800  /* data records may be on disk */
 327 #define HAMMER_INODE_BUFS       0x1000  /* dirty high level bps present */
 328 #define HAMMER_INODE_REFLUSH    0x2000  /* flush on dependancy / reflush */
 329 #define HAMMER_INODE_RECLAIM    0x4000  /* trying to reclaim */
 330 #define HAMMER_INODE_FLUSHW     0x8000  /* Someone waiting for flush */
 331
 332 #define HAMMER_INODE_TRUNCATED  0x00010000
 333 #define HAMMER_INODE_DELETING   0x00020000 /* inode delete request (frontend)*/
 334 #define HAMMER_INODE_RESIGNAL   0x00040000 /* re-signal on re-flush */
 335 #define HAMMER_INODE_ATIME      0x00100000 /* in-memory atime modified */
 336 #define HAMMER_INODE_MTIME      0x00200000 /* in-memory mtime modified */
 337 #define HAMMER_INODE_WOULDBLOCK 0x00400000 /* re-issue to new flush group */
 338
 339 #define HAMMER_INODE_MODMASK    (HAMMER_INODE_DDIRTY|                       \
 340                                  HAMMER_INODE_XDIRTY|HAMMER_INODE_BUFS|     \
 341                                  HAMMER_INODE_ATIME|HAMMER_INODE_MTIME|     \
 342                                  HAMMER_INODE_TRUNCATED|HAMMER_INODE_DELETING)
 343
 344 #define HAMMER_INODE_MODMASK_NOXDIRTY \
 345                                 (HAMMER_INODE_MODMASK & ~HAMMER_INODE_XDIRTY)
 346
 347 #define HAMMER_FLUSH_GROUP_SIZE 64
 348
 349 #define HAMMER_FLUSH_SIGNAL     0x0001
 350 #define HAMMER_FLUSH_RECURSION  0x0002
 351
 352 /*
 353  * Used by the inode reclaim code to pipeline reclaims and avoid
 354  * blowing out kernel memory or letting the flusher get too far
 355  * behind.
 356  */
 357 struct hammer_reclaim {
 358         TAILQ_ENTRY(hammer_reclaim) entry;
 359         int     okydoky;
 360 };
 361
 362 #define HAMMER_RECLAIM_FLUSH    2000
 363 #define HAMMER_RECLAIM_WAIT     4000
 364
 365 /*
 366  * Structure used to represent an unsynchronized record in-memory.  These
 367  * records typically represent directory entries.  Only non-historical
 368  * records are kept in-memory.
 369  *
 370  * Records are organized as a per-inode RB-Tree.  If the inode is not
 371  * on disk then neither are any records and the in-memory record tree
 372  * represents the entire contents of the inode.  If the inode is on disk
 373  * then the on-disk B-Tree is scanned in parallel with the in-memory
 374  * RB-Tree to synthesize the current state of the file.
 375  *
 376  * Records are also used to enforce the ordering of directory create/delete
 377  * operations.  A new inode will not be flushed to disk unless its related
 378  * directory entry is also being flushed at the same time.  A directory entry
 379  * will not be removed unless its related inode is also being removed at the
 380  * same time.
 381  */
 382 typedef enum hammer_record_type {
 383         HAMMER_MEM_RECORD_GENERAL,      /* misc record */
 384         HAMMER_MEM_RECORD_INODE,        /* inode record */
 385         HAMMER_MEM_RECORD_ADD,          /* positive memory cache record */
 386         HAMMER_MEM_RECORD_DEL,          /* negative delete-on-disk record */
 387         HAMMER_MEM_RECORD_DATA          /* bulk-data record w/on-disk ref */
 388 } hammer_record_type_t;
 389
 390 struct hammer_record {
 391         RB_ENTRY(hammer_record)         rb_node;
 392         TAILQ_ENTRY(hammer_record)      target_entry;
 393         hammer_inode_state_t            flush_state;
 394         hammer_flush_group_t            flush_group;
 395         hammer_record_type_t            type;
 396         struct hammer_lock              lock;
 397         struct hammer_reserve           *resv;
 398         struct hammer_inode             *ip;
 399         struct hammer_inode             *target_ip;
 400         struct hammer_btree_leaf_elm    leaf;
 401         union hammer_data_ondisk        *data;
 402         int                             flags;
 403         hammer_off_t                    zone2_offset;   /* direct-write only */
 404 };
 405
 406 typedef struct hammer_record *hammer_record_t;
 407
 408 /*
 409  * Record flags.  Note that FE can only be set by the frontend if the
 410  * record has not been interlocked by the backend w/ BE.
 411  */
 412 #define HAMMER_RECF_ALLOCDATA           0x0001
 413 #define HAMMER_RECF_ONRBTREE            0x0002
 414 #define HAMMER_RECF_DELETED_FE          0x0004  /* deleted (frontend) */
 415 #define HAMMER_RECF_DELETED_BE          0x0008  /* deleted (backend) */
 416 #define HAMMER_RECF_COMMITTED           0x0010  /* committed to the B-Tree */
 417 #define HAMMER_RECF_INTERLOCK_BE        0x0020  /* backend interlock */
 418 #define HAMMER_RECF_WANTED              0x0040  /* wanted by the frontend */
 419 #define HAMMER_RECF_CONVERT_DELETE      0x0100  /* special case */
 420 #define HAMMER_RECF_DIRECT_IO           0x0200  /* related direct I/O running*/
 421 #define HAMMER_RECF_DIRECT_WAIT         0x0400  /* related direct I/O running*/
 422 #define HAMMER_RECF_DIRECT_INVAL        0x0800  /* buffer alias invalidation */
 423
 424 /*
 425  * hammer_delete_at_cursor() flags
 426  */
 427 #define HAMMER_DELETE_ADJUST            0x0001
 428 #define HAMMER_DELETE_DESTROY           0x0002
 429
 430 /*
 431  * In-memory structures representing on-disk structures.
