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 1.61 2008/05/05 20:34:47 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/globaldata.h>
  53 #include <sys/lockf.h>
  54 #include <sys/buf.h>
  55 #include <sys/queue.h>
  56 #include <sys/globaldata.h>
  57
  58 #include <sys/buf2.h>
  59 #include <sys/signal2.h>
  60 #include "hammer_disk.h"
  61 #include "hammer_mount.h"
  62 #include "hammer_ioctl.h"
  63
  64 #if defined(_KERNEL) || defined(_KERNEL_STRUCTURES)
  65
  66 MALLOC_DECLARE(M_HAMMER);
  67
  68 struct hammer_mount;
  69
  70 /*
  71  * Key structure used for custom RB tree inode lookups.  This prototypes
  72  * the function hammer_ino_rb_tree_RB_LOOKUP_INFO(root, info).
  73  */
  74 typedef struct hammer_inode_info {
  75         int64_t         obj_id;         /* (key) object identifier */
  76         hammer_tid_t    obj_asof;       /* (key) snapshot transid or 0 */
  77 } *hammer_inode_info_t;
  78
  79 typedef enum hammer_transaction_type {
  80         HAMMER_TRANS_RO,
  81         HAMMER_TRANS_STD,
  82         HAMMER_TRANS_FLS
  83 } hammer_transaction_type_t;
  84
  85 /*
  86  * HAMMER Transaction tracking
  87  */
  88 struct hammer_transaction {
  89         hammer_transaction_type_t type;
  90         struct hammer_mount *hmp;
  91         hammer_tid_t    tid;
  92         hammer_tid_t    time;
  93         struct hammer_volume *rootvol;
  94 };
  95
  96 typedef struct hammer_transaction *hammer_transaction_t;
  97
  98 /*
  99  * HAMMER locks
 100  */
 101 struct hammer_lock {
 102         int     refs;           /* active references delay writes */
 103         int     lockcount;      /* lock count for exclusive/shared access */
 104         int     wanted;
 105         struct thread *locktd;
 106 };
 107
 108 static __inline int
 109 hammer_islocked(struct hammer_lock *lock)
 110 {
 111         return(lock->lockcount != 0);
 112 }
 113
 114 static __inline int
 115 hammer_isactive(struct hammer_lock *lock)
 116 {
 117         return(lock->refs != 0);
 118 }
 119
 120 static __inline int
 121 hammer_islastref(struct hammer_lock *lock)
 122 {
 123         return(lock->refs == 1);
 124 }
 125
 126 /*
 127  * Return if we specifically own the lock exclusively.
 128  */
 129 static __inline int
 130 hammer_lock_excl_owned(struct hammer_lock *lock, thread_t td)
 131 {
 132         if (lock->lockcount > 0 && lock->locktd == td)
 133                 return(1);
 134         return(0);
 135 }
 136
 137 /*
 138  * Flush state, used by various structures
 139  */
 140 typedef enum hammer_inode_state {
 141         HAMMER_FST_IDLE,
 142         HAMMER_FST_SETUP,
 143         HAMMER_FST_FLUSH
 144 } hammer_inode_state_t;
 145
 146 TAILQ_HEAD(hammer_record_list, hammer_record);
 147
 148 /*
 149  * Cache object ids.  A fixed number of objid cache structures are
 150  * created to reserve object id's for newly created files in multiples
 151  * of 100,000, localized to a particular directory, and recycled as
 152  * needed.  This allows parallel create operations in different
 153  * directories to retain fairly localized object ids which in turn
 154  * improves reblocking performance and layout.
 155  */
 156 #define OBJID_CACHE_SIZE        128
 157 #define OBJID_CACHE_BULK        100000
 158
 159 typedef struct hammer_objid_cache {
 160         TAILQ_ENTRY(hammer_objid_cache) entry;
 161         struct hammer_inode             *dip;
 162         hammer_tid_t                    next_tid;
 163         int                             count;
 164 } *hammer_objid_cache_t;
 165
 166 /*
 167  * Structure used to represent an inode in-memory.
 168  *
 169  * The record and data associated with an inode may be out of sync with
 170  * the disk (xDIRTY flags), or not even on the disk at all (ONDISK flag
 171  * clear).
 172  *
 173  * An inode may also hold a cache of unsynchronized records, used for
 174  * database and directories only.  Unsynchronized regular file data is
 175  * stored in the buffer cache.
 176  *
 177  * NOTE: A file which is created and destroyed within the initial
 178  * synchronization period can wind up not doing any disk I/O at all.
 179  *
 180  * Finally, an inode may cache numerous disk-referencing B-Tree cursors.
