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/priv.h>
  53 #include <sys/stat.h>
  54 #include <sys/globaldata.h>
  55 #include <sys/lockf.h>
  56 #include <sys/buf.h>
  57 #include <sys/queue.h>
  58 #include <sys/ktr.h>
  59 #include <sys/globaldata.h>
  60
  61 #include <sys/buf2.h>
  62 #include <sys/signal2.h>
  63 #include "hammer_disk.h"
  64 #include "hammer_mount.h"
  65 #include "hammer_ioctl.h"
  66
  67 #if defined(_KERNEL) || defined(_KERNEL_STRUCTURES)
  68
  69 MALLOC_DECLARE(M_HAMMER);
  70
  71 /*
  72  * Kernel trace
  73  */
  74 #if !defined(KTR_HAMMER)
  75 #define KTR_HAMMER      KTR_ALL
  76 #endif
  77 KTR_INFO_MASTER_EXTERN(hammer);
  78
  79 /*
  80  * Misc structures
  81  */
  82 struct hammer_mount;
  83
  84 /*
  85  * Key structure used for custom RB tree inode lookups.  This prototypes
  86  * the function hammer_ino_rb_tree_RB_LOOKUP_INFO(root, info).
  87  */
  88 typedef struct hammer_inode_info {
  89         int64_t         obj_id;         /* (key) object identifier */
  90         hammer_tid_t    obj_asof;       /* (key) snapshot transid or 0 */
  91         u_int32_t       obj_localization; /* (key) pseudo-fs */
  92         union {
  93                 struct hammer_btree_leaf_elm *leaf;
  94         } u;
  95 } *hammer_inode_info_t;
  96
  97 typedef enum hammer_transaction_type {
  98         HAMMER_TRANS_RO,
  99         HAMMER_TRANS_STD,
 100         HAMMER_TRANS_FLS
 101 } hammer_transaction_type_t;
 102
 103 /*
 104  * HAMMER Transaction tracking
 105  */
 106 struct hammer_transaction {
 107         hammer_transaction_type_t type;
 108         struct hammer_mount *hmp;
 109         hammer_tid_t    tid;
 110         u_int64_t       time;
 111         u_int32_t       time32;
 112         int             sync_lock_refs;
 113         int             flags;
 114         struct hammer_volume *rootvol;
 115 };
 116
 117 typedef struct hammer_transaction *hammer_transaction_t;
 118
 119 #define HAMMER_TRANSF_NEWINODE  0x0001
 120 #define HAMMER_TRANSF_DIDIO     0x0002
 121 #define HAMMER_TRANSF_CRCDOM    0x0004  /* EDOM on CRC error, less critical */
 122
 123 /*
 124  * HAMMER locks
 125  */
 126 struct hammer_lock {
 127         int     refs;           /* active references delay writes */
 128         int     lockcount;      /* lock count for exclusive/shared access */
 129         int     wanted;
 130         int     exwanted;       /* number of threads waiting for ex lock */
 131         struct thread *locktd;
 132 };
 133
 134 static __inline int
 135 hammer_islocked(struct hammer_lock *lock)
 136 {
 137         return(lock->lockcount != 0);
 138 }
 139
 140 static __inline int
 141 hammer_isactive(struct hammer_lock *lock)
 142 {
 143         return(lock->refs != 0);
 144 }
 145
 146 static __inline int
 147 hammer_islastref(struct hammer_lock *lock)
 148 {
 149         return(lock->refs == 1);
 150 }
 151
 152 /*
 153  * Return if we specifically own the lock exclusively.
 154  */
 155 static __inline int
 156 hammer_lock_excl_owned(struct hammer_lock *lock, thread_t td)
 157 {
 158         if (lock->lockcount > 0 && lock->locktd == td)
 159                 return(1);
 160         return(0);
 161 }
 162
 163 /*
 164  * Flush state, used by various structures
 165  */
 166 typedef enum hammer_inode_state {
 167         HAMMER_FST_IDLE,
 168         HAMMER_FST_SETUP,
 169         HAMMER_FST_FLUSH
 170 } hammer_inode_state_t;
 171
 172 TAILQ_HEAD(hammer_record_list, hammer_record);
 173
 174 /*
 175  * Pseudo-filesystem extended data tracking
 176  */
 177 struct hammer_pfs_rb_tree;
 178 struct hammer_pseudofs_inmem;
 179 RB_HEAD(hammer_pfs_rb_tree, hammer_pseudofs_inmem);
 180 RB_PROTOTYPE2(hammer_pfs_rb_tree, hammer_pseudofs_inmem, rb_node,
 181               hammer_pfs_rb_compare, u_int32_t);
 182
 183 struct hammer_pseudofs_inmem {
 184         RB_ENTRY(hammer_pseudofs_inmem) rb_node;
 185         struct hammer_lock      lock;
 186         u_int32_t               localization;
 187         hammer_tid_t            create_tid;
 188         int                     flags;
 189         udev_t                  fsid_udev;
 190         struct hammer_pseudofs_data pfsd;
 191 };
 192
 193 typedef struct hammer_pseudofs_inmem *hammer_pseudofs_inmem_t;
 194
 195 #define HAMMER_PFSM_DELETED     0x0001
 196
 197 /*
 198  * Cache object ids.  A fixed number of objid cache structures are
 199  * created to reserve object id's for newly created files in multiples
 200  * of 100,000, localized to a particular directory, and recycled as
 201  * needed.  This allows parallel create operations in different
 202  * directories to retain fairly localized object ids which in turn
 203  * improves reblocking performance and layout.
 204  */
 205 #define OBJID_CACHE_SIZE        1024
 206 #define OBJID_CACHE_BULK        100000
 207
 208 typedef struct hammer_objid_cache {
 209         TAILQ_ENTRY(hammer_objid_cache) entry;
 210         struct hammer_inode             *dip;
 211         hammer_tid_t                    next_tid;
 212         int                             count;
 213 } *hammer_objid_cache_t;
 214
 215 /*
 216  * Associate an inode with a B-Tree node to cache search start positions
 217  */
 218 typedef struct hammer_node_cache {
 219         TAILQ_ENTRY(hammer_node_cache)  entry;
 220         struct hammer_node              *node;
 221         struct hammer_inode             *ip;
 222 } *hammer_node_cache_t;
 223
 224 TAILQ_HEAD(hammer_node_cache_list, hammer_node_cache);
 225
 226 /*
 227  * Structure used to organize flush groups.  Flush groups must be
 228  * organized into chunks in order to avoid blowing out the UNDO FIFO.
 229  * Without this a 'sync' could end up flushing 50,000 inodes in a single
 230  * transaction.
 231  */
 232 struct hammer_flush_group {
 233         TAILQ_ENTRY(hammer_flush_group) flush_entry;
 234         TAILQ_HEAD(, hammer_inode)      flush_list;
 235         int                             unused01;       /* inode load */
 236         int                             total_count;    /* record load */
 237         int                             running;        /* group is running */
 238         int                             closed;
 239         int                             refs;
 240 };
 241
 242 typedef struct hammer_flush_group *hammer_flush_group_t;
 243
 244 TAILQ_HEAD(hammer_flush_group_list, hammer_flush_group);
 245
 246 /*
 247  * Structure used to represent an inode in-memory.
 248  *
 249  * The record and data associated with an inode may be out of sync with
 250  * the disk (xDIRTY flags), or not even on the disk at all (ONDISK flag
 251  * clear).
 252  *
 253  * An inode may also hold a cache of unsynchronized records, used for
 254  * database and directories only.  Unsynchronized regular file data is
 255  * stored in the buffer cache.
 256  *
 257  * NOTE: A file which is created and destroyed within the initial
 258  * synchronization period can wind up not doing any disk I/O at all.
 259  *
 260  * Finally, an inode may cache numerous disk-referencing B-Tree cursors.