 432  */
 433 struct hammer_volume;
 434 struct hammer_buffer;
 435 struct hammer_node;
 436 struct hammer_undo;
 437 struct hammer_reserve;
 438
 439 RB_HEAD(hammer_vol_rb_tree, hammer_volume);
 440 RB_HEAD(hammer_buf_rb_tree, hammer_buffer);
 441 RB_HEAD(hammer_nod_rb_tree, hammer_node);
 442 RB_HEAD(hammer_und_rb_tree, hammer_undo);
 443 RB_HEAD(hammer_res_rb_tree, hammer_reserve);
 444
 445 RB_PROTOTYPE2(hammer_vol_rb_tree, hammer_volume, rb_node,
 446               hammer_vol_rb_compare, int32_t);
 447 RB_PROTOTYPE2(hammer_buf_rb_tree, hammer_buffer, rb_node,
 448               hammer_buf_rb_compare, hammer_off_t);
 449 RB_PROTOTYPE2(hammer_nod_rb_tree, hammer_node, rb_node,
 450               hammer_nod_rb_compare, hammer_off_t);
 451 RB_PROTOTYPE2(hammer_und_rb_tree, hammer_undo, rb_node,
 452               hammer_und_rb_compare, hammer_off_t);
 453 RB_PROTOTYPE2(hammer_res_rb_tree, hammer_reserve, rb_node,
 454               hammer_res_rb_compare, hammer_off_t);
 455
 456 /*
 457  * IO management - embedded at the head of various in-memory structures
 458  *
 459  * VOLUME       - hammer_volume containing meta-data
 460  * META_BUFFER  - hammer_buffer containing meta-data
 461  * DATA_BUFFER  - hammer_buffer containing pure-data
 462  *
 463  * Dirty volume headers and dirty meta-data buffers are locked until the
 464  * flusher can sequence them out.  Dirty pure-data buffers can be written.
 465  * Clean buffers can be passively released.
 466  */
 467 typedef enum hammer_io_type {
 468         HAMMER_STRUCTURE_VOLUME,
 469         HAMMER_STRUCTURE_META_BUFFER,
 470         HAMMER_STRUCTURE_UNDO_BUFFER,
 471         HAMMER_STRUCTURE_DATA_BUFFER
 472 } hammer_io_type_t;
 473
 474 union hammer_io_structure;
 475 struct hammer_io;
 476
 477 struct worklist {
 478         LIST_ENTRY(worklist) node;
 479 };
 480
 481 TAILQ_HEAD(hammer_io_list, hammer_io);
 482 typedef struct hammer_io_list *hammer_io_list_t;
 483
 484 struct hammer_io {
 485         struct worklist         worklist;
 486         struct hammer_lock      lock;
 487         enum hammer_io_type     type;
 488         struct hammer_mount     *hmp;
 489         struct hammer_volume    *volume;
 490         TAILQ_ENTRY(hammer_io)  mod_entry; /* list entry if modified */
 491         hammer_io_list_t        mod_list;
 492         struct buf              *bp;
 493         int64_t                 offset;    /* zone-2 offset */
 494         int                     bytes;     /* buffer cache buffer size */
 495         int                     loading;   /* loading/unloading interlock */
 496         int                     modify_refs;
 497
 498         u_int           modified : 1;   /* bp's data was modified */
 499         u_int           released : 1;   /* bp released (w/ B_LOCKED set) */
 500         u_int           running : 1;    /* bp write IO in progress */
 501         u_int           waiting : 1;    /* someone is waiting on us */
 502         u_int           validated : 1;  /* ondisk has been validated */
 503         u_int           waitdep : 1;    /* flush waits for dependancies */
 504         u_int           recovered : 1;  /* has recovery ref */
 505         u_int           waitmod : 1;    /* waiting for modify_refs */
 506         u_int           reclaim : 1;    /* reclaim requested */
 507         u_int           gencrc : 1;     /* crc needs to be generated */
 508         u_int           ioerror : 1;    /* abort on io-error */
 509 };
 510
 511 typedef struct hammer_io *hammer_io_t;
 512
 513 #define HAMMER_CLUSTER_SIZE     (64 * 1024)
 514 #if HAMMER_CLUSTER_SIZE > MAXBSIZE
 515 #undef  HAMMER_CLUSTER_SIZE
 516 #define HAMMER_CLUSTER_SIZE     MAXBSIZE
 517 #endif
 518 #define HAMMER_CLUSTER_BUFS     (HAMMER_CLUSTER_SIZE / HAMMER_BUFSIZE)
 519
 520 /*
 521  * In-memory volume representing on-disk buffer
 522  */
 523 struct hammer_volume {
 524         struct hammer_io io;
 525         RB_ENTRY(hammer_volume) rb_node;
 526         struct hammer_volume_ondisk *ondisk;
 527         int32_t vol_no;
 528         int64_t nblocks;        /* note: special calculation for statfs */
 529         int64_t buffer_base;    /* base offset of buffer 0 */
 530         hammer_off_t maxbuf_off; /* Maximum buffer offset (zone-2) */
 531         hammer_off_t maxraw_off; /* Maximum raw offset for device */
 532         char    *vol_name;
 533         struct vnode *devvp;
 534         int     vol_flags;
 535 };
 536
 537 typedef struct hammer_volume *hammer_volume_t;
 538
 539 /*
 540  * In-memory buffer (other then volume, super-cluster, or cluster),
 541  * representing an on-disk buffer.
 542  */
 543 struct hammer_buffer {
 544         struct hammer_io io;
 545         RB_ENTRY(hammer_buffer) rb_node;
 546         void *ondisk;
 547         hammer_off_t zoneX_offset;
 548         hammer_off_t zone2_offset;
 549         struct hammer_reserve *resv;
 550         struct hammer_node_list clist;
 551 };
 552
 553 typedef struct hammer_buffer *hammer_buffer_t;
 554
 555 /*
 556  * In-memory B-Tree node, representing an on-disk B-Tree node.
 557  *
 558  * This is a hang-on structure which is backed by a hammer_buffer,
 559  * indexed by a hammer_cluster, and used for fine-grained locking of
 560  * B-Tree nodes in order to properly control lock ordering.  A hammer_buffer
 561  * can contain multiple nodes representing wildly disassociated portions
 562  * of the B-Tree so locking cannot be done on a buffer-by-buffer basis.
 563  *
 564  * This structure uses a cluster-relative index to reduce the number
 565  * of layers required to access it, and also because all on-disk B-Tree
 566  * references are cluster-relative offsets.