 181  */
 182 struct hammer_ino_rb_tree;
 183 struct hammer_inode;
 184 RB_HEAD(hammer_ino_rb_tree, hammer_inode);
 185 RB_PROTOTYPEX(hammer_ino_rb_tree, INFO, hammer_inode, rb_node,
 186               hammer_ino_rb_compare, hammer_inode_info_t);
 187
 188 struct hammer_rec_rb_tree;
 189 struct hammer_record;
 190 RB_HEAD(hammer_rec_rb_tree, hammer_record);
 191 RB_PROTOTYPEX(hammer_rec_rb_tree, INFO, hammer_record, rb_node,
 192               hammer_rec_rb_compare, hammer_base_elm_t);
 193
 194 TAILQ_HEAD(hammer_node_list, hammer_node);
 195
 196 struct hammer_inode {
 197         RB_ENTRY(hammer_inode)  rb_node;
 198         hammer_inode_state_t    flush_state;
 199         int                     flush_group;
 200         TAILQ_ENTRY(hammer_inode) flush_entry;
 201         struct hammer_record_list target_list;  /* target of dependant recs */
 202         u_int64_t               obj_id;         /* (key) object identifier */
 203         hammer_tid_t            obj_asof;       /* (key) snapshot or 0 */
 204         struct hammer_mount     *hmp;
 205         hammer_objid_cache_t    objid_cache;
 206         int                     flags;
 207         int                     error;          /* flush error */
 208         int                     cursor_ip_refs; /* sanity */
 209         struct vnode            *vp;
 210         struct lockf            advlock;
 211         struct hammer_lock      lock;           /* sync copy interlock */
 212         TAILQ_HEAD(, bio)       bio_list;       /* BIOs to flush out */
 213         TAILQ_HEAD(, bio)       bio_alt_list;   /* BIOs to flush out */
 214         off_t                   trunc_off;
 215         struct hammer_inode_record ino_rec;     /* in-memory cache */
 216         struct hammer_inode_data ino_data;      /* in-memory cache */
 217         struct hammer_rec_rb_tree rec_tree;     /* in-memory cache */
 218         struct hammer_node      *cache[2];      /* search initiate cache */
 219
 220         /*
 221          * When a demark is created to synchronize an inode to
 222          * disk, certain fields are copied so the front-end VOPs
 223          * can continue to run in parallel with the synchronization
 224          * occuring in the background.
 225          */
 226         int             sync_flags;             /* to-sync flags cache */
 227         off_t           sync_trunc_off;         /* to-sync truncation */
 228         struct hammer_inode_record sync_ino_rec;/* to-sync cache */
 229         struct hammer_inode_data sync_ino_data; /* to-sync cache */
 230 };
 231
 232 typedef struct hammer_inode *hammer_inode_t;
 233
 234 #define VTOI(vp)        ((struct hammer_inode *)(vp)->v_data)
 235
 236 #define HAMMER_INODE_DDIRTY     0x0001  /* in-memory ino_data is dirty */
 237 #define HAMMER_INODE_RDIRTY     0x0002  /* in-memory ino_rec is dirty */
 238 #define HAMMER_INODE_ITIMES     0x0004  /* in-memory mtime/atime modified */
 239 #define HAMMER_INODE_XDIRTY     0x0008  /* in-memory records */
 240 #define HAMMER_INODE_ONDISK     0x0010  /* inode is on-disk (else not yet) */
 241 #define HAMMER_INODE_FLUSH      0x0020  /* flush on last ref */
 242 #define HAMMER_INODE_DELETED    0x0080  /* inode delete (backend) */
 243 #define HAMMER_INODE_DELONDISK  0x0100  /* delete synchronized to disk */
 244 #define HAMMER_INODE_RO         0x0200  /* read-only (because of as-of) */
 245 #define HAMMER_INODE_VHELD      0x0400  /* vnode held on sync */
 246 #define HAMMER_INODE_DONDISK    0x0800  /* data records may be on disk */
 247 #define HAMMER_INODE_BUFS       0x1000  /* dirty high level bps present */
 248 #define HAMMER_INODE_REFLUSH    0x2000  /* pipelined flush during flush */
 249 #define HAMMER_INODE_WRITE_ALT  0x4000  /* strategy writes to alt bioq */
 250 #define HAMMER_INODE_FLUSHW     0x8000  /* Someone waiting for flush */
 251
 252 #define HAMMER_INODE_TRUNCATED  0x00010000
 253 #define HAMMER_INODE_DELETING   0x00020000 /* inode delete request (frontend)*/
 254 #define HAMMER_INODE_RESIGNAL   0x00040000 /* re-signal on re-flush */
 255 #define HAMMER_INODE_RESIGNAL   0x00040000 /* re-signal on re-flush */
 256
 257 #define HAMMER_INODE_MODMASK    (HAMMER_INODE_DDIRTY|HAMMER_INODE_RDIRTY| \
 258                                  HAMMER_INODE_XDIRTY|HAMMER_INODE_BUFS|   \
 259                                  HAMMER_INODE_ITIMES|HAMMER_INODE_TRUNCATED|\
 260                                  HAMMER_INODE_DELETING)
 261
 262 #define HAMMER_INODE_MODMASK_NOXDIRTY \
 263                                 (HAMMER_INODE_MODMASK & ~HAMMER_INODE_XDIRTY)
 264
 265 #define HAMMER_MAX_INODE_CURSORS        4
 266
 267 #define HAMMER_FLUSH_SIGNAL     0x0001
 268 #define HAMMER_FLUSH_RECURSION  0x0002
 269
 270 /*
 271  * Structure used to represent an unsynchronized record in-memory.  These
 272  * records typically represent directory entries.  Only non-historical
 273  * records are kept in-memory.
 274  *
 275  * Records are organized as a per-inode RB-Tree.  If the inode is not
 276  * on disk then neither are any records and the in-memory record tree
 277  * represents the entire contents of the inode.  If the inode is on disk
 278  * then the on-disk B-Tree is scanned in parallel with the in-memory
 279  * RB-Tree to synthesize the current state of the file.
 280  *
 281  * Records are also used to enforce the ordering of directory create/delete
 282  * operations.  A new inode will not be flushed to disk unless its related
 283  * directory entry is also being flushed at the same time.  A directory entry
 284  * will not be removed unless its related inode is also being removed at the
 285  * same time.