 261  */
 262 struct hammer_ino_rb_tree;
 263 struct hammer_inode;
 264 RB_HEAD(hammer_ino_rb_tree, hammer_inode);
 265 RB_PROTOTYPEX(hammer_ino_rb_tree, INFO, hammer_inode, rb_node,
 266               hammer_ino_rb_compare, hammer_inode_info_t);
 267
 268 struct hammer_rec_rb_tree;
 269 struct hammer_record;
 270 RB_HEAD(hammer_rec_rb_tree, hammer_record);
 271 RB_PROTOTYPEX(hammer_rec_rb_tree, INFO, hammer_record, rb_node,
 272               hammer_rec_rb_compare, hammer_btree_leaf_elm_t);
 273
 274 TAILQ_HEAD(hammer_node_list, hammer_node);
 275
 276 struct hammer_inode {
 277         RB_ENTRY(hammer_inode)  rb_node;
 278         hammer_inode_state_t    flush_state;
 279         hammer_flush_group_t    flush_group;
 280         TAILQ_ENTRY(hammer_inode) flush_entry;
 281         struct hammer_record_list target_list;  /* target of dependant recs */
 282         int64_t                 obj_id;         /* (key) object identifier */
 283         hammer_tid_t            obj_asof;       /* (key) snapshot or 0 */
 284         u_int32_t               obj_localization; /* (key) pseudo-fs */
 285         struct hammer_mount     *hmp;
 286         hammer_objid_cache_t    objid_cache;
 287         int                     flags;
 288         int                     error;          /* flush error */
 289         int                     cursor_ip_refs; /* sanity */
 290         int                     rsv_recs;
 291         struct vnode            *vp;
 292         hammer_pseudofs_inmem_t pfsm;
 293         struct lockf            advlock;
 294         struct hammer_lock      lock;           /* sync copy interlock */
 295         off_t                   trunc_off;
 296         struct hammer_btree_leaf_elm ino_leaf;  /* in-memory cache */
 297         struct hammer_inode_data ino_data;      /* in-memory cache */
 298         struct hammer_rec_rb_tree rec_tree;     /* in-memory cache */
 299         struct hammer_node_cache cache[2];      /* search initiate cache */
 300
 301         /*
 302          * When a demark is created to synchronize an inode to
 303          * disk, certain fields are copied so the front-end VOPs
 304          * can continue to run in parallel with the synchronization
 305          * occuring in the background.
 306          */
 307         int             sync_flags;             /* to-sync flags cache */
 308         off_t           sync_trunc_off;         /* to-sync truncation */
 309         off_t           save_trunc_off;         /* write optimization */
 310         struct hammer_btree_leaf_elm sync_ino_leaf; /* to-sync cache */
 311         struct hammer_inode_data sync_ino_data; /* to-sync cache */
 312 };
 313
 314 typedef struct hammer_inode *hammer_inode_t;
 315
 316 #define VTOI(vp)        ((struct hammer_inode *)(vp)->v_data)
 317
 318 #define HAMMER_INODE_DDIRTY     0x0001  /* in-memory ino_data is dirty */
 319                                         /* (not including atime/mtime) */
 320 #define HAMMER_INODE_RSV_INODES 0x0002  /* hmp->rsv_inodes bumped */
 321 #define HAMMER_INODE_CONN_DOWN  0x0004  /* include in downward recursion */
 322 #define HAMMER_INODE_XDIRTY     0x0008  /* in-memory records */
 323 #define HAMMER_INODE_ONDISK     0x0010  /* inode is on-disk (else not yet) */
 324 #define HAMMER_INODE_FLUSH      0x0020  /* flush on last ref */
 325 #define HAMMER_INODE_DELETED    0x0080  /* inode delete (backend) */
 326 #define HAMMER_INODE_DELONDISK  0x0100  /* delete synchronized to disk */
 327 #define HAMMER_INODE_RO         0x0200  /* read-only (because of as-of) */
 328 #define HAMMER_INODE_VHELD      0x0400  /* vnode held on sync */
 329 #define HAMMER_INODE_DONDISK    0x0800  /* data records may be on disk */
 330 #define HAMMER_INODE_BUFS       0x1000  /* dirty high level bps present */
 331 #define HAMMER_INODE_REFLUSH    0x2000  /* flush on dependancy / reflush */
 332 #define HAMMER_INODE_RECLAIM    0x4000  /* trying to reclaim */
 333 #define HAMMER_INODE_FLUSHW     0x8000  /* Someone waiting for flush */
 334
 335 #define HAMMER_INODE_TRUNCATED  0x00010000
 336 #define HAMMER_INODE_DELETING   0x00020000 /* inode delete request (frontend)*/
 337 #define HAMMER_INODE_RESIGNAL   0x00040000 /* re-signal on re-flush */
 338 #define HAMMER_INODE_ATIME      0x00100000 /* in-memory atime modified */
 339 #define HAMMER_INODE_MTIME      0x00200000 /* in-memory mtime modified */
 340 #define HAMMER_INODE_WOULDBLOCK 0x00400000 /* re-issue to new flush group */
 341 #define HAMMER_INODE_DUMMY      0x00800000 /* dummy inode covering bad file */
 342
 343 #define HAMMER_INODE_MODMASK    (HAMMER_INODE_DDIRTY|                       \
 344                                  HAMMER_INODE_XDIRTY|HAMMER_INODE_BUFS|     \
 345                                  HAMMER_INODE_ATIME|HAMMER_INODE_MTIME|     \
 346                                  HAMMER_INODE_TRUNCATED|HAMMER_INODE_DELETING)
 347
 348 #define HAMMER_INODE_MODMASK_NOXDIRTY \
 349                                 (HAMMER_INODE_MODMASK & ~HAMMER_INODE_XDIRTY)
 350
 351 #define HAMMER_FLUSH_GROUP_SIZE 64
 352
 353 #define HAMMER_FLUSH_SIGNAL     0x0001
 354 #define HAMMER_FLUSH_RECURSION  0x0002
 355
 356 /*
 357  * Used by the inode reclaim code to pipeline reclaims and avoid
 358  * blowing out kernel memory or letting the flusher get too far
 359  * behind.  The reclaim wakes up when count reaches 0 or the
 360  * timer expires.
 361  */
 362 struct hammer_reclaim {
 363         TAILQ_ENTRY(hammer_reclaim) entry;
 364         int     count;
 365 };
 366
 367 #define HAMMER_RECLAIM_FLUSH    2000
 368 #define HAMMER_RECLAIM_WAIT     4000
 369
 370 /*
 371  * Structure used to represent an unsynchronized record in-memory.  These
 372  * records typically represent directory entries.  Only non-historical
 373  * records are kept in-memory.
 374  *
 375  * Records are organized as a per-inode RB-Tree.  If the inode is not
 376  * on disk then neither are any records and the in-memory record tree
 377  * represents the entire contents of the inode.  If the inode is on disk
 378  * then the on-disk B-Tree is scanned in parallel with the in-memory
 379  * RB-Tree to synthesize the current state of the file.
 380  *
 381  * Records are also used to enforce the ordering of directory create/delete
 382  * operations.  A new inode will not be flushed to disk unless its related
 383  * directory entry is also being flushed at the same time.  A directory entry
 384  * will not be removed unless its related inode is also being removed at the
 385  * same time.
 386  */
 387 typedef enum hammer_record_type {
 388         HAMMER_MEM_RECORD_GENERAL,      /* misc record */
 389         HAMMER_MEM_RECORD_INODE,        /* inode record */
 390         HAMMER_MEM_RECORD_ADD,          /* positive memory cache record */
 391         HAMMER_MEM_RECORD_DEL,          /* negative delete-on-disk record */
 392         HAMMER_MEM_RECORD_DATA          /* bulk-data record w/on-disk ref */
 393 } hammer_record_type_t;
 394
 395 struct hammer_record {
 396         RB_ENTRY(hammer_record)         rb_node;
 397         TAILQ_ENTRY(hammer_record)      target_entry;
 398         hammer_inode_state_t            flush_state;
 399         hammer_flush_group_t            flush_group;
 400         hammer_record_type_t            type;
 401         struct hammer_lock              lock;
 402         struct hammer_reserve           *resv;
 403         struct hammer_inode             *ip;
 404         struct hammer_inode             *target_ip;
 405         struct hammer_btree_leaf_elm    leaf;
 406         union hammer_data_ondisk        *data;
 407         int                             flags;
 408         hammer_off_t                    zone2_offset;   /* direct-write only */
 409 };
 410
 411 typedef struct hammer_record *hammer_record_t;
 412
 413 /*
 414  * Record flags.  Note that FE can only be set by the frontend if the
 415  * record has not been interlocked by the backend w/ BE.