 567  */
 568 struct hammer_node {
 569         struct hammer_lock      lock;           /* node-by-node lock */
 570         TAILQ_ENTRY(hammer_node) entry;         /* per-buffer linkage */
 571         RB_ENTRY(hammer_node)   rb_node;        /* per-cluster linkage */
 572         hammer_off_t            node_offset;    /* full offset spec */
 573         struct hammer_mount     *hmp;
 574         struct hammer_buffer    *buffer;        /* backing buffer */
 575         hammer_node_ondisk_t    ondisk;         /* ptr to on-disk structure */
 576         TAILQ_HEAD(, hammer_cursor) cursor_list;  /* deadlock recovery */
 577         struct hammer_node_cache_list cache_list; /* passive caches */
 578         int                     flags;
 579         int                     loading;        /* load interlock */
 580 };
 581
 582 #define HAMMER_NODE_DELETED     0x0001
 583 #define HAMMER_NODE_FLUSH       0x0002
 584 #define HAMMER_NODE_CRCGOOD     0x0004
 585 #define HAMMER_NODE_NEEDSCRC    0x0008
 586 #define HAMMER_NODE_NEEDSMIRROR 0x0010
 587
 588 typedef struct hammer_node      *hammer_node_t;
 589
 590 /*
 591  * List of locked nodes.
 592  */
 593 struct hammer_node_locklist {
 594         struct hammer_node_locklist *next;
 595         hammer_node_t   node;
 596 };
 597
 598 typedef struct hammer_node_locklist *hammer_node_locklist_t;
 599
 600
 601 /*
 602  * Common I/O management structure - embedded in in-memory structures
 603  * which are backed by filesystem buffers.
 604  */
 605 union hammer_io_structure {
 606         struct hammer_io        io;
 607         struct hammer_volume    volume;
 608         struct hammer_buffer    buffer;
 609 };
 610
 611 typedef union hammer_io_structure *hammer_io_structure_t;
 612
 613 /*
 614  * The reserve structure prevents the blockmap from allocating
 615  * out of a reserved bigblock.  Such reservations are used by
 616  * the direct-write mechanism.
 617  *
 618  * The structure is also used to hold off on reallocations of
 619  * big blocks from the freemap until flush dependancies have
 620  * been dealt with.
 621  */
 622 struct hammer_reserve {
 623         RB_ENTRY(hammer_reserve) rb_node;
 624         TAILQ_ENTRY(hammer_reserve) delay_entry;
 625         int             flush_group;
 626         int             flags;
 627         int             refs;
 628         int             zone;
 629         int             append_off;
 630         hammer_off_t    zone_offset;
 631 };
 632
 633 typedef struct hammer_reserve *hammer_reserve_t;
 634
 635 #define HAMMER_RESF_ONDELAY     0x0001
 636 #define HAMMER_RESF_LAYER2FREE  0x0002
 637
 638 #include "hammer_cursor.h"
 639
 640 /*
 641  * The undo structure tracks recent undos to avoid laying down duplicate
 642  * undos within a flush group, saving us a significant amount of overhead.
 643  *
 644  * This is strictly a heuristic.
 645  */
 646 #define HAMMER_MAX_UNDOS                1024
 647 #define HAMMER_MAX_FLUSHERS             4
 648
 649 struct hammer_undo {
 650         RB_ENTRY(hammer_undo)   rb_node;
 651         TAILQ_ENTRY(hammer_undo) lru_entry;
 652         hammer_off_t            offset;
 653         int                     bytes;
 654 };
 655
 656 typedef struct hammer_undo *hammer_undo_t;
 657
 658 struct hammer_flusher_info;
 659 TAILQ_HEAD(hammer_flusher_info_list, hammer_flusher_info);
 660
 661 struct hammer_flusher {
 662         int             signal;         /* flusher thread sequencer */
 663         int             act;            /* currently active flush group */
 664         int             done;           /* set to act when complete */
 665         int             next;           /* next flush group */
 666         int             group_lock;     /* lock sequencing of the next flush */
 667         int             exiting;        /* request master exit */
 668         thread_t        td;             /* master flusher thread */
 669         hammer_tid_t    tid;            /* last flushed transaction id */
 670         int             finalize_want;          /* serialize finalization */
 671         struct hammer_lock finalize_lock;       /* serialize finalization */
 672         struct hammer_transaction trans;        /* shared transaction */
 673         struct hammer_flusher_info_list run_list;
 674         struct hammer_flusher_info_list ready_list;
 675 };
 676
 677 /*
 678  * Internal hammer mount data structure
 679  */
 680 struct hammer_mount {
 681         struct mount *mp;
 682         /*struct vnode *rootvp;*/
 683         struct hammer_ino_rb_tree rb_inos_root;
 684         struct hammer_vol_rb_tree rb_vols_root;
 685         struct hammer_nod_rb_tree rb_nods_root;
 686         struct hammer_und_rb_tree rb_undo_root;
 687         struct hammer_res_rb_tree rb_resv_root;
 688         struct hammer_buf_rb_tree rb_bufs_root;
 689         struct hammer_pfs_rb_tree rb_pfsm_root;
 690         struct hammer_volume *rootvol;
 691         struct hammer_base_elm root_btree_beg;
 692         struct hammer_base_elm root_btree_end;
 693
 694         struct malloc_type      *m_misc;
 695         struct malloc_type      *m_inodes;
 696
 697         int     flags;          /* HAMMER_MOUNT_xxx flags */
 698         int     hflags;
 699         int     ronly;
 700         int     nvolumes;
 701         int     volume_iterator;
 702         int     master_id;      /* -1 or 0-15 - clustering and mirroring */
 703         int     version;        /* hammer filesystem version to use */
 704         int     rsv_inodes;     /* reserved space due to dirty inodes */
 705         int64_t rsv_databytes;  /* reserved space due to record data */
 706         int     rsv_recs;       /* reserved space due to dirty records */
 707         int     rsv_fromdelay;  /* bigblocks reserved due to flush delay */
 708         int     undo_rec_limit; /* based on size of undo area */
 709         int     last_newrecords;
 710         int     count_newrecords;
 711
 712         int     inode_reclaims; /* inodes pending reclaim by flusher */
 713         int     count_inodes;   /* total number of inodes */
 714         int     count_iqueued;  /* inodes queued to flusher */
 715
 716         struct hammer_flusher flusher;
 717
 718         u_int   check_interrupt;
 719         uuid_t  fsid;
 720         struct hammer_io_list volu_list;        /* dirty undo buffers */
 721         struct hammer_io_list undo_list;        /* dirty undo buffers */
 722         struct hammer_io_list data_list;        /* dirty data buffers */
 723         struct hammer_io_list alt_data_list;    /* dirty data buffers */
 724         struct hammer_io_list meta_list;        /* dirty meta bufs    */
 725         struct hammer_io_list lose_list;        /* loose buffers      */
 726         int     locked_dirty_space;             /* meta/volu count    */
 727         int     io_running_space;
 728         int     objid_cache_count;
 729         int     error;                          /* critical I/O error */
 730         struct krate    krate;                  /* rate limited kprintf */
 731         hammer_tid_t    asof;                   /* snapshot mount */
 732         hammer_tid_t    next_tid;
 733         hammer_tid_t    flush_tid1;             /* flusher tid sequencing */
 734         hammer_tid_t    flush_tid2;             /* flusher tid sequencing */
 735         int64_t copy_stat_freebigblocks;        /* number of free bigblocks */
 736
 737         struct netexport export;
 738         struct hammer_lock sync_lock;
 739         struct hammer_lock free_lock;
 740         struct hammer_lock undo_lock;
 741         struct hammer_lock blkmap_lock;
 742         struct hammer_blockmap  blockmap[HAMMER_MAX_ZONES];
 743         struct hammer_undo      undos[HAMMER_MAX_UNDOS];
 744         int                     undo_alloc;
 745         TAILQ_HEAD(, hammer_undo)  undo_lru_list;
 746         TAILQ_HEAD(, hammer_reserve) delay_list;
 747         struct hammer_flush_group_list  flush_group_list;
 748         hammer_flush_group_t    next_flush_group;
 749         TAILQ_HEAD(, hammer_objid_cache) objid_cache_list;
 750         TAILQ_HEAD(, hammer_reclaim) reclaim_list;
 751 };
 752
 753 typedef struct hammer_mount     *hammer_mount_t;
 754
 755 #define HAMMER_MOUNT_CRITICAL_ERROR     0x0001
 756
 757 struct hammer_sync_info {
 758         int error;
 759         int waitfor;
 760 };
 761
 762 #endif
 763
 764 /*
 765  * checkspace slop (8MB chunks), higher numbers are more conservative.