 286  */
 287 typedef enum hammer_record_type {
 288         HAMMER_MEM_RECORD_GENERAL,      /* misc record */
 289         HAMMER_MEM_RECORD_ADD,          /* positive memory cache record */
 290         HAMMER_MEM_RECORD_DEL           /* negative delete-on-disk record */
 291 } hammer_record_type_t;
 292
 293 struct hammer_record {
 294         RB_ENTRY(hammer_record)         rb_node;
 295         TAILQ_ENTRY(hammer_record)      target_entry;
 296         hammer_inode_state_t            flush_state;
 297         int                             flush_group;
 298         hammer_record_type_t            type;
 299         struct hammer_lock              lock;
 300         struct hammer_inode             *ip;
 301         struct hammer_inode             *target_ip;
 302         union hammer_record_ondisk      rec;
 303         union hammer_data_ondisk        *data;
 304         int                             flags;
 305 };
 306
 307 typedef struct hammer_record *hammer_record_t;
 308
 309 /*
 310  * Record flags.  Note that FE can only be set by the frontend if the
 311  * record has not been interlocked by the backend w/ BE.
 312  */
 313 #define HAMMER_RECF_ALLOCDATA           0x0001
 314 #define HAMMER_RECF_ONRBTREE            0x0002
 315 #define HAMMER_RECF_DELETED_FE          0x0004  /* deleted (frontend) */
 316 #define HAMMER_RECF_DELETED_BE          0x0008  /* deleted (backend) */
 317 #define HAMMER_RECF_INBAND              0x0010
 318 #define HAMMER_RECF_INTERLOCK_BE        0x0020  /* backend interlock */
 319 #define HAMMER_RECF_WANTED              0x0040
 320 #define HAMMER_RECF_CONVERT_DELETE      0x0100 /* special case */
 321
 322 /*
 323  * In-memory structures representing on-disk structures.
 324  */
 325 struct hammer_volume;
 326 struct hammer_buffer;
 327 struct hammer_node;
 328 struct hammer_undo;
 329 RB_HEAD(hammer_vol_rb_tree, hammer_volume);
 330 RB_HEAD(hammer_buf_rb_tree, hammer_buffer);
 331 RB_HEAD(hammer_nod_rb_tree, hammer_node);
 332 RB_HEAD(hammer_und_rb_tree, hammer_undo);
 333
 334 RB_PROTOTYPE2(hammer_vol_rb_tree, hammer_volume, rb_node,
 335               hammer_vol_rb_compare, int32_t);
 336 RB_PROTOTYPE2(hammer_buf_rb_tree, hammer_buffer, rb_node,
 337               hammer_buf_rb_compare, hammer_off_t);
 338 RB_PROTOTYPE2(hammer_nod_rb_tree, hammer_node, rb_node,
 339               hammer_nod_rb_compare, hammer_off_t);
 340 RB_PROTOTYPE2(hammer_und_rb_tree, hammer_undo, rb_node,
 341               hammer_und_rb_compare, hammer_off_t);
 342
 343 /*
 344  * IO management - embedded at the head of various in-memory structures
 345  *
 346  * VOLUME       - hammer_volume containing meta-data
 347  * META_BUFFER  - hammer_buffer containing meta-data
 348  * DATA_BUFFER  - hammer_buffer containing pure-data
 349  *
 350  * Dirty volume headers and dirty meta-data buffers are locked until the
 351  * flusher can sequence them out.  Dirty pure-data buffers can be written.
 352  * Clean buffers can be passively released.
 353  */
 354 typedef enum hammer_io_type {
 355         HAMMER_STRUCTURE_VOLUME,
 356         HAMMER_STRUCTURE_META_BUFFER,
 357         HAMMER_STRUCTURE_UNDO_BUFFER,
 358         HAMMER_STRUCTURE_DATA_BUFFER
 359 } hammer_io_type_t;
 360
 361 union hammer_io_structure;
 362 struct hammer_io;
 363
 364 struct worklist {
 365         LIST_ENTRY(worklist) node;
 366 };
 367
 368 TAILQ_HEAD(hammer_io_list, hammer_io);
 369 typedef struct hammer_io_list *hammer_io_list_t;
 370
 371 struct hammer_io {
 372         struct worklist         worklist;
 373         struct hammer_lock      lock;
 374         enum hammer_io_type     type;
 375         struct hammer_mount     *hmp;
 376         TAILQ_ENTRY(hammer_io)  mod_entry; /* list entry if modified */
 377         hammer_io_list_t        mod_list;
 378         struct buf              *bp;
 379         int64_t                 offset;
 380         int                     loading;   /* loading/unloading interlock */
 381         int                     modify_refs;
 382
 383         u_int           modified : 1;   /* bp's data was modified */
 384         u_int           released : 1;   /* bp released (w/ B_LOCKED set) */
 385         u_int           running : 1;    /* bp write IO in progress */
 386         u_int           waiting : 1;    /* someone is waiting on us */
 387         u_int           validated : 1;  /* ondisk has been validated */
 388         u_int           flush : 1;      /* flush on last release */
 389         u_int           waitdep : 1;    /* flush waits for dependancies */
 390 };
 391
 392 typedef struct hammer_io *hammer_io_t;
 393
 394 /*
 395  * In-memory volume representing on-disk buffer
 396  */
 397 struct hammer_volume {
 398         struct hammer_io io;
 399         RB_ENTRY(hammer_volume) rb_node;
 400         struct hammer_buf_rb_tree rb_bufs_root;
 401         struct hammer_volume_ondisk *ondisk;
 402         int32_t vol_no;
 403         int64_t nblocks;        /* note: special calculation for statfs */
 404         int64_t buffer_base;    /* base offset of buffer 0 */
 405         hammer_off_t maxbuf_off; /* Maximum buffer offset */
 406         char    *vol_name;
 407         struct vnode *devvp;
 408         int     vol_flags;
 409 };
 410
 411 typedef struct hammer_volume *hammer_volume_t;
 412
 413 /*
 414  * In-memory buffer (other then volume, super-cluster, or cluster),
 415  * representing an on-disk buffer.