 416  */
 417 #define HAMMER_RECF_ALLOCDATA           0x0001
 418 #define HAMMER_RECF_ONRBTREE            0x0002
 419 #define HAMMER_RECF_DELETED_FE          0x0004  /* deleted (frontend) */
 420 #define HAMMER_RECF_DELETED_BE          0x0008  /* deleted (backend) */
 421 #define HAMMER_RECF_COMMITTED           0x0010  /* committed to the B-Tree */
 422 #define HAMMER_RECF_INTERLOCK_BE        0x0020  /* backend interlock */
 423 #define HAMMER_RECF_WANTED              0x0040  /* wanted by the frontend */
 424 #define HAMMER_RECF_CONVERT_DELETE      0x0100  /* special case */
 425 #define HAMMER_RECF_DIRECT_IO           0x0200  /* related direct I/O running*/
 426 #define HAMMER_RECF_DIRECT_WAIT         0x0400  /* related direct I/O running*/
 427 #define HAMMER_RECF_DIRECT_INVAL        0x0800  /* buffer alias invalidation */
 428
 429 /*
 430  * hammer_delete_at_cursor() flags
 431  */
 432 #define HAMMER_DELETE_ADJUST            0x0001
 433 #define HAMMER_DELETE_DESTROY           0x0002
 434
 435 /*
 436  * In-memory structures representing on-disk structures.
 437  */
 438 struct hammer_volume;
 439 struct hammer_buffer;
 440 struct hammer_node;
 441 struct hammer_undo;
 442 struct hammer_reserve;
 443
 444 RB_HEAD(hammer_vol_rb_tree, hammer_volume);
 445 RB_HEAD(hammer_buf_rb_tree, hammer_buffer);
 446 RB_HEAD(hammer_nod_rb_tree, hammer_node);
 447 RB_HEAD(hammer_und_rb_tree, hammer_undo);
 448 RB_HEAD(hammer_res_rb_tree, hammer_reserve);
 449
 450 RB_PROTOTYPE2(hammer_vol_rb_tree, hammer_volume, rb_node,
 451               hammer_vol_rb_compare, int32_t);
 452 RB_PROTOTYPE2(hammer_buf_rb_tree, hammer_buffer, rb_node,
 453               hammer_buf_rb_compare, hammer_off_t);
 454 RB_PROTOTYPE2(hammer_nod_rb_tree, hammer_node, rb_node,
 455               hammer_nod_rb_compare, hammer_off_t);
 456 RB_PROTOTYPE2(hammer_und_rb_tree, hammer_undo, rb_node,
 457               hammer_und_rb_compare, hammer_off_t);
 458 RB_PROTOTYPE2(hammer_res_rb_tree, hammer_reserve, rb_node,
 459               hammer_res_rb_compare, hammer_off_t);
 460
 461 /*
 462  * IO management - embedded at the head of various in-memory structures
 463  *
 464  * VOLUME       - hammer_volume containing meta-data
 465  * META_BUFFER  - hammer_buffer containing meta-data
 466  * DATA_BUFFER  - hammer_buffer containing pure-data
 467  *
 468  * Dirty volume headers and dirty meta-data buffers are locked until the
 469  * flusher can sequence them out.  Dirty pure-data buffers can be written.
 470  * Clean buffers can be passively released.
 471  */
 472 typedef enum hammer_io_type {
 473         HAMMER_STRUCTURE_VOLUME,
 474         HAMMER_STRUCTURE_META_BUFFER,
 475         HAMMER_STRUCTURE_UNDO_BUFFER,
 476         HAMMER_STRUCTURE_DATA_BUFFER
 477 } hammer_io_type_t;
 478
 479 union hammer_io_structure;
 480 struct hammer_io;
 481
 482 struct worklist {
 483         LIST_ENTRY(worklist) node;
 484 };
 485
 486 TAILQ_HEAD(hammer_io_list, hammer_io);
 487 typedef struct hammer_io_list *hammer_io_list_t;
 488
 489 struct hammer_io {
 490         struct worklist         worklist;
 491         struct hammer_lock      lock;
 492         enum hammer_io_type     type;
 493         struct hammer_mount     *hmp;
 494         struct hammer_volume    *volume;
 495         TAILQ_ENTRY(hammer_io)  mod_entry; /* list entry if modified */
 496         hammer_io_list_t        mod_list;
 497         struct buf              *bp;
 498         int64_t                 offset;    /* zone-2 offset */
 499         int                     bytes;     /* buffer cache buffer size */
 500         int                     loading;   /* loading/unloading interlock */
 501         int                     modify_refs;
 502
 503         u_int           modified : 1;   /* bp's data was modified */
 504         u_int           released : 1;   /* bp released (w/ B_LOCKED set) */
 505         u_int           running : 1;    /* bp write IO in progress */
 506         u_int           waiting : 1;    /* someone is waiting on us */
 507         u_int           validated : 1;  /* ondisk has been validated */
 508         u_int           waitdep : 1;    /* flush waits for dependancies */
 509         u_int           recovered : 1;  /* has recovery ref */
 510         u_int           waitmod : 1;    /* waiting for modify_refs */
 511         u_int           reclaim : 1;    /* reclaim requested */
 512         u_int           gencrc : 1;     /* crc needs to be generated */
 513         u_int           ioerror : 1;    /* abort on io-error */
 514 };
 515
 516 typedef struct hammer_io *hammer_io_t;
 517
 518 #define HAMMER_CLUSTER_SIZE     (64 * 1024)
 519 #if HAMMER_CLUSTER_SIZE > MAXBSIZE
 520 #undef  HAMMER_CLUSTER_SIZE
 521 #define HAMMER_CLUSTER_SIZE     MAXBSIZE
 522 #endif
 523 #define HAMMER_CLUSTER_BUFS     (HAMMER_CLUSTER_SIZE / HAMMER_BUFSIZE)
 524
 525 /*
 526  * In-memory volume representing on-disk buffer
 527  */
 528 struct hammer_volume {
 529         struct hammer_io io;
 530         RB_ENTRY(hammer_volume) rb_node;
 531         struct hammer_volume_ondisk *ondisk;
 532         int32_t vol_no;
 533         int64_t nblocks;        /* note: special calculation for statfs */
 534         int64_t buffer_base;    /* base offset of buffer 0 */
 535         hammer_off_t maxbuf_off; /* Maximum buffer offset (zone-2) */
 536         hammer_off_t maxraw_off; /* Maximum raw offset for device */
 537         char    *vol_name;
 538         struct vnode *devvp;
 539         int     vol_flags;
 540 };
 541
 542 typedef struct hammer_volume *hammer_volume_t;
 543
 544 /*
 545  * In-memory buffer (other then volume, super-cluster, or cluster),
 546  * representing an on-disk buffer.
 547  */
 548 struct hammer_buffer {
 549         struct hammer_io io;
 550         RB_ENTRY(hammer_buffer) rb_node;
 551         void *ondisk;
 552         hammer_off_t zoneX_offset;
 553         hammer_off_t zone2_offset;
 554         struct hammer_reserve *resv;
 555         struct hammer_node_list clist;
 556 };
 557
 558 typedef struct hammer_buffer *hammer_buffer_t;
 559
 560 /*
 561  * In-memory B-Tree node, representing an on-disk B-Tree node.
 562  *
 563  * This is a hang-on structure which is backed by a hammer_buffer,
 564  * indexed by a hammer_cluster, and used for fine-grained locking of
 565  * B-Tree nodes in order to properly control lock ordering.  A hammer_buffer
 566  * can contain multiple nodes representing wildly disassociated portions
 567  * of the B-Tree so locking cannot be done on a buffer-by-buffer basis.
 568  *
 569  * This structure uses a cluster-relative index to reduce the number
 570  * of layers required to access it, and also because all on-disk B-Tree
 571  * references are cluster-relative offsets.