 766  */
 767 #define HAMMER_CHKSPC_REBLOCK   25
 768 #define HAMMER_CHKSPC_MIRROR    20
 769 #define HAMMER_CHKSPC_WRITE     20
 770 #define HAMMER_CHKSPC_CREATE    20
 771 #define HAMMER_CHKSPC_REMOVE    10
 772 #define HAMMER_CHKSPC_EMERGENCY 0
 773
 774 #if defined(_KERNEL)
 775
 776 extern struct vop_ops hammer_vnode_vops;
 777 extern struct vop_ops hammer_spec_vops;
 778 extern struct vop_ops hammer_fifo_vops;
 779 extern struct bio_ops hammer_bioops;
 780
 781 extern int hammer_debug_io;
 782 extern int hammer_debug_general;
 783 extern int hammer_debug_debug;
 784 extern int hammer_debug_inode;
 785 extern int hammer_debug_locks;
 786 extern int hammer_debug_btree;
 787 extern int hammer_debug_tid;
 788 extern int hammer_debug_recover;
 789 extern int hammer_debug_recover_faults;
 790 extern int hammer_cluster_enable;
 791 extern int hammer_count_fsyncs;
 792 extern int hammer_count_inodes;
 793 extern int hammer_count_iqueued;
 794 extern int hammer_count_reclaiming;
 795 extern int hammer_count_records;
 796 extern int hammer_count_record_datas;
 797 extern int hammer_count_volumes;
 798 extern int hammer_count_buffers;
 799 extern int hammer_count_nodes;
 800 extern int64_t hammer_count_extra_space_used;
 801 extern int64_t hammer_stats_btree_lookups;
 802 extern int64_t hammer_stats_btree_searches;
 803 extern int64_t hammer_stats_btree_inserts;
 804 extern int64_t hammer_stats_btree_deletes;
 805 extern int64_t hammer_stats_btree_elements;
 806 extern int64_t hammer_stats_btree_splits;
 807 extern int64_t hammer_stats_btree_iterations;
 808 extern int64_t hammer_stats_record_iterations;
 809 extern int64_t hammer_stats_file_read;
 810 extern int64_t hammer_stats_file_write;
 811 extern int64_t hammer_stats_file_iopsr;
 812 extern int64_t hammer_stats_file_iopsw;
 813 extern int64_t hammer_stats_disk_read;
 814 extern int64_t hammer_stats_disk_write;
 815 extern int64_t hammer_stats_inode_flushes;
 816 extern int64_t hammer_stats_commits;
 817 extern int hammer_count_dirtybufspace;
 818 extern int hammer_count_refedbufs;
 819 extern int hammer_count_reservations;
 820 extern int hammer_count_io_running_read;
 821 extern int hammer_count_io_running_write;
 822 extern int hammer_count_io_locked;
 823 extern int hammer_limit_dirtybufspace;
 824 extern int hammer_limit_iqueued;
 825 extern int hammer_limit_recs;
 826 extern int hammer_bio_count;
 827 extern int hammer_verify_zone;
 828 extern int hammer_verify_data;
 829 extern int hammer_write_mode;
 830 extern int hammer_autoflush;
 831 extern int64_t hammer_contention_count;
 832
 833 void    hammer_critical_error(hammer_mount_t hmp, hammer_inode_t ip,
 834                         int error, const char *msg);
 835 int     hammer_vop_inactive(struct vop_inactive_args *);
 836 int     hammer_vop_reclaim(struct vop_reclaim_args *);
 837 int     hammer_get_vnode(struct hammer_inode *ip, struct vnode **vpp);
 838 struct hammer_inode *hammer_get_inode(hammer_transaction_t trans,
 839                         hammer_inode_t dip, int64_t obj_id,
 840                         hammer_tid_t asof, u_int32_t localization,
 841                         int flags, int *errorp);
 842 void    hammer_scan_inode_snapshots(hammer_mount_t hmp,
 843                         hammer_inode_info_t iinfo,
 844                         int (*callback)(hammer_inode_t ip, void *data),
 845                         void *data);
 846 void    hammer_put_inode(struct hammer_inode *ip);
 847 void    hammer_put_inode_ref(struct hammer_inode *ip);
 848 void    hammer_inode_waitreclaims(hammer_mount_t hmp);
 849
 850 int     hammer_unload_volume(hammer_volume_t volume, void *data __unused);
 851 int     hammer_adjust_volume_mode(hammer_volume_t volume, void *data __unused);
 852
 853 int     hammer_unload_buffer(hammer_buffer_t buffer, void *data __unused);
 854 int     hammer_install_volume(hammer_mount_t hmp, const char *volname,
 855                         struct vnode *devvp);
 856 int     hammer_mountcheck_volumes(hammer_mount_t hmp);
 857
 858 int     hammer_mem_add(hammer_record_t record);
 859 int     hammer_ip_lookup(hammer_cursor_t cursor);
 860 int     hammer_ip_first(hammer_cursor_t cursor);
 861 int     hammer_ip_next(hammer_cursor_t cursor);
 862 int     hammer_ip_resolve_data(hammer_cursor_t cursor);
 863 int     hammer_ip_delete_record(hammer_cursor_t cursor, hammer_inode_t ip,
 864                         hammer_tid_t tid);
 865 int     hammer_delete_at_cursor(hammer_cursor_t cursor, int delete_flags,
 866                         hammer_tid_t delete_tid, u_int32_t delete_ts,
 867                         int track, int64_t *stat_bytes);
 868 int     hammer_ip_check_directory_empty(hammer_transaction_t trans,
 869                         hammer_inode_t ip);
 870 int     hammer_sync_hmp(hammer_mount_t hmp, int waitfor);
 871 int     hammer_queue_inodes_flusher(hammer_mount_t hmp, int waitfor);
 872
 873 hammer_record_t
 874         hammer_alloc_mem_record(hammer_inode_t ip, int data_len);
 875 void    hammer_flush_record_done(hammer_record_t record, int error);
 876 void    hammer_wait_mem_record_ident(hammer_record_t record, const char *ident);
 877 void    hammer_rel_mem_record(hammer_record_t record);
 878
 879 int     hammer_cursor_up(hammer_cursor_t cursor);
 880 int     hammer_cursor_up_locked(hammer_cursor_t cursor);
