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 1.72 2008/06/02 20:19:03 dillon Exp $
  35  */
  36 /*
  37  * This header file contains structures used internally by the HAMMERFS
  38  * implementation.  See hammer_disk.h for on-disk structures.
  39  */
  40
  41 #include <sys/param.h>
  42 #include <sys/types.h>
  43 #include <sys/kernel.h>
  44 #include <sys/conf.h>
  45 #include <sys/systm.h>
  46 #include <sys/tree.h>
  47 #include <sys/malloc.h>
  48 #include <sys/mount.h>
  49 #include <sys/mountctl.h>
  50 #include <sys/vnode.h>
  51 #include <sys/proc.h>
  52 #include <sys/stat.h>
  53 #include <sys/globaldata.h>
  54 #include <sys/lockf.h>
  55 #include <sys/buf.h>
  56 #include <sys/queue.h>
  57 #include <sys/globaldata.h>
  58
  59 #include <sys/buf2.h>
  60 #include <sys/signal2.h>
  61 #include "hammer_disk.h"
  62 #include "hammer_mount.h"
  63 #include "hammer_ioctl.h"
  64
  65 #if defined(_KERNEL) || defined(_KERNEL_STRUCTURES)
  66
  67 MALLOC_DECLARE(M_HAMMER);
  68
  69 struct hammer_mount;
  70
  71 /*
  72  * Key structure used for custom RB tree inode lookups.  This prototypes
  73  * the function hammer_ino_rb_tree_RB_LOOKUP_INFO(root, info).
  74  */
  75 typedef struct hammer_inode_info {
  76         int64_t         obj_id;         /* (key) object identifier */
  77         hammer_tid_t    obj_asof;       /* (key) snapshot transid or 0 */
  78 } *hammer_inode_info_t;
  79
  80 typedef enum hammer_transaction_type {
  81         HAMMER_TRANS_RO,
  82         HAMMER_TRANS_STD,
  83         HAMMER_TRANS_FLS
  84 } hammer_transaction_type_t;
  85
  86 /*
  87  * HAMMER Transaction tracking
  88  */
  89 struct hammer_transaction {
  90         hammer_transaction_type_t type;
  91         struct hammer_mount *hmp;
  92         hammer_tid_t    tid;
  93         hammer_tid_t    time;
  94         int             sync_lock_refs;
  95         struct hammer_volume *rootvol;
  96 };
  97
  98 typedef struct hammer_transaction *hammer_transaction_t;
  99
 100 /*
 101  * HAMMER locks
 102  */
 103 struct hammer_lock {
 104         int     refs;           /* active references delay writes */
 105         int     lockcount;      /* lock count for exclusive/shared access */
 106         int     wanted;
 107         struct thread *locktd;
 108 };
 109
 110 static __inline int
 111 hammer_islocked(struct hammer_lock *lock)
 112 {
 113         return(lock->lockcount != 0);
 114 }
 115
 116 static __inline int
 117 hammer_isactive(struct hammer_lock *lock)
 118 {
 119         return(lock->refs != 0);
 120 }
 121
 122 static __inline int
 123 hammer_islastref(struct hammer_lock *lock)
 124 {
 125         return(lock->refs == 1);
 126 }
 127
 128 /*
 129  * Return if we specifically own the lock exclusively.
 130  */
 131 static __inline int
 132 hammer_lock_excl_owned(struct hammer_lock *lock, thread_t td)
 133 {
 134         if (lock->lockcount > 0 && lock->locktd == td)
 135                 return(1);
 136         return(0);
 137 }
 138
 139 /*
 140  * Flush state, used by various structures
 141  */
 142 typedef enum hammer_inode_state {
 143         HAMMER_FST_IDLE,
 144         HAMMER_FST_SETUP,
 145         HAMMER_FST_FLUSH
 146 } hammer_inode_state_t;
 147
 148 TAILQ_HEAD(hammer_record_list, hammer_record);
 149
 150 /*
 151  * Cache object ids.  A fixed number of objid cache structures are
 152  * created to reserve object id's for newly created files in multiples
 153  * of 100,000, localized to a particular directory, and recycled as
 154  * needed.  This allows parallel create operations in different
 155  * directories to retain fairly localized object ids which in turn
 156  * improves reblocking performance and layout.