 572  */
 573 struct hammer_node {
 574         struct hammer_lock      lock;           /* node-by-node lock */
 575         TAILQ_ENTRY(hammer_node) entry;         /* per-buffer linkage */
 576         RB_ENTRY(hammer_node)   rb_node;        /* per-cluster linkage */
 577         hammer_off_t            node_offset;    /* full offset spec */
 578         struct hammer_mount     *hmp;
 579         struct hammer_buffer    *buffer;        /* backing buffer */
 580         hammer_node_ondisk_t    ondisk;         /* ptr to on-disk structure */
 581         TAILQ_HEAD(, hammer_cursor) cursor_list;  /* deadlock recovery */
 582         struct hammer_node_cache_list cache_list; /* passive caches */
 583         int                     flags;
 584         int                     loading;        /* load interlock */
 585 };
 586
 587 #define HAMMER_NODE_DELETED     0x0001
 588 #define HAMMER_NODE_FLUSH       0x0002
 589 #define HAMMER_NODE_CRCGOOD     0x0004
 590 #define HAMMER_NODE_NEEDSCRC    0x0008
 591 #define HAMMER_NODE_NEEDSMIRROR 0x0010
 592 #define HAMMER_NODE_CRCBAD      0x0020
 593
 594 #define HAMMER_NODE_CRCANY      (HAMMER_NODE_CRCGOOD | HAMMER_NODE_CRCBAD)
 595
 596 typedef struct hammer_node      *hammer_node_t;
 597
 598 /*
 599  * List of locked nodes.  This structure is used to lock potentially large
 600  * numbers of nodes as an aid for complex B-Tree operations.
 601  */
 602 struct hammer_node_lock;
 603 TAILQ_HEAD(hammer_node_lock_list, hammer_node_lock);
 604
 605 struct hammer_node_lock {
 606         TAILQ_ENTRY(hammer_node_lock) entry;
 607         struct hammer_node_lock_list  list;
 608         struct hammer_node_lock       *parent;
 609         hammer_node_t   node;
 610         hammer_node_ondisk_t copy;      /* copy of on-disk data */
 611         int             index;          /* index of this node in parent */
 612         int             count;          /* count children */
 613         int             flags;
 614 };
 615
 616 typedef struct hammer_node_lock *hammer_node_lock_t;
 617
 618 #define HAMMER_NODE_LOCK_UPDATED        0x0001
 619
 620 /*
 621  * Common I/O management structure - embedded in in-memory structures
 622  * which are backed by filesystem buffers.
 623  */
 624 union hammer_io_structure {
 625         struct hammer_io        io;
 626         struct hammer_volume    volume;
 627         struct hammer_buffer    buffer;
 628 };
 629
 630 typedef union hammer_io_structure *hammer_io_structure_t;
 631
 632 /*
 633  * The reserve structure prevents the blockmap from allocating
 634  * out of a reserved bigblock.  Such reservations are used by
 635  * the direct-write mechanism.
 636  *
 637  * The structure is also used to hold off on reallocations of
 638  * big blocks from the freemap until flush dependancies have
 639  * been dealt with.
 640  */
 641 struct hammer_reserve {
 642         RB_ENTRY(hammer_reserve) rb_node;
 643         TAILQ_ENTRY(hammer_reserve) delay_entry;
 644         int             flush_group;
 645         int             flags;
 646         int             refs;
 647         int             zone;
 648         int             append_off;
 649         hammer_off_t    zone_offset;
 650 };
 651
 652 typedef struct hammer_reserve *hammer_reserve_t;
 653
 654 #define HAMMER_RESF_ONDELAY     0x0001
 655 #define HAMMER_RESF_LAYER2FREE  0x0002
 656
 657 #include "hammer_cursor.h"
 658
 659 /*
 660  * The undo structure tracks recent undos to avoid laying down duplicate
 661  * undos within a flush group, saving us a significant amount of overhead.
 662  *
 663  * This is strictly a heuristic.
 664  */
 665 #define HAMMER_MAX_UNDOS                1024
 666 #define HAMMER_MAX_FLUSHERS             4
 667
 668 struct hammer_undo {
 669         RB_ENTRY(hammer_undo)   rb_node;
 670         TAILQ_ENTRY(hammer_undo) lru_entry;
 671         hammer_off_t            offset;
 672         int                     bytes;
 673 };
 674
 675 typedef struct hammer_undo *hammer_undo_t;
 676
 677 struct hammer_flusher_info;
 678 TAILQ_HEAD(hammer_flusher_info_list, hammer_flusher_info);
 679
 680 struct hammer_flusher {
 681         int             signal;         /* flusher thread sequencer */
 682         int             act;            /* currently active flush group */
 683         int             done;           /* set to act when complete */
 684         int             next;           /* next flush group */
 685         int             group_lock;     /* lock sequencing of the next flush */
 686         int             exiting;        /* request master exit */
 687         thread_t        td;             /* master flusher thread */
 688         hammer_tid_t    tid;            /* last flushed transaction id */
 689         int             finalize_want;          /* serialize finalization */
 690         struct hammer_lock finalize_lock;       /* serialize finalization */
 691         struct hammer_transaction trans;        /* shared transaction */
 692         struct hammer_flusher_info_list run_list;
 693         struct hammer_flusher_info_list ready_list;
 694 };
 695
 696 /*
 697  * Internal hammer mount data structure
 698  */
 699 struct hammer_mount {
 700         struct mount *mp;
 701         /*struct vnode *rootvp;*/
 702         struct hammer_ino_rb_tree rb_inos_root;
 703         struct hammer_vol_rb_tree rb_vols_root;
 704         struct hammer_nod_rb_tree rb_nods_root;
 705         struct hammer_und_rb_tree rb_undo_root;
 706         struct hammer_res_rb_tree rb_resv_root;
 707         struct hammer_buf_rb_tree rb_bufs_root;
 708         struct hammer_pfs_rb_tree rb_pfsm_root;
 709         struct hammer_volume *rootvol;
 710         struct hammer_base_elm root_btree_beg;
 711         struct hammer_base_elm root_btree_end;
 712
 713         struct malloc_type      *m_misc;
 714         struct malloc_type      *m_inodes;
 715
 716         int     flags;          /* HAMMER_MOUNT_xxx flags */
 717         int     hflags;
 718         int     ronly;
 719         int     nvolumes;
 720         int     volume_iterator;
 721         int     master_id;      /* -1 or 0-15 - clustering and mirroring */
 722         int     version;        /* hammer filesystem version to use */
 723         int     rsv_inodes;     /* reserved space due to dirty inodes */
 724         int64_t rsv_databytes;  /* reserved space due to record data */
 725         int     rsv_recs;       /* reserved space due to dirty records */
 726         int     rsv_fromdelay;  /* bigblocks reserved due to flush delay */
 727         int     undo_rec_limit; /* based on size of undo area */
 728         int     last_newrecords;
 729         int     count_newrecords;
 730
 731         int     inode_reclaims; /* inodes pending reclaim by flusher */
 732         int     count_inodes;   /* total number of inodes */
 733         int     count_iqueued;  /* inodes queued to flusher */
 734
 735         struct hammer_flusher flusher;
 736
 737         u_int   check_interrupt;
 738         uuid_t  fsid;
 739         struct hammer_io_list volu_list;        /* dirty undo buffers */
 740         struct hammer_io_list undo_list;        /* dirty undo buffers */
 741         struct hammer_io_list data_list;        /* dirty data buffers */
 742         struct hammer_io_list alt_data_list;    /* dirty data buffers */
 743         struct hammer_io_list meta_list;        /* dirty meta bufs    */
 744         struct hammer_io_list lose_list;        /* loose buffers      */
 745         int     locked_dirty_space;             /* meta/volu count    */
 746         int     io_running_space;
 747         int     objid_cache_count;
 748         int     error;                          /* critical I/O error */
 749         struct krate    krate;                  /* rate limited kprintf */
 750         hammer_tid_t    asof;                   /* snapshot mount */
 751         hammer_tid_t    next_tid;
 752         hammer_tid_t    flush_tid1;             /* flusher tid sequencing */
 753         hammer_tid_t    flush_tid2;             /* flusher tid sequencing */
 754         int64_t copy_stat_freebigblocks;        /* number of free bigblocks */
 755
 756         struct netexport export;
 757         struct hammer_lock sync_lock;
 758         struct hammer_lock free_lock;
 759         struct hammer_lock undo_lock;
 760         struct hammer_lock blkmap_lock;
 761         struct hammer_blockmap  blockmap[HAMMER_MAX_ZONES];
 762         struct hammer_undo      undos[HAMMER_MAX_UNDOS];
 763         int                     undo_alloc;
 764         TAILQ_HEAD(, hammer_undo)  undo_lru_list;
 765         TAILQ_HEAD(, hammer_reserve) delay_list;
 766         struct hammer_flush_group_list  flush_group_list;
 767         hammer_flush_group_t    next_flush_group;
 768         TAILQ_HEAD(, hammer_objid_cache) objid_cache_list;
 769         TAILQ_HEAD(, hammer_reclaim) reclaim_list;
 770 };
 771
 772 typedef struct hammer_mount     *hammer_mount_t;
 773
 774 #define HAMMER_MOUNT_CRITICAL_ERROR     0x0001
 775 #define HAMMER_MOUNT_FLUSH_RECOVERY     0x0002
 776
 777 struct hammer_sync_info {
 778         int error;
 779         int waitfor;
 780 };
 781
 782 #endif
 783
 784 /*
 785  * checkspace slop (8MB chunks), higher numbers are more conservative.