 881 int     hammer_cursor_down(hammer_cursor_t cursor);
 882 int     hammer_cursor_upgrade(hammer_cursor_t cursor);
 883 int     hammer_cursor_upgrade_node(hammer_cursor_t cursor);
 884 void    hammer_cursor_downgrade(hammer_cursor_t cursor);
 885 int     hammer_cursor_seek(hammer_cursor_t cursor, hammer_node_t node,
 886                         int index);
 887 void    hammer_lock_ex_ident(struct hammer_lock *lock, const char *ident);
 888 int     hammer_lock_ex_try(struct hammer_lock *lock);
 889 void    hammer_lock_sh(struct hammer_lock *lock);
 890 int     hammer_lock_sh_try(struct hammer_lock *lock);
 891 int     hammer_lock_upgrade(struct hammer_lock *lock);
 892 void    hammer_lock_downgrade(struct hammer_lock *lock);
 893 int     hammer_lock_status(struct hammer_lock *lock);
 894 void    hammer_unlock(struct hammer_lock *lock);
 895 void    hammer_ref(struct hammer_lock *lock);
 896 void    hammer_unref(struct hammer_lock *lock);
 897
 898 void    hammer_sync_lock_ex(hammer_transaction_t trans);
 899 void    hammer_sync_lock_sh(hammer_transaction_t trans);
 900 int     hammer_sync_lock_sh_try(hammer_transaction_t trans);
 901 void    hammer_sync_unlock(hammer_transaction_t trans);
 902
 903 u_int32_t hammer_to_unix_xid(uuid_t *uuid);
 904 void hammer_guid_to_uuid(uuid_t *uuid, u_int32_t guid);
 905 void    hammer_time_to_timespec(u_int64_t xtime, struct timespec *ts);
 906 u_int64_t hammer_timespec_to_time(struct timespec *ts);
 907 int     hammer_str_to_tid(const char *str, int *ispfsp,
 908                         hammer_tid_t *tidp, u_int32_t *localizationp);
 909 int     hammer_is_atatext(const char *name, int len);
 910 hammer_tid_t hammer_alloc_objid(hammer_mount_t hmp, hammer_inode_t dip);
 911 void hammer_clear_objid(hammer_inode_t dip);
 912 void hammer_destroy_objid_cache(hammer_mount_t hmp);
 913
 914 int hammer_enter_undo_history(hammer_mount_t hmp, hammer_off_t offset,
 915                         int bytes);
 916 void hammer_clear_undo_history(hammer_mount_t hmp);
 917 enum vtype hammer_get_vnode_type(u_int8_t obj_type);
 918 int hammer_get_dtype(u_int8_t obj_type);
 919 u_int8_t hammer_get_obj_type(enum vtype vtype);
 920 int64_t hammer_directory_namekey(hammer_inode_t dip, const void *name, int len,
 921                         u_int32_t *max_iterationsp);
 922 int     hammer_nohistory(hammer_inode_t ip);
 923
 924 int     hammer_init_cursor(hammer_transaction_t trans, hammer_cursor_t cursor,
 925                         hammer_node_cache_t cache, hammer_inode_t ip);
 926 void    hammer_normalize_cursor(hammer_cursor_t cursor);
 927 void    hammer_done_cursor(hammer_cursor_t cursor);
 928 int     hammer_recover_cursor(hammer_cursor_t cursor);
 929 void    hammer_unlock_cursor(hammer_cursor_t cursor, int also_ip);
 930 int     hammer_lock_cursor(hammer_cursor_t cursor, int also_ip);
 931 hammer_cursor_t hammer_push_cursor(hammer_cursor_t ocursor);
 932 void    hammer_pop_cursor(hammer_cursor_t ocursor, hammer_cursor_t ncursor);
 933
 934 void    hammer_cursor_replaced_node(hammer_node_t onode, hammer_node_t nnode);
 935 void    hammer_cursor_removed_node(hammer_node_t onode, hammer_node_t parent,
 936                         int index);
 937 void    hammer_cursor_split_node(hammer_node_t onode, hammer_node_t nnode,
 938                         int index);
 939 void    hammer_cursor_inserted_element(hammer_node_t node, int index);
 940 void    hammer_cursor_deleted_element(hammer_node_t node, int index);
 941
 942 int     hammer_btree_lookup(hammer_cursor_t cursor);
 943 int     hammer_btree_first(hammer_cursor_t cursor);
 944 int     hammer_btree_last(hammer_cursor_t cursor);
 945 int     hammer_btree_extract(hammer_cursor_t cursor, int flags);
 946 int     hammer_btree_iterate(hammer_cursor_t cursor);
 947 int     hammer_btree_iterate_reverse(hammer_cursor_t cursor);
 948 int     hammer_btree_insert(hammer_cursor_t cursor,
 949                             hammer_btree_leaf_elm_t elm, int *doprop);
 950 int     hammer_btree_delete(hammer_cursor_t cursor);
 951 void    hammer_btree_do_propagation(hammer_cursor_t cursor,
 952                             hammer_pseudofs_inmem_t pfsm,
 953                             hammer_btree_leaf_elm_t leaf);
 954 int     hammer_btree_cmp(hammer_base_elm_t key1, hammer_base_elm_t key2);
 955 int     hammer_btree_chkts(hammer_tid_t ts, hammer_base_elm_t key);
 956 int     hammer_btree_correct_rhb(hammer_cursor_t cursor, hammer_tid_t tid);
 957 int     hammer_btree_correct_lhb(hammer_cursor_t cursor, hammer_tid_t tid);
 958
 959 int     btree_set_parent(hammer_transaction_t trans, hammer_node_t node,
 960                         hammer_btree_elm_t elm);
 961 int     hammer_btree_lock_children(hammer_cursor_t cursor,
 962                         struct hammer_node_locklist **locklistp);
 963 void    hammer_btree_unlock_children(hammer_cursor_t cursor,
 964                         struct hammer_node_locklist **locklistp);
 965 int     hammer_btree_search_node(hammer_base_elm_t elm, hammer_node_ondisk_t node);
 966 hammer_node_t hammer_btree_get_parent(hammer_node_t node, int *parent_indexp,
 967                         int *errorp, int try_exclusive);
 968
 969 void    hammer_print_btree_node(hammer_node_ondisk_t ondisk);
 970 void    hammer_print_btree_elm(hammer_btree_elm_t elm, u_int8_t type, int i);
 971