 157  */
 158 #define OBJID_CACHE_SIZE        128
 159 #define OBJID_CACHE_BULK        100000
 160
 161 typedef struct hammer_objid_cache {
 162         TAILQ_ENTRY(hammer_objid_cache) entry;
 163         struct hammer_inode             *dip;
 164         hammer_tid_t                    next_tid;
 165         int                             count;
 166 } *hammer_objid_cache_t;
 167
 168 /*
 169  * Structure used to represent an inode in-memory.
 170  *
 171  * The record and data associated with an inode may be out of sync with
 172  * the disk (xDIRTY flags), or not even on the disk at all (ONDISK flag
 173  * clear).
 174  *
 175  * An inode may also hold a cache of unsynchronized records, used for
 176  * database and directories only.  Unsynchronized regular file data is
 177  * stored in the buffer cache.
 178  *
 179  * NOTE: A file which is created and destroyed within the initial
 180  * synchronization period can wind up not doing any disk I/O at all.
 181  *
 182  * Finally, an inode may cache numerous disk-referencing B-Tree cursors.
 183  */
 184 struct hammer_ino_rb_tree;
 185 struct hammer_inode;
 186 RB_HEAD(hammer_ino_rb_tree, hammer_inode);
 187 RB_PROTOTYPEX(hammer_ino_rb_tree, INFO, hammer_inode, rb_node,
 188               hammer_ino_rb_compare, hammer_inode_info_t);
 189
 190 struct hammer_rec_rb_tree;
 191 struct hammer_record;
 192 RB_HEAD(hammer_rec_rb_tree, hammer_record);
 193 RB_PROTOTYPEX(hammer_rec_rb_tree, INFO, hammer_record, rb_node,
 194               hammer_rec_rb_compare, hammer_base_elm_t);
 195
 196 TAILQ_HEAD(hammer_node_list, hammer_node);
 197
 198 struct hammer_inode {
 199         RB_ENTRY(hammer_inode)  rb_node;
 200         hammer_inode_state_t    flush_state;
 201         int                     flush_group;
 202         TAILQ_ENTRY(hammer_inode) flush_entry;
 203         struct hammer_record_list target_list;  /* target of dependant recs */
 204         u_int64_t               obj_id;         /* (key) object identifier */
 205         hammer_tid_t            obj_asof;       /* (key) snapshot or 0 */
 206         struct hammer_mount     *hmp;
 207         hammer_objid_cache_t    objid_cache;
 208         int                     flags;
 209         int                     error;          /* flush error */
 210         int                     cursor_ip_refs; /* sanity */
 211         int                     rsv_databufs;   /* counts towards hmp */
 212         struct vnode            *vp;
 213         struct lockf            advlock;
 214         struct hammer_lock      lock;           /* sync copy interlock */
 215         TAILQ_HEAD(, bio)       bio_list;       /* BIOs to flush out */
 216         TAILQ_HEAD(, bio)       bio_alt_list;   /* BIOs to flush out */
 217         off_t                   trunc_off;
 218         struct hammer_btree_leaf_elm ino_leaf;  /* in-memory cache */
 219         struct hammer_inode_data ino_data;      /* in-memory cache */
 220         struct hammer_rec_rb_tree rec_tree;     /* in-memory cache */
 221         struct hammer_node      *cache[2];      /* search initiate cache */
 222
 223         /*
 224          * When a demark is created to synchronize an inode to
 225          * disk, certain fields are copied so the front-end VOPs
 226          * can continue to run in parallel with the synchronization
 227          * occuring in the background.