 786  */
 787 #define HAMMER_CHKSPC_REBLOCK   25
 788 #define HAMMER_CHKSPC_MIRROR    20
 789 #define HAMMER_CHKSPC_WRITE     20
 790 #define HAMMER_CHKSPC_CREATE    20
 791 #define HAMMER_CHKSPC_REMOVE    10
 792 #define HAMMER_CHKSPC_EMERGENCY 0
 793
 794 #if defined(_KERNEL)
 795
 796 extern struct vop_ops hammer_vnode_vops;
 797 extern struct vop_ops hammer_spec_vops;
 798 extern struct vop_ops hammer_fifo_vops;
 799 extern struct bio_ops hammer_bioops;
 800
 801 extern int hammer_debug_io;
 802 extern int hammer_debug_general;
 803 extern int hammer_debug_debug;
 804 extern int hammer_debug_inode;
 805 extern int hammer_debug_locks;
 806 extern int hammer_debug_btree;
 807 extern int hammer_debug_tid;
 808 extern int hammer_debug_recover;
 809 extern int hammer_debug_recover_faults;
 810 extern int hammer_cluster_enable;
 811 extern int hammer_count_fsyncs;
 812 extern int hammer_count_inodes;
 813 extern int hammer_count_iqueued;
 814 extern int hammer_count_reclaiming;
 815 extern int hammer_count_records;
 816 extern int hammer_count_record_datas;
 817 extern int hammer_count_volumes;
 818 extern int hammer_count_buffers;
 819 extern int hammer_count_nodes;
 820 extern int64_t hammer_count_extra_space_used;
 821 extern int64_t hammer_stats_btree_lookups;
 822 extern int64_t hammer_stats_btree_searches;
 823 extern int64_t hammer_stats_btree_inserts;
 824 extern int64_t hammer_stats_btree_deletes;
 825 extern int64_t hammer_stats_btree_elements;
 826 extern int64_t hammer_stats_btree_splits;
 827 extern int64_t hammer_stats_btree_iterations;
 828 extern int64_t hammer_stats_record_iterations;
 829 extern int64_t hammer_stats_file_read;
 830 extern int64_t hammer_stats_file_write;
 831 extern int64_t hammer_stats_file_iopsr;
 832 extern int64_t hammer_stats_file_iopsw;
 833 extern int64_t hammer_stats_disk_read;
 834 extern int64_t hammer_stats_disk_write;
 835 extern int64_t hammer_stats_inode_flushes;
 836 extern int64_t hammer_stats_commits;
 837 extern int hammer_count_dirtybufspace;
 838 extern int hammer_count_refedbufs;
 839 extern int hammer_count_reservations;
 840 extern int hammer_count_io_running_read;
 841 extern int hammer_count_io_running_write;
 842 extern int hammer_count_io_locked;
 843 extern int hammer_limit_dirtybufspace;
 844 extern int hammer_limit_recs;
 845 extern int hammer_bio_count;
 846 extern int hammer_verify_zone;
 847 extern int hammer_verify_data;
 848 extern int hammer_write_mode;
 849 extern int hammer_autoflush;
 850 extern int64_t hammer_contention_count;
 851
 852 void    hammer_critical_error(hammer_mount_t hmp, hammer_inode_t ip,
 853                         int error, const char *msg);
 854 int     hammer_vop_inactive(struct vop_inactive_args *);
 855 int     hammer_vop_reclaim(struct vop_reclaim_args *);
 856 int     hammer_get_vnode(struct hammer_inode *ip, struct vnode **vpp);
 857 struct hammer_inode *hammer_get_inode(hammer_transaction_t trans,
 858                         hammer_inode_t dip, int64_t obj_id,
 859                         hammer_tid_t asof, u_int32_t localization,
 860                         int flags, int *errorp);
 861 struct hammer_inode *hammer_get_dummy_inode(hammer_transaction_t trans,
 862                         hammer_inode_t dip, int64_t obj_id,
 863                         hammer_tid_t asof, u_int32_t localization,
 864                         int flags, int *errorp);
 865 void    hammer_scan_inode_snapshots(hammer_mount_t hmp,
 866                         hammer_inode_info_t iinfo,
 867                         int (*callback)(hammer_inode_t ip, void *data),
 868                         void *data);
 869 void    hammer_put_inode(struct hammer_inode *ip);
 870 void    hammer_put_inode_ref(struct hammer_inode *ip);
 871 void    hammer_inode_waitreclaims(hammer_mount_t hmp);
 872 void    hammer_inode_waithard(hammer_mount_t hmp);
 873
 874 int     hammer_unload_volume(hammer_volume_t volume, void *data __unused);
 875 int     hammer_adjust_volume_mode(hammer_volume_t volume, void *data __unused);
 876
 877 int     hammer_unload_buffer(hammer_buffer_t buffer, void *data __unused);
 878 int     hammer_install_volume(hammer_mount_t hmp, const char *volname,
 879                         struct vnode *devvp);
 880 int     hammer_mountcheck_volumes(hammer_mount_t hmp);
 881
 882 int     hammer_mem_add(hammer_record_t record);
 883 int     hammer_ip_lookup(hammer_cursor_t cursor);
 884 int     hammer_ip_first(hammer_cursor_t cursor);
 885 int     hammer_ip_next(hammer_cursor_t cursor);
 886 int     hammer_ip_resolve_data(hammer_cursor_t cursor);
 887 int     hammer_ip_delete_record(hammer_cursor_t cursor, hammer_inode_t ip,
 888                         hammer_tid_t tid);
 889 int     hammer_delete_at_cursor(hammer_cursor_t cursor, int delete_flags,
 890                         hammer_tid_t delete_tid, u_int32_t delete_ts,
 891                         int track, int64_t *stat_bytes);
 892 int     hammer_ip_check_directory_empty(hammer_transaction_t trans,
 893                         hammer_inode_t ip);
 894 int     hammer_sync_hmp(hammer_mount_t hmp, int waitfor);
 895 int     hammer_queue_inodes_flusher(hammer_mount_t hmp, int waitfor);
 896
 897 hammer_record_t
 898         hammer_alloc_mem_record(hammer_inode_t ip, int data_len);
 899 void    hammer_flush_record_done(hammer_record_t record, int error);
 900 void    hammer_wait_mem_record_ident(hammer_record_t record, const char *ident);
 901 void    hammer_rel_mem_record(hammer_record_t record);
 902
 903 int     hammer_cursor_up(hammer_cursor_t cursor);
 904 int     hammer_cursor_up_locked(hammer_cursor_t cursor);
 905 int     hammer_cursor_down(hammer_cursor_t cursor);
 906 int     hammer_cursor_upgrade(hammer_cursor_t cursor);
 907 int     hammer_cursor_upgrade_node(hammer_cursor_t cursor);
 908 void    hammer_cursor_downgrade(hammer_cursor_t cursor);
 909 int     hammer_cursor_seek(hammer_cursor_t cursor, hammer_node_t node,
 910                         int index);
 911 void    hammer_lock_ex_ident(struct hammer_lock *lock, const char *ident);
 912 int     hammer_lock_ex_try(struct hammer_lock *lock);
 913 void    hammer_lock_sh(struct hammer_lock *lock);
 914 int     hammer_lock_sh_try(struct hammer_lock *lock);
 915 int     hammer_lock_upgrade(struct hammer_lock *lock);
 916 void    hammer_lock_downgrade(struct hammer_lock *lock);
 917 int     hammer_lock_status(struct hammer_lock *lock);
 918 void    hammer_unlock(struct hammer_lock *lock);
 919 void    hammer_ref(struct hammer_lock *lock);
 920 void    hammer_unref(struct hammer_lock *lock);
 921
 922 void    hammer_sync_lock_ex(hammer_transaction_t trans);
 923 void    hammer_sync_lock_sh(hammer_transaction_t trans);
 924 int     hammer_sync_lock_sh_try(hammer_transaction_t trans);
 925 void    hammer_sync_unlock(hammer_transaction_t trans);
 926
 927 u_int32_t hammer_to_unix_xid(uuid_t *uuid);
 928 void hammer_guid_to_uuid(uuid_t *uuid, u_int32_t guid);
 929 void    hammer_time_to_timespec(u_int64_t xtime, struct timespec *ts);