 972 void    *hammer_bread(struct hammer_mount *hmp, hammer_off_t off,
 973                         int *errorp, struct hammer_buffer **bufferp);
 974 void    *hammer_bnew(struct hammer_mount *hmp, hammer_off_t off,
 975                         int *errorp, struct hammer_buffer **bufferp);
 976 void    *hammer_bread_ext(struct hammer_mount *hmp, hammer_off_t off, int bytes,
 977                         int *errorp, struct hammer_buffer **bufferp);
 978 void    *hammer_bnew_ext(struct hammer_mount *hmp, hammer_off_t off, int bytes,
 979                         int *errorp, struct hammer_buffer **bufferp);
 980
 981 hammer_volume_t hammer_get_root_volume(hammer_mount_t hmp, int *errorp);
 982
 983 hammer_volume_t hammer_get_volume(hammer_mount_t hmp,
 984                         int32_t vol_no, int *errorp);
 985 hammer_buffer_t hammer_get_buffer(hammer_mount_t hmp, hammer_off_t buf_offset,
 986                         int bytes, int isnew, int *errorp);
 987 void            hammer_sync_buffers(hammer_mount_t hmp,
 988                         hammer_off_t base_offset, int bytes);
 989 void            hammer_del_buffers(hammer_mount_t hmp,
 990                         hammer_off_t base_offset,
 991                         hammer_off_t zone2_offset, int bytes);
 992
 993 int             hammer_ref_volume(hammer_volume_t volume);
 994 int             hammer_ref_buffer(hammer_buffer_t buffer);
 995 void            hammer_flush_buffer_nodes(hammer_buffer_t buffer);
 996
 997 void            hammer_rel_volume(hammer_volume_t volume, int flush);
 998 void            hammer_rel_buffer(hammer_buffer_t buffer, int flush);
 999
1000 int             hammer_vfs_export(struct mount *mp, int op,
1001                         const struct export_args *export);
1002 hammer_node_t   hammer_get_node(hammer_mount_t hmp, hammer_off_t node_offset,
1003                         int isnew, int *errorp);
1004 void            hammer_ref_node(hammer_node_t node);
1005 hammer_node_t   hammer_ref_node_safe(struct hammer_mount *hmp,
1006                         hammer_node_cache_t cache, int *errorp);
1007 void            hammer_rel_node(hammer_node_t node);
1008 void            hammer_delete_node(hammer_transaction_t trans,
1009                         hammer_node_t node);
1010 void            hammer_cache_node(hammer_node_cache_t cache,
1011                         hammer_node_t node);
1012 void            hammer_uncache_node(hammer_node_cache_t cache);
1013 void            hammer_flush_node(hammer_node_t node);
1014
1015 void hammer_dup_buffer(struct hammer_buffer **bufferp,
1016                         struct hammer_buffer *buffer);
1017 hammer_node_t hammer_alloc_btree(hammer_transaction_t trans, int *errorp);
1018 void *hammer_alloc_data(hammer_transaction_t trans, int32_t data_len,
1019                         u_int16_t rec_type, hammer_off_t *data_offsetp,
1020                         struct hammer_buffer **data_bufferp, int *errorp);
1021
1022 int hammer_generate_undo(hammer_transaction_t trans, hammer_io_t io,
1023                         hammer_off_t zone1_offset, void *base, int len);
1024
1025 void hammer_put_volume(struct hammer_volume *volume, int flush);
1026 void hammer_put_buffer(struct hammer_buffer *buffer, int flush);
1027
1028 hammer_off_t hammer_freemap_alloc(hammer_transaction_t trans,
1029                         hammer_off_t owner, int *errorp);
1030 void hammer_freemap_free(hammer_transaction_t trans, hammer_off_t phys_offset,
1031                         hammer_off_t owner, int *errorp);
1032 int hammer_checkspace(hammer_mount_t hmp, int slop);
1033 hammer_off_t hammer_blockmap_alloc(hammer_transaction_t trans, int zone,
1034                         int bytes, int *errorp);
1035 hammer_reserve_t hammer_blockmap_reserve(hammer_mount_t hmp, int zone,
1036                         int bytes, hammer_off_t *zone_offp, int *errorp);
1037 void hammer_blockmap_reserve_undo(hammer_mount_t hmp, hammer_reserve_t resv,
1038                         hammer_off_t zone_offset, int bytes);
1039 void hammer_blockmap_reserve_complete(hammer_mount_t hmp,
1040                         hammer_reserve_t resv);
1041 void hammer_reserve_clrdelay(hammer_mount_t hmp, hammer_reserve_t resv);
1042 void hammer_blockmap_free(hammer_transaction_t trans,
1043                         hammer_off_t bmap_off, int bytes);
1044 int hammer_blockmap_finalize(hammer_transaction_t trans,
1045                         hammer_reserve_t resv,
1046                         hammer_off_t bmap_off, int bytes);
1047 int hammer_blockmap_getfree(hammer_mount_t hmp, hammer_off_t bmap_off,
1048                         int *curp, int *errorp);
1049 hammer_off_t hammer_blockmap_lookup(hammer_mount_t hmp, hammer_off_t bmap_off,
1050                         int *errorp);
1051 hammer_off_t hammer_undo_lookup(hammer_mount_t hmp, hammer_off_t bmap_off,
1052                         int *errorp);
1053 int64_t hammer_undo_used(hammer_transaction_t trans);
1054 int64_t hammer_undo_space(hammer_transaction_t trans);
1055 int64_t hammer_undo_max(hammer_mount_t hmp);
1056
1057 void hammer_start_transaction(struct hammer_transaction *trans,
1058                               struct hammer_mount *hmp);
1059 void hammer_simple_transaction(struct hammer_transaction *trans,
1060                               struct hammer_mount *hmp);