 228          */
 229         int             sync_flags;             /* to-sync flags cache */
 230         off_t           sync_trunc_off;         /* to-sync truncation */
 231         struct hammer_btree_leaf_elm sync_ino_leaf; /* to-sync cache */
 232         struct hammer_inode_data sync_ino_data; /* to-sync cache */
 233 };
 234
 235 typedef struct hammer_inode *hammer_inode_t;
 236
 237 #define VTOI(vp)        ((struct hammer_inode *)(vp)->v_data)
 238
 239 #define HAMMER_INODE_DDIRTY     0x0001  /* in-memory ino_data is dirty */
 240 #define HAMMER_INODE_RSV_INODES 0x0002  /* hmp->rsv_inodes bumped */
 241 #define HAMMER_INODE_ITIMES     0x0004  /* in-memory mtime/atime modified */
 242 #define HAMMER_INODE_XDIRTY     0x0008  /* in-memory records */
 243 #define HAMMER_INODE_ONDISK     0x0010  /* inode is on-disk (else not yet) */
 244 #define HAMMER_INODE_FLUSH      0x0020  /* flush on last ref */
 245 #define HAMMER_INODE_DELETED    0x0080  /* inode delete (backend) */
 246 #define HAMMER_INODE_DELONDISK  0x0100  /* delete synchronized to disk */
 247 #define HAMMER_INODE_RO         0x0200  /* read-only (because of as-of) */
 248 #define HAMMER_INODE_VHELD      0x0400  /* vnode held on sync */
 249 #define HAMMER_INODE_DONDISK    0x0800  /* data records may be on disk */
 250 #define HAMMER_INODE_BUFS       0x1000  /* dirty high level bps present */
 251 #define HAMMER_INODE_REFLUSH    0x2000  /* pipelined flush during flush */
 252 #define HAMMER_INODE_WRITE_ALT  0x4000  /* strategy writes to alt bioq */
 253 #define HAMMER_INODE_FLUSHW     0x8000  /* Someone waiting for flush */
 254
 255 #define HAMMER_INODE_TRUNCATED  0x00010000
 256 #define HAMMER_INODE_DELETING   0x00020000 /* inode delete request (frontend)*/
 257 #define HAMMER_INODE_RESIGNAL   0x00040000 /* re-signal on re-flush */
 258
 259 #define HAMMER_INODE_MODMASK    (HAMMER_INODE_DDIRTY|                       \
 260                                  HAMMER_INODE_XDIRTY|HAMMER_INODE_BUFS|     \
 261                                  HAMMER_INODE_ITIMES|HAMMER_INODE_TRUNCATED|\
 262                                  HAMMER_INODE_DELETING)
 263
 264 #define HAMMER_INODE_MODMASK_NOXDIRTY \
 265                                 (HAMMER_INODE_MODMASK & ~HAMMER_INODE_XDIRTY)
 266
 267 #define HAMMER_MAX_INODE_CURSORS        4
 268
 269 #define HAMMER_FLUSH_SIGNAL     0x0001
 270 #define HAMMER_FLUSH_RECURSION  0x0002
 271
 272 /*
 273  * Structure used to represent an unsynchronized record in-memory.  These
 274  * records typically represent directory entries.  Only non-historical
 275  * records are kept in-memory.
 276  *
 277  * Records are organized as a per-inode RB-Tree.  If the inode is not
 278  * on disk then neither are any records and the in-memory record tree
 279  * represents the entire contents of the inode.  If the inode is on disk
 280  * then the on-disk B-Tree is scanned in parallel with the in-memory
 281  * RB-Tree to synthesize the current state of the file.
 282  *
 283  * Records are also used to enforce the ordering of directory create/delete
 284  * operations.  A new inode will not be flushed to disk unless its related
 285  * directory entry is also being flushed at the same time.  A directory entry
 286  * will not be removed unless its related inode is also being removed at the
 287  * same time.
 288  */
 289 typedef enum hammer_record_type {
 290         HAMMER_MEM_RECORD_GENERAL,      /* misc record */
 291         HAMMER_MEM_RECORD_INODE,        /* inode record */
 292         HAMMER_MEM_RECORD_ADD,          /* positive memory cache record */
 293         HAMMER_MEM_RECORD_DEL           /* negative delete-on-disk record */
 294 } hammer_record_type_t;
 295
 296 struct hammer_record {
 297         RB_ENTRY(hammer_record)         rb_node;
 298         TAILQ_ENTRY(hammer_record)      target_entry;
 299         hammer_inode_state_t            flush_state;
 300         int                             flush_group;
 301         hammer_record_type_t            type;
 302         struct hammer_lock              lock;
 303         struct hammer_inode             *ip;
 304         struct hammer_inode             *target_ip;
 305         struct hammer_btree_leaf_elm    leaf;
 306         union hammer_data_ondisk        *data;
 307         int                             flags;
 308 };
 309
 310 typedef struct hammer_record *hammer_record_t;
 311
 312 /*
 313  * Record flags.  Note that FE can only be set by the frontend if the
 314  * record has not been interlocked by the backend w/ BE.