 930 u_int64_t hammer_timespec_to_time(struct timespec *ts);
 931 int     hammer_str_to_tid(const char *str, int *ispfsp,
 932                         hammer_tid_t *tidp, u_int32_t *localizationp);
 933 int     hammer_is_atatext(const char *name, int len);
 934 hammer_tid_t hammer_alloc_objid(hammer_mount_t hmp, hammer_inode_t dip);
 935 void hammer_clear_objid(hammer_inode_t dip);
 936 void hammer_destroy_objid_cache(hammer_mount_t hmp);
 937
 938 int hammer_enter_undo_history(hammer_mount_t hmp, hammer_off_t offset,
 939                         int bytes);
 940 void hammer_clear_undo_history(hammer_mount_t hmp);
 941 enum vtype hammer_get_vnode_type(u_int8_t obj_type);
 942 int hammer_get_dtype(u_int8_t obj_type);
 943 u_int8_t hammer_get_obj_type(enum vtype vtype);
 944 int64_t hammer_directory_namekey(hammer_inode_t dip, const void *name, int len,
 945                         u_int32_t *max_iterationsp);
 946 int     hammer_nohistory(hammer_inode_t ip);
 947
 948 int     hammer_init_cursor(hammer_transaction_t trans, hammer_cursor_t cursor,
 949                         hammer_node_cache_t cache, hammer_inode_t ip);
 950 void    hammer_normalize_cursor(hammer_cursor_t cursor);
 951 void    hammer_done_cursor(hammer_cursor_t cursor);
 952 int     hammer_recover_cursor(hammer_cursor_t cursor);
 953 void    hammer_unlock_cursor(hammer_cursor_t cursor);
 954 int     hammer_lock_cursor(hammer_cursor_t cursor);
 955 hammer_cursor_t hammer_push_cursor(hammer_cursor_t ocursor);
 956 void    hammer_pop_cursor(hammer_cursor_t ocursor, hammer_cursor_t ncursor);
 957
 958 void    hammer_cursor_replaced_node(hammer_node_t onode, hammer_node_t nnode);
 959 void    hammer_cursor_removed_node(hammer_node_t onode, hammer_node_t parent,
 960                         int index);
 961 void    hammer_cursor_split_node(hammer_node_t onode, hammer_node_t nnode,
 962                         int index);
 963 void    hammer_cursor_moved_element(hammer_node_t onode, hammer_node_t nnode,
 964                         int oindex, int nindex);
 965 void    hammer_cursor_parent_changed(hammer_node_t node, hammer_node_t oparent,
 966                         hammer_node_t nparent, int nindex);
 967 void    hammer_cursor_inserted_element(hammer_node_t node, int index);
 968 void    hammer_cursor_deleted_element(hammer_node_t node, int index);
 969
 970 int     hammer_btree_lookup(hammer_cursor_t cursor);
 971 int     hammer_btree_first(hammer_cursor_t cursor);
 972 int     hammer_btree_last(hammer_cursor_t cursor);
 973 int     hammer_btree_extract(hammer_cursor_t cursor, int flags);
 974 int     hammer_btree_iterate(hammer_cursor_t cursor);
 975 int     hammer_btree_iterate_reverse(hammer_cursor_t cursor);
 976 int     hammer_btree_insert(hammer_cursor_t cursor,
 977                             hammer_btree_leaf_elm_t elm, int *doprop);
 978 int     hammer_btree_delete(hammer_cursor_t cursor);
 979 void    hammer_btree_do_propagation(hammer_cursor_t cursor,
 980                             hammer_pseudofs_inmem_t pfsm,
 981                             hammer_btree_leaf_elm_t leaf);
 982 int     hammer_btree_cmp(hammer_base_elm_t key1, hammer_base_elm_t key2);
 983 int     hammer_btree_chkts(hammer_tid_t ts, hammer_base_elm_t key);
 984 int     hammer_btree_correct_rhb(hammer_cursor_t cursor, hammer_tid_t tid);
 985 int     hammer_btree_correct_lhb(hammer_cursor_t cursor, hammer_tid_t tid);
 986
 987 int     btree_set_parent(hammer_transaction_t trans, hammer_node_t node,
 988                         hammer_btree_elm_t elm);
 989 void    hammer_node_lock_init(hammer_node_lock_t parent, hammer_node_t node);
 990 int     hammer_btree_lock_children(hammer_cursor_t cursor, int depth,
 991                         hammer_node_lock_t parent);
 992 void    hammer_btree_lock_copy(hammer_cursor_t cursor,
 993                         hammer_node_lock_t parent);
 994 int     hammer_btree_sync_copy(hammer_cursor_t cursor,
 995                         hammer_node_lock_t parent);
 996 void    hammer_btree_unlock_children(hammer_cursor_t cursor,
 997                         hammer_node_lock_t parent);
 998 int     hammer_btree_search_node(hammer_base_elm_t elm, hammer_node_ondisk_t node);
 999 hammer_node_t hammer_btree_get_parent(hammer_transaction_t trans,
1000                         hammer_node_t node, int *parent_indexp,
1001                         int *errorp, int try_exclusive);
1002
1003 void    hammer_print_btree_node(hammer_node_ondisk_t ondisk);
1004 void    hammer_print_btree_elm(hammer_btree_elm_t elm, u_int8_t type, int i);
1005
1006 void    *hammer_bread(struct hammer_mount *hmp, hammer_off_t off,
1007                         int *errorp, struct hammer_buffer **bufferp);
1008 void    *hammer_bnew(struct hammer_mount *hmp, hammer_off_t off,
1009                         int *errorp, struct hammer_buffer **bufferp);
1010 void    *hammer_bread_ext(struct hammer_mount *hmp, hammer_off_t off, int bytes,
1011                         int *errorp, struct hammer_buffer **bufferp);
1012 void    *hammer_bnew_ext(struct hammer_mount *hmp, hammer_off_t off, int bytes,
1013                         int *errorp, struct hammer_buffer **bufferp);
1014
1015 hammer_volume_t hammer_get_root_volume(hammer_mount_t hmp, int *errorp);
1016
1017 hammer_volume_t hammer_get_volume(hammer_mount_t hmp,
1018                         int32_t vol_no, int *errorp);
1019 hammer_buffer_t hammer_get_buffer(hammer_mount_t hmp, hammer_off_t buf_offset,
1020                         int bytes, int isnew, int *errorp);
1021 void            hammer_sync_buffers(hammer_mount_t hmp,
1022                         hammer_off_t base_offset, int bytes);
1023 int             hammer_del_buffers(hammer_mount_t hmp,
1024                         hammer_off_t base_offset,
1025                         hammer_off_t zone2_offset, int bytes,
1026                         int report_conflicts);
1027
1028 int             hammer_ref_volume(hammer_volume_t volume);
1029 int             hammer_ref_buffer(hammer_buffer_t buffer);
1030 void            hammer_flush_buffer_nodes(hammer_buffer_t buffer);
1031
1032 void            hammer_rel_volume(hammer_volume_t volume, int flush);
1033 void            hammer_rel_buffer(hammer_buffer_t buffer, int flush);
1034
1035 int             hammer_vfs_export(struct mount *mp, int op,
1036                         const struct export_args *export);
1037 hammer_node_t   hammer_get_node(hammer_transaction_t trans,
1038                         hammer_off_t node_offset, int isnew, int *errorp);
1039 void            hammer_ref_node(hammer_node_t node);
1040 hammer_node_t   hammer_ref_node_safe(hammer_transaction_t trans,
1041                         hammer_node_cache_t cache, int *errorp);
1042 void            hammer_rel_node(hammer_node_t node);
1043 void            hammer_delete_node(hammer_transaction_t trans,
1044                         hammer_node_t node);
1045 void            hammer_cache_node(hammer_node_cache_t cache,
1046                         hammer_node_t node);