1061 void hammer_start_transaction_fls(struct hammer_transaction *trans,
1062                                   struct hammer_mount *hmp);
1063 void hammer_done_transaction(struct hammer_transaction *trans);
1064
1065 void hammer_modify_inode(hammer_inode_t ip, int flags);
1066 void hammer_flush_inode(hammer_inode_t ip, int flags);
1067 void hammer_flush_inode_done(hammer_inode_t ip, int error);
1068 void hammer_wait_inode(hammer_inode_t ip);
1069
1070 int  hammer_create_inode(struct hammer_transaction *trans, struct vattr *vap,
1071                         struct ucred *cred, struct hammer_inode *dip,
1072                         hammer_pseudofs_inmem_t pfsm,
1073                         struct hammer_inode **ipp);
1074 void hammer_rel_inode(hammer_inode_t ip, int flush);
1075 int hammer_reload_inode(hammer_inode_t ip, void *arg __unused);
1076 int hammer_ino_rb_compare(hammer_inode_t ip1, hammer_inode_t ip2);
1077 int hammer_destroy_inode_callback(hammer_inode_t ip, void *data __unused);
1078
1079 int hammer_sync_inode(hammer_transaction_t trans, hammer_inode_t ip);
1080 void hammer_test_inode(hammer_inode_t dip);
1081 void hammer_inode_unloadable_check(hammer_inode_t ip, int getvp);
1082
1083 int  hammer_ip_add_directory(struct hammer_transaction *trans,
1084                         hammer_inode_t dip, const char *name, int bytes,
1085                         hammer_inode_t nip);
1086 int  hammer_ip_del_directory(struct hammer_transaction *trans,
1087                         hammer_cursor_t cursor, hammer_inode_t dip,
1088                         hammer_inode_t ip);
1089 hammer_record_t hammer_ip_add_bulk(hammer_inode_t ip, off_t file_offset,
1090                         void *data, int bytes, int *errorp);
1091 int  hammer_ip_frontend_trunc(struct hammer_inode *ip, off_t file_size);
1092 int  hammer_ip_add_record(struct hammer_transaction *trans,
1093                         hammer_record_t record);
1094 int  hammer_ip_delete_range(hammer_cursor_t cursor, hammer_inode_t ip,
1095                         int64_t ran_beg, int64_t ran_end, int truncating);
1096 int  hammer_ip_delete_clean(hammer_cursor_t cursor, hammer_inode_t ip,
1097                         int *countp);
1098 int  hammer_ip_sync_data(hammer_cursor_t cursor, hammer_inode_t ip,
1099                         int64_t offset, void *data, int bytes);
1100 int  hammer_ip_sync_record(hammer_transaction_t trans, hammer_record_t rec);
1101 int  hammer_ip_sync_record_cursor(hammer_cursor_t cursor, hammer_record_t rec);
1102 hammer_pseudofs_inmem_t  hammer_load_pseudofs(hammer_transaction_t trans,
1103                         u_int32_t localization, int *errorp);
1104 int  hammer_mkroot_pseudofs(hammer_transaction_t trans, struct ucred *cred,
1105                         hammer_pseudofs_inmem_t pfsm);
1106 int  hammer_save_pseudofs(hammer_transaction_t trans,
1107                         hammer_pseudofs_inmem_t pfsm);
1108 int  hammer_unload_pseudofs(hammer_transaction_t trans, u_int32_t localization);
1109 void hammer_rel_pseudofs(hammer_mount_t hmp, hammer_pseudofs_inmem_t pfsm);
1110 int hammer_ioctl(hammer_inode_t ip, u_long com, caddr_t data, int fflag,
1111                         struct ucred *cred);
1112
1113 void hammer_io_init(hammer_io_t io, hammer_volume_t volume,
1114                         enum hammer_io_type type);
1115 int hammer_io_read(struct vnode *devvp, struct hammer_io *io,
1116                         hammer_off_t limit);
1117 int hammer_io_new(struct vnode *devvp, struct hammer_io *io);
1118 void hammer_io_inval(hammer_volume_t volume, hammer_off_t zone2_offset);
1119 struct buf *hammer_io_release(struct hammer_io *io, int flush);
1120 void hammer_io_flush(struct hammer_io *io);
1121 void hammer_io_wait(struct hammer_io *io);
1122 void hammer_io_waitdep(struct hammer_io *io);
1123 void hammer_io_wait_all(hammer_mount_t hmp, const char *ident);
1124 int hammer_io_direct_read(hammer_mount_t hmp, struct bio *bio,
1125                         hammer_btree_leaf_elm_t leaf);
1126 int hammer_io_direct_write(hammer_mount_t hmp, hammer_record_t record,
1127                         struct bio *bio);
1128 void hammer_io_direct_wait(hammer_record_t record);
1129 void hammer_io_direct_uncache(hammer_mount_t hmp, hammer_btree_leaf_elm_t leaf);
1130 void hammer_io_write_interlock(hammer_io_t io);
1131 void hammer_io_done_interlock(hammer_io_t io);
1132 void hammer_io_clear_modify(struct hammer_io *io, int inval);
1133 void hammer_io_clear_modlist(struct hammer_io *io);
1134 void hammer_io_flush_sync(hammer_mount_t hmp);
1135
1136 void hammer_modify_volume(hammer_transaction_t trans, hammer_volume_t volume,
1137                         void *base, int len);
1138 void hammer_modify_buffer(hammer_transaction_t trans, hammer_buffer_t buffer,
1139                         void *base, int len);
1140 void hammer_modify_volume_done(hammer_volume_t volume);
1141 void hammer_modify_buffer_done(hammer_buffer_t buffer);
1142
1143 int hammer_ioc_reblock(hammer_transaction_t trans, hammer_inode_t ip,
1144                         struct hammer_ioc_reblock *reblock);
1145 int hammer_ioc_prune(hammer_transaction_t trans, hammer_inode_t ip,
1146                         struct hammer_ioc_prune *prune);
1147 int hammer_ioc_mirror_read(hammer_transaction_t trans, hammer_inode_t ip,
1148                         struct hammer_ioc_mirror_rw *mirror);