 315  */
 316 #define HAMMER_RECF_ALLOCDATA           0x0001
 317 #define HAMMER_RECF_ONRBTREE            0x0002
 318 #define HAMMER_RECF_DELETED_FE          0x0004  /* deleted (frontend) */
 319 #define HAMMER_RECF_DELETED_BE          0x0008  /* deleted (backend) */
 320 #define HAMMER_RECF_UNUSED0010          0x0010
 321 #define HAMMER_RECF_INTERLOCK_BE        0x0020  /* backend interlock */
 322 #define HAMMER_RECF_WANTED              0x0040
 323 #define HAMMER_RECF_CONVERT_DELETE      0x0100 /* special case */
 324
 325 /*
 326  * In-memory structures representing on-disk structures.
 327  */
 328 struct hammer_volume;
 329 struct hammer_buffer;
 330 struct hammer_node;
 331 struct hammer_undo;
 332 RB_HEAD(hammer_vol_rb_tree, hammer_volume);
 333 RB_HEAD(hammer_buf_rb_tree, hammer_buffer);
 334 RB_HEAD(hammer_nod_rb_tree, hammer_node);
 335 RB_HEAD(hammer_und_rb_tree, hammer_undo);
 336
 337 RB_PROTOTYPE2(hammer_vol_rb_tree, hammer_volume, rb_node,
 338               hammer_vol_rb_compare, int32_t);
 339 RB_PROTOTYPE2(hammer_buf_rb_tree, hammer_buffer, rb_node,
 340               hammer_buf_rb_compare, hammer_off_t);
 341 RB_PROTOTYPE2(hammer_nod_rb_tree, hammer_node, rb_node,
 342               hammer_nod_rb_compare, hammer_off_t);
 343 RB_PROTOTYPE2(hammer_und_rb_tree, hammer_undo, rb_node,
 344               hammer_und_rb_compare, hammer_off_t);
 345
 346 /*
 347  * IO management - embedded at the head of various in-memory structures
 348  *
 349  * VOLUME       - hammer_volume containing meta-data
 350  * META_BUFFER  - hammer_buffer containing meta-data
 351  * DATA_BUFFER  - hammer_buffer containing pure-data
 352  *
 353  * Dirty volume headers and dirty meta-data buffers are locked until the
 354  * flusher can sequence them out.  Dirty pure-data buffers can be written.
 355  * Clean buffers can be passively released.
 356  */
 357 typedef enum hammer_io_type {
 358         HAMMER_STRUCTURE_VOLUME,
 359         HAMMER_STRUCTURE_META_BUFFER,
 360         HAMMER_STRUCTURE_UNDO_BUFFER,
 361         HAMMER_STRUCTURE_DATA_BUFFER
 362 } hammer_io_type_t;
 363
 364 union hammer_io_structure;
 365 struct hammer_io;
 366
 367 struct worklist {
 368         LIST_ENTRY(worklist) node;
 369 };
 370
 371 TAILQ_HEAD(hammer_io_list, hammer_io);
 372 typedef struct hammer_io_list *hammer_io_list_t;
 373
 374 struct hammer_io {
 375         struct worklist         worklist;
 376         struct hammer_lock      lock;
 377         enum hammer_io_type     type;
 378         struct hammer_mount     *hmp;
 379         TAILQ_ENTRY(hammer_io)  mod_entry; /* list entry if modified */
 380         hammer_io_list_t        mod_list;
 381         struct buf              *bp;
 382         int64_t                 offset;
 383         int                     loading;   /* loading/unloading interlock */
 384         int                     modify_refs;
 385
 386         u_int           modified : 1;   /* bp's data was modified */
 387         u_int           released : 1;   /* bp released (w/ B_LOCKED set) */
 388         u_int           running : 1;    /* bp write IO in progress */
 389         u_int           waiting : 1;    /* someone is waiting on us */
 390         u_int           validated : 1;  /* ondisk has been validated */
 391         u_int           waitdep : 1;    /* flush waits for dependancies */
 392 };
 393
 394 typedef struct hammer_io *hammer_io_t;
 395
 396 /*
 397  * In-memory volume representing on-disk buffer
 398  */
 399 struct hammer_volume {
 400         struct hammer_io io;
 401         RB_ENTRY(hammer_volume) rb_node;
 402         struct hammer_buf_rb_tree rb_bufs_root;
 403         struct hammer_volume_ondisk *ondisk;
 404         int32_t vol_no;
 405         int64_t nblocks;        /* note: special calculation for statfs */
 406         int64_t buffer_base;    /* base offset of buffer 0 */
 407         hammer_off_t maxbuf_off; /* Maximum buffer offset (zone-2) */
 408         hammer_off_t maxraw_off; /* Maximum raw offset for device */
 409         char    *vol_name;
 410         struct vnode *devvp;
 411         int     vol_flags;
 412 };
 413
 414 typedef struct hammer_volume *hammer_volume_t;
 415
 416 /*
 417  * In-memory buffer (other then volume, super-cluster, or cluster),
 418  * representing an on-disk buffer.