 416  */
 417 struct hammer_buffer {
 418         struct hammer_io io;
 419         RB_ENTRY(hammer_buffer) rb_node;
 420         void *ondisk;
 421         struct hammer_volume *volume;
 422         hammer_off_t zone2_offset;
 423         hammer_off_t zoneX_offset;
 424         struct hammer_node_list clist;
 425 };
 426
 427 typedef struct hammer_buffer *hammer_buffer_t;
 428
 429 /*
 430  * In-memory B-Tree node, representing an on-disk B-Tree node.
 431  *
 432  * This is a hang-on structure which is backed by a hammer_buffer,
 433  * indexed by a hammer_cluster, and used for fine-grained locking of
 434  * B-Tree nodes in order to properly control lock ordering.  A hammer_buffer
 435  * can contain multiple nodes representing wildly disassociated portions
 436  * of the B-Tree so locking cannot be done on a buffer-by-buffer basis.
 437  *
 438  * This structure uses a cluster-relative index to reduce the number
 439  * of layers required to access it, and also because all on-disk B-Tree
 440  * references are cluster-relative offsets.
 441  */
 442 struct hammer_node {
 443         struct hammer_lock      lock;           /* node-by-node lock */
 444         TAILQ_ENTRY(hammer_node) entry;         /* per-buffer linkage */
 445         RB_ENTRY(hammer_node)   rb_node;        /* per-cluster linkage */
 446         hammer_off_t            node_offset;    /* full offset spec */
 447         struct hammer_mount     *hmp;
 448         struct hammer_buffer    *buffer;        /* backing buffer */
 449         hammer_node_ondisk_t    ondisk;         /* ptr to on-disk structure */
 450         struct hammer_node      **cache1;       /* passive cache(s) */
 451         struct hammer_node      **cache2;
 452         int                     flags;
 453         int                     loading;        /* load interlock */
 454 };
 455
 456 #define HAMMER_NODE_DELETED     0x0001
 457 #define HAMMER_NODE_FLUSH       0x0002
 458
 459 typedef struct hammer_node      *hammer_node_t;
 460
 461 /*
 462  * List of locked nodes.
 463  */
 464 struct hammer_node_locklist {
 465         struct hammer_node_locklist *next;
 466         hammer_node_t   node;
 467 };
 468
 469 typedef struct hammer_node_locklist *hammer_node_locklist_t;
 470
 471
 472 /*
 473  * Common I/O management structure - embedded in in-memory structures
 474  * which are backed by filesystem buffers.
 475  */
 476 union hammer_io_structure {
 477         struct hammer_io        io;
 478         struct hammer_volume    volume;
 479         struct hammer_buffer    buffer;
 480 };
 481
 482 typedef union hammer_io_structure *hammer_io_structure_t;
 483
 484 /*
 485  * Allocation holes are recorded for a short period of time in an attempt
 486  * to use up the space.
 487  */
 488
 489 #define HAMMER_MAX_HOLES        8
 490
 491 struct hammer_hole;
 492
 493 struct hammer_holes {
 494         TAILQ_HEAD(, hammer_hole) list;
 495         int     count;
 496 };
 497
 498 typedef struct hammer_holes *hammer_holes_t;
 499
 500 struct hammer_hole {
 501         TAILQ_ENTRY(hammer_hole) entry;
 502         hammer_off_t    offset;
 503         int             bytes;
 504 };
 505
 506 typedef struct hammer_hole *hammer_hole_t;
 507
 508 #include "hammer_cursor.h"
 509
 510 /*
 511  * Undo history tracking
 512  */
 513 #define HAMMER_MAX_UNDOS        256
 514
 515 struct hammer_undo {
 516         RB_ENTRY(hammer_undo)   rb_node;
 517         TAILQ_ENTRY(hammer_undo) lru_entry;
 518         hammer_off_t            offset;
 519         int                     bytes;
 520 };
 521
 522 typedef struct hammer_undo *hammer_undo_t;
 523
 524 /*
 525  * Internal hammer mount data structure
 526  */
 527 struct hammer_mount {
 528         struct mount *mp;
 529         /*struct vnode *rootvp;*/
 530         struct hammer_ino_rb_tree rb_inos_root;
 531         struct hammer_vol_rb_tree rb_vols_root;
 532         struct hammer_nod_rb_tree rb_nods_root;
 533         struct hammer_und_rb_tree rb_undo_root;
 534         struct hammer_volume *rootvol;
 535         struct hammer_base_elm root_btree_beg;
 536         struct hammer_base_elm root_btree_end;
 537         char    *zbuf;  /* HAMMER_BUFSIZE bytes worth of all-zeros */
 538         int     hflags;
 539         int     ronly;
 540         int     nvolumes;
 541         int     volume_iterator;
 542         int     flusher_signal; /* flusher thread sequencer */
 543         int     flusher_act;    /* currently active flush group */
 544         int     flusher_done;   /* set to act when complete */
 545         int     flusher_next;   /* next flush group */
 546         int     flusher_lock;   /* lock sequencing of the next flush */
 547         int     flusher_exiting;
 548         hammer_off_t flusher_undo_start; /* UNDO window for flushes */
 549         int     reclaim_count;
 550         thread_t flusher_td;
 551         u_int   check_interrupt;
 552         uuid_t  fsid;
 553         udev_t  fsid_udev;
 554         struct hammer_io_list volu_list;        /* dirty undo buffers */