1149 int hammer_ioc_mirror_write(hammer_transaction_t trans, hammer_inode_t ip,
1150                         struct hammer_ioc_mirror_rw *mirror);
1151 int hammer_ioc_set_pseudofs(hammer_transaction_t trans, hammer_inode_t ip,
1152                         struct ucred *cred, struct hammer_ioc_pseudofs_rw *pfs);
1153 int hammer_ioc_get_pseudofs(hammer_transaction_t trans, hammer_inode_t ip,
1154                         struct hammer_ioc_pseudofs_rw *pfs);
1155 int hammer_ioc_destroy_pseudofs(hammer_transaction_t trans, hammer_inode_t ip,
1156                         struct hammer_ioc_pseudofs_rw *pfs);
1157 int hammer_ioc_downgrade_pseudofs(hammer_transaction_t trans, hammer_inode_t ip,
1158                         struct hammer_ioc_pseudofs_rw *pfs);
1159 int hammer_ioc_upgrade_pseudofs(hammer_transaction_t trans, hammer_inode_t ip,
1160                         struct hammer_ioc_pseudofs_rw *pfs);
1161 int hammer_ioc_wait_pseudofs(hammer_transaction_t trans, hammer_inode_t ip,
1162                         struct hammer_ioc_pseudofs_rw *pfs);
1163
1164 int hammer_signal_check(hammer_mount_t hmp);
1165
1166 void hammer_flusher_create(hammer_mount_t hmp);
1167 void hammer_flusher_destroy(hammer_mount_t hmp);
1168 void hammer_flusher_sync(hammer_mount_t hmp);
1169 int  hammer_flusher_async(hammer_mount_t hmp, hammer_flush_group_t flg);
1170 int  hammer_flusher_async_one(hammer_mount_t hmp);
1171 void hammer_flusher_wait(hammer_mount_t hmp, int seq);
1172 int  hammer_flusher_meta_limit(hammer_mount_t hmp);
1173 int  hammer_flusher_meta_halflimit(hammer_mount_t hmp);
1174 int  hammer_flusher_undo_exhausted(hammer_transaction_t trans, int quarter);
1175 void hammer_flusher_clean_loose_ios(hammer_mount_t hmp);
1176 void hammer_flusher_finalize(hammer_transaction_t trans, int final);
1177 int  hammer_flusher_haswork(hammer_mount_t hmp);
1178
1179
1180 int hammer_recover(hammer_mount_t hmp, hammer_volume_t rootvol);
1181 void hammer_recover_flush_buffers(hammer_mount_t hmp,
1182                         hammer_volume_t root_volume, int final);
1183
1184 void hammer_crc_set_blockmap(hammer_blockmap_t blockmap);
1185 void hammer_crc_set_volume(hammer_volume_ondisk_t ondisk);
1186 void hammer_crc_set_leaf(void *data, hammer_btree_leaf_elm_t leaf);
1187
1188 int hammer_crc_test_blockmap(hammer_blockmap_t blockmap);
1189 int hammer_crc_test_volume(hammer_volume_ondisk_t ondisk);
1190 int hammer_crc_test_btree(hammer_node_ondisk_t ondisk);
1191 int hammer_crc_test_leaf(void *data, hammer_btree_leaf_elm_t leaf);
1192 void hkprintf(const char *ctl, ...);
1193 udev_t hammer_fsid_to_udev(uuid_t *uuid);
1194
1195
1196 int hammer_blocksize(int64_t file_offset);
1197 int64_t hammer_blockdemarc(int64_t file_offset1, int64_t file_offset2);
1198
1199 #endif
1200
1201 static __inline void
1202 hammer_wait_mem_record(hammer_record_t record)
1203 {
1204         hammer_wait_mem_record_ident(record, "hmmwai");
1205 }
1206
1207 static __inline void
1208 hammer_lock_ex(struct hammer_lock *lock)
1209 {
1210         hammer_lock_ex_ident(lock, "hmrlck");
1211 }
1212
1213 /*
1214  * Indicate that a B-Tree node is being modified.
1215  */
1216 static __inline void
1217 hammer_modify_node_noundo(hammer_transaction_t trans, hammer_node_t node)
1218 {
1219         hammer_modify_buffer(trans, node->buffer, NULL, 0);
1220 }
1221
1222 static __inline void
1223 hammer_modify_node_all(hammer_transaction_t trans, struct hammer_node *node)
1224 {
1225         hammer_modify_buffer(trans, node->buffer,
1226                              node->ondisk, sizeof(*node->ondisk));
1227 }
1228
1229 static __inline void
1230 hammer_modify_node(hammer_transaction_t trans, hammer_node_t node,
1231                    void *base, int len)
1232 {
1233         hammer_crc_t *crcptr;
1234
1235         KKASSERT((char *)base >= (char *)node->ondisk &&
1236                  (char *)base + len <=
1237                     (char *)node->ondisk + sizeof(*node->ondisk));
1238         hammer_modify_buffer(trans, node->buffer, base, len);
1239         crcptr = &node->ondisk->crc;
1240         hammer_modify_buffer(trans, node->buffer, crcptr, sizeof(hammer_crc_t));
1241         --node->buffer->io.modify_refs; /* only want one ref */
1242 }
1243
1244 /*
1245  * Indicate that the specified modifications have been completed.
1246  *
1247  * Do not try to generate the crc here, it's very expensive to do and a
1248  * sequence of insertions or deletions can result in many calls to this
1249  * function on the same node.
1250  */
1251 static __inline void
1252 hammer_modify_node_done(hammer_node_t node)
1253 {
1254         node->flags |= HAMMER_NODE_CRCGOOD;
1255         if ((node->flags & HAMMER_NODE_NEEDSCRC) == 0) {
1256                 node->flags |= HAMMER_NODE_NEEDSCRC;
1257                 node->buffer->io.gencrc = 1;
1258                 hammer_ref_node(node);
1259         }
1260         hammer_modify_buffer_done(node->buffer);
1261 }
1262
1263 #define hammer_modify_volume_field(trans, vol, field)           \
1264         hammer_modify_volume(trans, vol, &(vol)->ondisk->field, \
1265                              sizeof((vol)->ondisk->field))
1266
1267 #define hammer_modify_node_field(trans, node, field)            \
1268         hammer_modify_node(trans, node, &(node)->ondisk->field, \
1269                              sizeof((node)->ondisk->field))
1270