1047 void            hammer_uncache_node(hammer_node_cache_t cache);
1048 void            hammer_flush_node(hammer_node_t node);
1049
1050 void hammer_dup_buffer(struct hammer_buffer **bufferp,
1051                         struct hammer_buffer *buffer);
1052 hammer_node_t hammer_alloc_btree(hammer_transaction_t trans, int *errorp);
1053 void *hammer_alloc_data(hammer_transaction_t trans, int32_t data_len,
1054                         u_int16_t rec_type, hammer_off_t *data_offsetp,
1055                         struct hammer_buffer **data_bufferp, int *errorp);
1056
1057 int hammer_generate_undo(hammer_transaction_t trans, hammer_io_t io,
1058                         hammer_off_t zone1_offset, void *base, int len);
1059
1060 void hammer_put_volume(struct hammer_volume *volume, int flush);
1061 void hammer_put_buffer(struct hammer_buffer *buffer, int flush);
1062
1063 hammer_off_t hammer_freemap_alloc(hammer_transaction_t trans,
1064                         hammer_off_t owner, int *errorp);
1065 void hammer_freemap_free(hammer_transaction_t trans, hammer_off_t phys_offset,
1066                         hammer_off_t owner, int *errorp);
1067 int hammer_checkspace(hammer_mount_t hmp, int slop);
1068 hammer_off_t hammer_blockmap_alloc(hammer_transaction_t trans, int zone,
1069                         int bytes, int *errorp);
1070 hammer_reserve_t hammer_blockmap_reserve(hammer_mount_t hmp, int zone,
1071                         int bytes, hammer_off_t *zone_offp, int *errorp);
1072 void hammer_blockmap_reserve_undo(hammer_mount_t hmp, hammer_reserve_t resv,
1073                         hammer_off_t zone_offset, int bytes);
1074 void hammer_blockmap_reserve_complete(hammer_mount_t hmp,
1075                         hammer_reserve_t resv);
1076 void hammer_reserve_clrdelay(hammer_mount_t hmp, hammer_reserve_t resv);
1077 void hammer_blockmap_free(hammer_transaction_t trans,
1078                         hammer_off_t bmap_off, int bytes);
1079 int hammer_blockmap_finalize(hammer_transaction_t trans,
1080                         hammer_reserve_t resv,
1081                         hammer_off_t bmap_off, int bytes);
1082 int hammer_blockmap_getfree(hammer_mount_t hmp, hammer_off_t bmap_off,
1083                         int *curp, int *errorp);
1084 hammer_off_t hammer_blockmap_lookup(hammer_mount_t hmp, hammer_off_t bmap_off,
1085                         int *errorp);
1086 hammer_off_t hammer_undo_lookup(hammer_mount_t hmp, hammer_off_t bmap_off,
1087                         int *errorp);
1088 int64_t hammer_undo_used(hammer_transaction_t trans);
1089 int64_t hammer_undo_space(hammer_transaction_t trans);
1090 int64_t hammer_undo_max(hammer_mount_t hmp);
1091
1092 void hammer_start_transaction(struct hammer_transaction *trans,
1093                               struct hammer_mount *hmp);
1094 void hammer_simple_transaction(struct hammer_transaction *trans,
1095                               struct hammer_mount *hmp);
1096 void hammer_start_transaction_fls(struct hammer_transaction *trans,
1097                                   struct hammer_mount *hmp);
1098 void hammer_done_transaction(struct hammer_transaction *trans);
1099
1100 void hammer_modify_inode(hammer_inode_t ip, int flags);
1101 void hammer_flush_inode(hammer_inode_t ip, int flags);
1102 void hammer_flush_inode_done(hammer_inode_t ip, int error);
1103 void hammer_wait_inode(hammer_inode_t ip);
1104
1105 int  hammer_create_inode(struct hammer_transaction *trans, struct vattr *vap,
1106                         struct ucred *cred, struct hammer_inode *dip,
1107                         hammer_pseudofs_inmem_t pfsm,
1108                         struct hammer_inode **ipp);
1109 void hammer_rel_inode(hammer_inode_t ip, int flush);
1110 int hammer_reload_inode(hammer_inode_t ip, void *arg __unused);
1111 int hammer_ino_rb_compare(hammer_inode_t ip1, hammer_inode_t ip2);
1112 int hammer_destroy_inode_callback(hammer_inode_t ip, void *data __unused);
1113
1114 int hammer_sync_inode(hammer_transaction_t trans, hammer_inode_t ip);
1115 void hammer_test_inode(hammer_inode_t dip);
1116 void hammer_inode_unloadable_check(hammer_inode_t ip, int getvp);
1117
1118 int  hammer_ip_add_directory(struct hammer_transaction *trans,
1119                         hammer_inode_t dip, const char *name, int bytes,
1120                         hammer_inode_t nip);
1121 int  hammer_ip_del_directory(struct hammer_transaction *trans,
1122                         hammer_cursor_t cursor, hammer_inode_t dip,
1123                         hammer_inode_t ip);
1124 hammer_record_t hammer_ip_add_bulk(hammer_inode_t ip, off_t file_offset,
1125                         void *data, int bytes, int *errorp);
1126 int  hammer_ip_frontend_trunc(struct hammer_inode *ip, off_t file_size);
1127 int  hammer_ip_add_record(struct hammer_transaction *trans,
1128                         hammer_record_t record);
1129 int  hammer_ip_delete_range(hammer_cursor_t cursor, hammer_inode_t ip,
1130                         int64_t ran_beg, int64_t ran_end, int truncating);
1131 int  hammer_ip_delete_clean(hammer_cursor_t cursor, hammer_inode_t ip,
1132                         int *countp);
1133 int  hammer_ip_sync_data(hammer_cursor_t cursor, hammer_inode_t ip,
1134                         int64_t offset, void *data, int bytes);
1135 int  hammer_ip_sync_record(hammer_transaction_t trans, hammer_record_t rec);
1136 int  hammer_ip_sync_record_cursor(hammer_cursor_t cursor, hammer_record_t rec);
1137 hammer_pseudofs_inmem_t  hammer_load_pseudofs(hammer_transaction_t trans,
1138                         u_int32_t localization, int *errorp);
1139 int  hammer_mkroot_pseudofs(hammer_transaction_t trans, struct ucred *cred,
1140                         hammer_pseudofs_inmem_t pfsm);
1141 int  hammer_save_pseudofs(hammer_transaction_t trans,
1142                         hammer_pseudofs_inmem_t pfsm);
1143 int  hammer_unload_pseudofs(hammer_transaction_t trans, u_int32_t localization);
1144 void hammer_rel_pseudofs(hammer_mount_t hmp, hammer_pseudofs_inmem_t pfsm);
1145 int hammer_ioctl(hammer_inode_t ip, u_long com, caddr_t data, int fflag,
1146                         struct ucred *cred);
1147
1148 void hammer_io_init(hammer_io_t io, hammer_volume_t volume,
1149                         enum hammer_io_type type);
1150 int hammer_io_read(struct vnode *devvp, struct hammer_io *io,
1151                         hammer_off_t limit);
1152 int hammer_io_new(struct vnode *devvp, struct hammer_io *io);
1153 int hammer_io_inval(hammer_volume_t volume, hammer_off_t zone2_offset);
1154 struct buf *hammer_io_release(struct hammer_io *io, int flush);
1155 void hammer_io_flush(struct hammer_io *io);
1156 void hammer_io_wait(struct hammer_io *io);
1157 void hammer_io_waitdep(struct hammer_io *io);
1158 void hammer_io_wait_all(hammer_mount_t hmp, const char *ident);
1159 int hammer_io_direct_read(hammer_mount_t hmp, struct bio *bio,
1160                         hammer_btree_leaf_elm_t leaf);
1161 int hammer_io_direct_write(hammer_mount_t hmp, hammer_record_t record,