 555         struct hammer_io_list undo_list;        /* dirty undo buffers */
 556         struct hammer_io_list data_list;        /* dirty data buffers */
 557         struct hammer_io_list meta_list;        /* dirty meta bufs    */
 558         struct hammer_io_list lose_list;        /* loose buffers      */
 559         int     locked_dirty_count;             /* meta/volu count    */
 560         int     io_running_count;
 561         int     objid_cache_count;
 562         hammer_tid_t asof;
 563         hammer_off_t next_tid;
 564         u_int32_t namekey_iterator;
 565         hammer_off_t zone_limits[HAMMER_MAX_ZONES];
 566         struct netexport export;
 567         struct hammer_lock sync_lock;
 568         struct hammer_lock free_lock;
 569         struct lock blockmap_lock;
 570         struct hammer_blockmap  blockmap[HAMMER_MAX_ZONES];
 571         struct hammer_holes     holes[HAMMER_MAX_ZONES];
 572         struct hammer_undo      undos[HAMMER_MAX_UNDOS];
 573         int                     undo_alloc;
 574         TAILQ_HEAD(, hammer_undo)  undo_lru_list;
 575         TAILQ_HEAD(, hammer_inode) flush_list;
 576         TAILQ_HEAD(, hammer_objid_cache) objid_cache_list;
 577 };
 578
 579 typedef struct hammer_mount     *hammer_mount_t;
 580
 581 struct hammer_sync_info {
 582         int error;
 583         int waitfor;
 584 };
 585
 586 #endif
 587
 588 #if defined(_KERNEL)
 589
 590 extern struct vop_ops hammer_vnode_vops;
 591 extern struct vop_ops hammer_spec_vops;
 592 extern struct vop_ops hammer_fifo_vops;
 593 extern struct bio_ops hammer_bioops;
 594
 595 extern int hammer_debug_general;
 596 extern int hammer_debug_inode;
 597 extern int hammer_debug_locks;
 598 extern int hammer_debug_btree;
 599 extern int hammer_debug_tid;
 600 extern int hammer_debug_recover;
 601 extern int hammer_debug_recover_faults;
 602 extern int hammer_count_inodes;
 603 extern int hammer_count_records;
 604 extern int hammer_count_record_datas;
 605 extern int hammer_count_volumes;
 606 extern int hammer_count_buffers;
 607 extern int hammer_count_nodes;
 608 extern int hammer_count_dirtybufs;
 609 extern int hammer_limit_dirtybufs;
 610 extern int hammer_bio_count;
 611 extern int64_t hammer_contention_count;
 612
 613 int     hammer_vop_inactive(struct vop_inactive_args *);
 614 int     hammer_vop_reclaim(struct vop_reclaim_args *);
 615 int     hammer_get_vnode(struct hammer_inode *ip, struct vnode **vpp);
 616 struct hammer_inode *hammer_get_inode(hammer_transaction_t trans,
 617                         struct hammer_node **cache,
 618                         u_int64_t obj_id, hammer_tid_t asof, int flags,
 619                         int *errorp);
 620 void    hammer_put_inode(struct hammer_inode *ip);
 621 void    hammer_put_inode_ref(struct hammer_inode *ip);
 622
 623 int     hammer_unload_volume(hammer_volume_t volume, void *data __unused);
 624 int     hammer_unload_buffer(hammer_buffer_t buffer, void *data __unused);
 625 int     hammer_install_volume(hammer_mount_t hmp, const char *volname);
 626
 627 int     hammer_ip_lookup(hammer_cursor_t cursor, hammer_inode_t ip);
 628 int     hammer_ip_first(hammer_cursor_t cursor);
 629 int     hammer_ip_next(hammer_cursor_t cursor);
 630 int     hammer_ip_resolve_record_and_data(hammer_cursor_t cursor);
 631 int     hammer_ip_resolve_data(hammer_cursor_t cursor);
 632 int     hammer_ip_delete_record(hammer_cursor_t cursor, hammer_tid_t tid);
 633 int     hammer_delete_at_cursor(hammer_cursor_t cursor, int64_t *stat_bytes);
 634 int     hammer_ip_check_directory_empty(hammer_transaction_t trans,
 635                         hammer_cursor_t parent_cursor, hammer_inode_t ip);
 636 int     hammer_sync_hmp(hammer_mount_t hmp, int waitfor);
 637
 638 hammer_record_t
 639         hammer_alloc_mem_record(hammer_inode_t ip);
 640 void    hammer_flush_record_done(hammer_record_t record, int error);
 641 void    hammer_wait_mem_record(hammer_record_t record);
 642 void    hammer_rel_mem_record(hammer_record_t record);
 643
 644 int     hammer_cursor_up(hammer_cursor_t cursor);
 645 int     hammer_cursor_down(hammer_cursor_t cursor);
 646 int     hammer_cursor_upgrade(hammer_cursor_t cursor);
 647 void    hammer_cursor_downgrade(hammer_cursor_t cursor);
 648 int     hammer_cursor_seek(hammer_cursor_t cursor, hammer_node_t node,
 649                         int index);
 650 void    hammer_lock_ex(struct hammer_lock *lock);
 651 int     hammer_lock_ex_try(struct hammer_lock *lock);
 652 void    hammer_lock_sh(struct hammer_lock *lock);
 653 int     hammer_lock_upgrade(struct hammer_lock *lock);
 654 void    hammer_lock_downgrade(struct hammer_lock *lock);
 655 void    hammer_unlock(struct hammer_lock *lock);
 656 void    hammer_ref(struct hammer_lock *lock);
 657 void    hammer_unref(struct hammer_lock *lock);
 658
 659 u_int32_t hammer_to_unix_xid(uuid_t *uuid);
 660 void hammer_guid_to_uuid(uuid_t *uuid, u_int32_t guid);
 661 void    hammer_to_timespec(hammer_tid_t tid, struct timespec *ts);
 662 hammer_tid_t hammer_timespec_to_transid(struct timespec *ts);
 663 hammer_tid_t hammer_now_tid(void);
 664 hammer_tid_t hammer_str_to_tid(const char *str);
 665 hammer_tid_t hammer_alloc_objid(hammer_transaction_t trans, hammer_inode_t dip);
 666 void hammer_clear_objid(hammer_inode_t dip);
 667 void hammer_destroy_objid_cache(hammer_mount_t hmp);
 668
 669 int hammer_enter_undo_history(hammer_mount_t hmp, hammer_off_t offset,
 670                               int bytes);
 671 void hammer_clear_undo_history(hammer_mount_t hmp);
 672 enum vtype hammer_get_vnode_type(u_int8_t obj_type);
 673 int hammer_get_dtype(u_int8_t obj_type);
 674 u_int8_t hammer_get_obj_type(enum vtype vtype);
 675 int64_t hammer_directory_namekey(void *name, int len);
 676
 677 int     hammer_init_cursor(hammer_transaction_t trans, hammer_cursor_t cursor,
 678                            struct hammer_node **cache, hammer_inode_t ip);
 679 int     hammer_reinit_cursor(hammer_cursor_t cursor);
 680 void    hammer_normalize_cursor(hammer_cursor_t cursor);
 681 void    hammer_done_cursor(hammer_cursor_t cursor);
 682 void    hammer_mem_done(hammer_cursor_t cursor);
 683
 684 int     hammer_btree_lookup(hammer_cursor_t cursor);
 685 int     hammer_btree_first(hammer_cursor_t cursor);
 686 int     hammer_btree_last(hammer_cursor_t cursor);
 687 int     hammer_btree_extract(hammer_cursor_t cursor, int flags);
 688 int     hammer_btree_iterate(hammer_cursor_t cursor);
 689 int     hammer_btree_iterate_reverse(hammer_cursor_t cursor);
 690 int     hammer_btree_insert(hammer_cursor_t cursor, hammer_btree_elm_t elm);
 691 int     hammer_btree_delete(hammer_cursor_t cursor);
 692 int     hammer_btree_cmp(hammer_base_elm_t key1, hammer_base_elm_t key2);
 693 int     hammer_btree_chkts(hammer_tid_t ts, hammer_base_elm_t key);
 694 int     hammer_btree_correct_rhb(hammer_cursor_t cursor, hammer_tid_t tid);
 695 int     hammer_btree_correct_lhb(hammer_cursor_t cursor, hammer_tid_t tid);
 696
 697
 698 int     hammer_btree_lock_children(hammer_cursor_t cursor,
 699                         struct hammer_node_locklist **locklistp);
 700
 701 void    hammer_print_btree_node(hammer_node_ondisk_t ondisk);
 702 void    hammer_print_btree_elm(hammer_btree_elm_t elm, u_int8_t type, int i);
 703
 704 void    *hammer_bread(struct hammer_mount *hmp, hammer_off_t off,
 705                         int *errorp, struct hammer_buffer **bufferp);
 706 void    *hammer_bnew(struct hammer_mount *hmp, hammer_off_t off,
 707                         int *errorp, struct hammer_buffer **bufferp);
 708
 709 hammer_volume_t hammer_get_root_volume(hammer_mount_t hmp, int *errorp);
 710 int     hammer_dowrite(hammer_cursor_t cursor, hammer_inode_t ip,
 711                         struct bio *bio);
 712
 713 hammer_volume_t hammer_get_volume(hammer_mount_t hmp,
 714                         int32_t vol_no, int *errorp);
 715 hammer_buffer_t hammer_get_buffer(hammer_mount_t hmp,
 716                         hammer_off_t buf_offset, int isnew, int *errorp);
 717 void    hammer_uncache_buffer(struct hammer_mount *hmp, hammer_off_t off);
 718
 719 int             hammer_ref_volume(hammer_volume_t volume);
 720 int             hammer_ref_buffer(hammer_buffer_t buffer);
 721 void            hammer_flush_buffer_nodes(hammer_buffer_t buffer);
 722
 723 void            hammer_rel_volume(hammer_volume_t volume, int flush);
 724 void            hammer_rel_buffer(hammer_buffer_t buffer, int flush);
 725
 726 int             hammer_vfs_export(struct mount *mp, int op,
 727                         const struct export_args *export);
 728 hammer_node_t   hammer_get_node(hammer_mount_t hmp, hammer_off_t node_offset,
 729                         int isnew, int *errorp);
 730 void            hammer_ref_node(hammer_node_t node);
 731 hammer_node_t   hammer_ref_node_safe(struct hammer_mount *hmp,
 732                         struct hammer_node **cache, int *errorp);
 733 void            hammer_rel_node(hammer_node_t node);
 734 void            hammer_delete_node(hammer_transaction_t trans,
 735                         hammer_node_t node);
 736 void            hammer_cache_node(hammer_node_t node,
 737                         struct hammer_node **cache);
 738 void            hammer_uncache_node(struct hammer_node **cache);
 739 void            hammer_flush_node(hammer_node_t node);
 740
 741 void hammer_dup_buffer(struct hammer_buffer **bufferp,
 742                         struct hammer_buffer *buffer);
 743 hammer_node_t hammer_alloc_btree(hammer_transaction_t trans, int *errorp);
 744 void *hammer_alloc_record(hammer_transaction_t trans,
 745                         hammer_off_t *rec_offp, u_int16_t rec_type,
 746                         struct hammer_buffer **rec_bufferp,
 747                         int32_t data_len, void **datap,
 748                         hammer_off_t *data_offp,
 749                         struct hammer_buffer **data_bufferp, int *errorp);
 750 void *hammer_alloc_data(hammer_transaction_t trans, int32_t data_len,
 751                         hammer_off_t *data_offsetp,