1162                         struct bio *bio);
1163 void hammer_io_direct_wait(hammer_record_t record);
1164 void hammer_io_direct_uncache(hammer_mount_t hmp, hammer_btree_leaf_elm_t leaf);
1165 void hammer_io_write_interlock(hammer_io_t io);
1166 void hammer_io_done_interlock(hammer_io_t io);
1167 void hammer_io_clear_modify(struct hammer_io *io, int inval);
1168 void hammer_io_clear_modlist(struct hammer_io *io);
1169 void hammer_io_flush_sync(hammer_mount_t hmp);
1170
1171 void hammer_modify_volume(hammer_transaction_t trans, hammer_volume_t volume,
1172                         void *base, int len);
1173 void hammer_modify_buffer(hammer_transaction_t trans, hammer_buffer_t buffer,
1174                         void *base, int len);
1175 void hammer_modify_volume_done(hammer_volume_t volume);
1176 void hammer_modify_buffer_done(hammer_buffer_t buffer);
1177
1178 int hammer_ioc_reblock(hammer_transaction_t trans, hammer_inode_t ip,
1179                         struct hammer_ioc_reblock *reblock);
1180 int hammer_ioc_rebalance(hammer_transaction_t trans, hammer_inode_t ip,
1181                         struct hammer_ioc_rebalance *rebal);
1182 int hammer_ioc_prune(hammer_transaction_t trans, hammer_inode_t ip,
1183                         struct hammer_ioc_prune *prune);
1184 int hammer_ioc_mirror_read(hammer_transaction_t trans, hammer_inode_t ip,
1185                         struct hammer_ioc_mirror_rw *mirror);
1186 int hammer_ioc_mirror_write(hammer_transaction_t trans, hammer_inode_t ip,
1187                         struct hammer_ioc_mirror_rw *mirror);
1188 int hammer_ioc_set_pseudofs(hammer_transaction_t trans, hammer_inode_t ip,
1189                         struct ucred *cred, struct hammer_ioc_pseudofs_rw *pfs);
1190 int hammer_ioc_get_pseudofs(hammer_transaction_t trans, hammer_inode_t ip,
1191                         struct hammer_ioc_pseudofs_rw *pfs);
1192 int hammer_ioc_destroy_pseudofs(hammer_transaction_t trans, hammer_inode_t ip,
1193                         struct hammer_ioc_pseudofs_rw *pfs);
1194 int hammer_ioc_downgrade_pseudofs(hammer_transaction_t trans, hammer_inode_t ip,
1195                         struct hammer_ioc_pseudofs_rw *pfs);
1196 int hammer_ioc_upgrade_pseudofs(hammer_transaction_t trans, hammer_inode_t ip,
1197                         struct hammer_ioc_pseudofs_rw *pfs);
1198 int hammer_ioc_wait_pseudofs(hammer_transaction_t trans, hammer_inode_t ip,
1199                         struct hammer_ioc_pseudofs_rw *pfs);
1200
1201 int hammer_signal_check(hammer_mount_t hmp);
1202
1203 void hammer_flusher_create(hammer_mount_t hmp);
1204 void hammer_flusher_destroy(hammer_mount_t hmp);
1205 void hammer_flusher_sync(hammer_mount_t hmp);
1206 int  hammer_flusher_async(hammer_mount_t hmp, hammer_flush_group_t flg);
1207 int  hammer_flusher_async_one(hammer_mount_t hmp);
1208 void hammer_flusher_wait(hammer_mount_t hmp, int seq);
1209 void hammer_flusher_wait_next(hammer_mount_t hmp);
1210 int  hammer_flusher_meta_limit(hammer_mount_t hmp);
1211 int  hammer_flusher_meta_halflimit(hammer_mount_t hmp);
1212 int  hammer_flusher_undo_exhausted(hammer_transaction_t trans, int quarter);
1213 void hammer_flusher_clean_loose_ios(hammer_mount_t hmp);
1214 void hammer_flusher_finalize(hammer_transaction_t trans, int final);
1215 int  hammer_flusher_haswork(hammer_mount_t hmp);
1216
1217
1218 int hammer_recover(hammer_mount_t hmp, hammer_volume_t rootvol);
1219 void hammer_recover_flush_buffers(hammer_mount_t hmp,
1220                         hammer_volume_t root_volume, int final);
1221
1222 void hammer_crc_set_blockmap(hammer_blockmap_t blockmap);
1223 void hammer_crc_set_volume(hammer_volume_ondisk_t ondisk);
1224 void hammer_crc_set_leaf(void *data, hammer_btree_leaf_elm_t leaf);
1225
1226 int hammer_crc_test_blockmap(hammer_blockmap_t blockmap);
1227 int hammer_crc_test_volume(hammer_volume_ondisk_t ondisk);
1228 int hammer_crc_test_btree(hammer_node_ondisk_t ondisk);
1229 int hammer_crc_test_leaf(void *data, hammer_btree_leaf_elm_t leaf);
1230 void hkprintf(const char *ctl, ...);
1231 udev_t hammer_fsid_to_udev(uuid_t *uuid);
1232
1233
1234 int hammer_blocksize(int64_t file_offset);
1235 int64_t hammer_blockdemarc(int64_t file_offset1, int64_t file_offset2);
1236
1237 #endif
1238
1239 static __inline void
1240 hammer_wait_mem_record(hammer_record_t record)
1241 {
1242         hammer_wait_mem_record_ident(record, "hmmwai");
1243 }
1244
1245 static __inline void
1246 hammer_lock_ex(struct hammer_lock *lock)
1247 {
1248         hammer_lock_ex_ident(lock, "hmrlck");
1249 }
1250
1251 /*
1252  * Indicate that a B-Tree node is being modified.
1253  */
1254 static __inline void
1255 hammer_modify_node_noundo(hammer_transaction_t trans, hammer_node_t node)
1256 {
1257         KKASSERT((node->flags & HAMMER_NODE_CRCBAD) == 0);
1258         hammer_modify_buffer(trans, node->buffer, NULL, 0);
1259 }
1260
1261 static __inline void
1262 hammer_modify_node_all(hammer_transaction_t trans, struct hammer_node *node)
1263 {
1264         KKASSERT((node->flags & HAMMER_NODE_CRCBAD) == 0);
1265         hammer_modify_buffer(trans, node->buffer,
1266                              node->ondisk, sizeof(*node->ondisk));
1267 }
1268
1269 static __inline void
1270 hammer_modify_node(hammer_transaction_t trans, hammer_node_t node,
1271                    void *base, int len)
1272 {
1273         hammer_crc_t *crcptr;
1274
1275         KKASSERT((char *)base >= (char *)node->ondisk &&
1276                  (char *)base + len <=
1277                     (char *)node->ondisk + sizeof(*node->ondisk));
1278         KKASSERT((node->flags & HAMMER_NODE_CRCBAD) == 0);
1279         hammer_modify_buffer(trans, node->buffer, base, len);
1280         crcptr = &node->ondisk->crc;
1281         hammer_modify_buffer(trans, node->buffer, crcptr, sizeof(hammer_crc_t));
1282         --node->buffer->io.modify_refs; /* only want one ref */
1283 }
1284
1285 /*
1286  * Indicate that the specified modifications have been completed.
1287  *
1288  * Do not try to generate the crc here, it's very expensive to do and a
1289  * sequence of insertions or deletions can result in many calls to this
1290  * function on the same node.
1291  */
1292 static __inline void
1293 hammer_modify_node_done(hammer_node_t node)
1294 {
1295         node->flags |= HAMMER_NODE_CRCGOOD;
1296         if ((node->flags & HAMMER_NODE_NEEDSCRC) == 0) {
1297                 node->flags |= HAMMER_NODE_NEEDSCRC;
1298                 node->buffer->io.gencrc = 1;
1299                 hammer_ref_node(node);
1300         }
1301         hammer_modify_buffer_done(node->buffer);
1302 }
1303
1304 #define hammer_modify_volume_field(trans, vol, field)           \
1305         hammer_modify_volume(trans, vol, &(vol)->ondisk->field, \
1306                              sizeof((vol)->ondisk->field))
1307
1308 #define hammer_modify_node_field(trans, node, field)            \
1309         hammer_modify_node(trans, node, &(node)->ondisk->field, \
1310                              sizeof((node)->ondisk->field))
1311