 752                         struct hammer_buffer **data_bufferp, int *errorp);
 753
 754 int hammer_generate_undo(hammer_transaction_t trans, hammer_io_t io,
 755                         hammer_off_t zone1_offset, void *base, int len);
 756
 757 void hammer_put_volume(struct hammer_volume *volume, int flush);
 758 void hammer_put_buffer(struct hammer_buffer *buffer, int flush);
 759
 760 hammer_off_t hammer_freemap_alloc(hammer_transaction_t trans,
 761                         hammer_off_t owner, int *errorp);
 762 void hammer_freemap_free(hammer_transaction_t trans, hammer_off_t phys_offset,
 763                         hammer_off_t owner, int *errorp);
 764 hammer_off_t hammer_blockmap_alloc(hammer_transaction_t trans, int zone,
 765                         int bytes, int *errorp);
 766 void hammer_blockmap_free(hammer_transaction_t trans,
 767                         hammer_off_t bmap_off, int bytes);
 768 int hammer_blockmap_getfree(hammer_mount_t hmp, hammer_off_t bmap_off,
 769                         int *curp, int *errorp);
 770 hammer_off_t hammer_blockmap_lookup(hammer_mount_t hmp, hammer_off_t bmap_off,
 771                         int *errorp);
 772 hammer_off_t hammer_undo_lookup(hammer_mount_t hmp, hammer_off_t bmap_off,
 773                         int *errorp);
 774 int64_t hammer_undo_used(hammer_mount_t hmp);
 775 int64_t hammer_undo_space(hammer_mount_t hmp);
 776 int64_t hammer_undo_max(hammer_mount_t hmp);
 777
 778
 779 void hammer_start_transaction(struct hammer_transaction *trans,
 780                               struct hammer_mount *hmp);
 781 void hammer_simple_transaction(struct hammer_transaction *trans,
 782                               struct hammer_mount *hmp);
 783 void hammer_start_transaction_fls(struct hammer_transaction *trans,
 784                                   struct hammer_mount *hmp);
 785 void hammer_done_transaction(struct hammer_transaction *trans);
 786
 787 void hammer_modify_inode(struct hammer_transaction *trans,
 788                         hammer_inode_t ip, int flags);
 789 void hammer_flush_inode(hammer_inode_t ip, int flags);
 790 void hammer_flush_inode_done(hammer_inode_t ip);
 791 void hammer_wait_inode(hammer_inode_t ip);
 792
 793 int  hammer_create_inode(struct hammer_transaction *trans, struct vattr *vap,
 794                         struct ucred *cred, struct hammer_inode *dip,
 795                         struct hammer_inode **ipp);
 796 void hammer_rel_inode(hammer_inode_t ip, int flush);
 797 int hammer_sync_inode(hammer_inode_t ip);
 798 void hammer_test_inode(hammer_inode_t ip);
 799 void hammer_inode_unloadable_check(hammer_inode_t ip, int getvp);
 800
 801 int  hammer_ip_add_directory(struct hammer_transaction *trans,
 802                         hammer_inode_t dip, struct namecache *ncp,
 803                         hammer_inode_t nip);
 804 int  hammer_ip_del_directory(struct hammer_transaction *trans,
 805                         hammer_cursor_t cursor, hammer_inode_t dip,
 806                         hammer_inode_t ip);
 807 int  hammer_ip_add_record(struct hammer_transaction *trans,
 808                         hammer_record_t record);
 809 int  hammer_ip_delete_range(hammer_cursor_t cursor, hammer_inode_t ip,
 810                         int64_t ran_beg, int64_t ran_end);
 811 int  hammer_ip_delete_range_all(hammer_cursor_t cursor, hammer_inode_t ip,
 812                         int *countp);
 813 int  hammer_ip_sync_data(hammer_cursor_t cursor, hammer_inode_t ip,
 814                         int64_t offset, void *data, int bytes);
 815 int  hammer_ip_sync_record(hammer_transaction_t trans, hammer_record_t rec);
 816 int  hammer_ip_sync_record_cursor(hammer_cursor_t cursor, hammer_record_t rec);
 817
 818 int hammer_ioctl(hammer_inode_t ip, u_long com, caddr_t data, int fflag,
 819                         struct ucred *cred);
 820
 821 void hammer_io_init(hammer_io_t io, hammer_mount_t hmp,
 822                         enum hammer_io_type type);
 823 void hammer_io_reinit(hammer_io_t io, enum hammer_io_type type);
 824 int hammer_io_read(struct vnode *devvp, struct hammer_io *io);
 825 int hammer_io_new(struct vnode *devvp, struct hammer_io *io);
 826 void hammer_io_release(struct hammer_io *io);
 827 void hammer_io_flush(struct hammer_io *io);
 828 int hammer_io_checkflush(hammer_io_t io);
 829 void hammer_io_clear_modify(struct hammer_io *io);
 830 void hammer_io_waitdep(struct hammer_io *io);
 831
 832 void hammer_modify_volume(hammer_transaction_t trans, hammer_volume_t volume,
 833                         void *base, int len);
 834 void hammer_modify_buffer(hammer_transaction_t trans, hammer_buffer_t buffer,
 835                         void *base, int len);
 836 void hammer_modify_volume_done(hammer_volume_t volume);
 837 void hammer_modify_buffer_done(hammer_buffer_t buffer);
 838
 839 int hammer_ioc_reblock(hammer_transaction_t trans, hammer_inode_t ip,
 840                         struct hammer_ioc_reblock *reblock);
 841
 842 void hammer_init_holes(hammer_mount_t hmp, hammer_holes_t holes);
 843 void hammer_free_holes(hammer_mount_t hmp, hammer_holes_t holes);
 844 int hammer_signal_check(hammer_mount_t hmp);
 845
 846 void hammer_flusher_create(hammer_mount_t hmp);
 847 void hammer_flusher_destroy(hammer_mount_t hmp);
 848 void hammer_flusher_sync(hammer_mount_t hmp);
 849 void hammer_flusher_async(hammer_mount_t hmp);
 850
 851 int hammer_recover(hammer_mount_t hmp, hammer_volume_t rootvol);
 852
 853 void hammer_crc_set_blockmap(hammer_blockmap_t blockmap);
 854 void hammer_crc_set_volume(hammer_volume_ondisk_t ondisk);
 855
 856 int hammer_crc_test_blockmap(hammer_blockmap_t blockmap);
 857 int hammer_crc_test_volume(hammer_volume_ondisk_t ondisk);
 858 int hammer_crc_test_record(hammer_record_ondisk_t ondisk);
 859 int hammer_crc_test_btree(hammer_node_ondisk_t ondisk);
 860
 861 #endif
 862
 863 static __inline void
 864 hammer_modify_node_noundo(hammer_transaction_t trans, hammer_node_t node)
 865 {
 866         hammer_modify_buffer(trans, node->buffer, NULL, 0);
 867 }
 868
 869 static __inline void
 870 hammer_modify_node_all(hammer_transaction_t trans, struct hammer_node *node)
 871 {
 872         hammer_modify_buffer(trans, node->buffer,
 873                              node->ondisk, sizeof(*node->ondisk));
 874 }
 875
 876 static __inline void
 877 hammer_modify_node(hammer_transaction_t trans, hammer_node_t node,
 878                    void *base, int len)
 879 {
 880         hammer_crc_t *crcptr;
 881
 882         KKASSERT((char *)base >= (char *)node->ondisk &&
 883                  (char *)base + len <=
 884                     (char *)node->ondisk + sizeof(*node->ondisk));
 885         hammer_modify_buffer(trans, node->buffer, base, len);
 886         crcptr = &node->ondisk->crc;
 887         hammer_modify_buffer(trans, node->buffer, crcptr, sizeof(hammer_crc_t));
 888         --node->buffer->io.modify_refs; /* only want one ref */
 889 }
 890
 891 static __inline void
 892 hammer_modify_node_done(hammer_node_t node)
 893 {
 894         node->ondisk->crc = crc32(&node->ondisk->crc + 1, HAMMER_BTREE_CRCSIZE);
 895         hammer_modify_buffer_done(node->buffer);
 896 }
 897
 898 static __inline void
 899 hammer_modify_record_noundo(hammer_transaction_t trans, hammer_buffer_t buffer,
 900                             hammer_record_ondisk_t rec __unused)
 901 {
 902         hammer_modify_buffer(trans, buffer, NULL, 0);
 903 }
 904
 905 static __inline void
 906 hammer_modify_record_all(hammer_transaction_t trans, hammer_buffer_t buffer,
 907                          hammer_record_ondisk_t rec)
 908 {
 909         KKASSERT((char *)rec >= (char *)buffer->ondisk &&
 910                  (char *)(rec + 1) <= (char *)buffer->ondisk + HAMMER_BUFSIZE);
 911         hammer_modify_buffer(trans, buffer, rec, sizeof(*rec));
 912 }
 913
 914 static __inline void
 915 hammer_modify_record(hammer_transaction_t trans, hammer_buffer_t buffer,
 916                      hammer_record_ondisk_t rec, void *base, int len,
 917                      int dodelete)
 918 {
 919         KKASSERT((char *)base >= (char *)buffer->ondisk &&
 920                  (char *)base + len <= (char *)buffer->ondisk + HAMMER_BUFSIZE);
 921         KKASSERT((char *)rec >= (char *)buffer->ondisk &&
 922                  (char *)(rec + 1) <= (char *)buffer->ondisk + HAMMER_BUFSIZE);
 923
 924         /*
 925          * Due to undo overheads it is more efficient to just undo the whole
 926          * record.
 927          */
 928         hammer_modify_buffer(trans, buffer, rec, sizeof(*rec));
 929 #if 0
 930         hammer_modify_buffer(trans, buffer, base, len);
 931         hammer_modify_buffer(trans, buffer, &rec->base.rec_crc,
 932                                      sizeof(rec->base.rec_crc));
 933         --node->buffer->io.modify_refs; /* only want one ref */
 934         if (dodelete) {
 935                 hammer_modify_buffer(trans, buffer, &rec->base.signature,
 936                                      sizeof(rec->base.signature));
 937                 --node->buffer->io.modify_refs; /* only want one ref */
 938         }
 939 #endif
 940 }
 941
 942 static __inline void
 943 hammer_modify_record_done(hammer_buffer_t buffer, hammer_record_ondisk_t rec)
 944 {
 945         rec->base.rec_crc = crc32(&rec->base.rec_crc + 1,
 946                                   HAMMER_RECORD_CRCSIZE);
 947         hammer_modify_buffer_done(buffer);
 948 }
 949
 950 #define hammer_modify_volume_field(trans, vol, field)           \
 951         hammer_modify_volume(trans, vol, &(vol)->ondisk->field, \
 952                              sizeof((vol)->ondisk->field))
 953
 954 #define hammer_modify_node_field(trans, node, field)            \
 955         hammer_modify_node(trans, node, &(node)->ondisk->field, \
 956                              sizeof((node)->ondisk->field))
 957
 958 #define hammer_modify_record_field(trans, buffer, rec, field, dodelete) \
 959         hammer_modify_record(trans, buffer, rec, &(rec)->field,         \
 960         sizeof((rec)->field), dodelete)
 961