 419  */
 420 struct hammer_buffer {
 421         struct hammer_io io;
 422         RB_ENTRY(hammer_buffer) rb_node;
 423         void *ondisk;
 424         struct hammer_volume *volume;
 425         hammer_off_t zone2_offset;
 426         hammer_off_t zoneX_offset;
 427         struct hammer_node_list clist;
 428 };
 429
 430 typedef struct hammer_buffer *hammer_buffer_t;
 431
 432 /*
 433  * In-memory B-Tree node, representing an on-disk B-Tree node.
 434  *
 435  * This is a hang-on structure which is backed by a hammer_buffer,
 436  * indexed by a hammer_cluster, and used for fine-grained locking of
 437  * B-Tree nodes in order to properly control lock ordering.  A hammer_buffer
 438  * can contain multiple nodes representing wildly disassociated portions
 439  * of the B-Tree so locking cannot be done on a buffer-by-buffer basis.
 440  *
 441  * This structure uses a cluster-relative index to reduce the number
 442  * of layers required to access it, and also because all on-disk B-Tree
 443  * references are cluster-relative offsets.
 444  */
 445 struct hammer_node {
 446         struct hammer_lock      lock;           /* node-by-node lock */
 447         TAILQ_ENTRY(hammer_node) entry;         /* per-buffer linkage */
 448         RB_ENTRY(hammer_node)   rb_node;        /* per-cluster linkage */
 449         hammer_off_t            node_offset;    /* full offset spec */
 450         struct hammer_mount     *hmp;
 451         struct hammer_buffer    *buffer;        /* backing buffer */
 452         hammer_node_ondisk_t    ondisk;         /* ptr to on-disk structure */
 453         struct hammer_node      **cache1;       /* passive cache(s) */
 454         struct hammer_node      **cache2;
 455         int                     flags;
 456         int                     loading;        /* load interlock */
 457 };
 458
 459 #define HAMMER_NODE_DELETED     0x0001
 460 #define HAMMER_NODE_FLUSH       0x0002
 461
 462 typedef struct hammer_node      *hammer_node_t;
 463
 464 /*
 465  * List of locked nodes.
 466  */
 467 struct hammer_node_locklist {
 468         struct hammer_node_locklist *next;
 469         hammer_node_t   node;
 470 };
 471
 472 typedef struct hammer_node_locklist *hammer_node_locklist_t;
 473
 474
 475 /*
 476  * Common I/O management structure - embedded in in-memory structures
 477  * which are backed by filesystem buffers.
 478  */
 479 union hammer_io_structure {
 480         struct hammer_io        io;
 481         struct hammer_volume    volume;
 482         struct hammer_buffer    buffer;
 483 };
 484
 485 typedef union hammer_io_structure *hammer_io_structure_t;
 486
 487 /*
 488  * Allocation holes are recorded for a short period of time in an attempt
 489  * to use up the space.
 490  */
 491
 492 #define HAMMER_MAX_HOLES        8
 493
 494 struct hammer_hole;
 495
 496 struct hammer_holes {
 497         TAILQ_HEAD(, hammer_hole) list;
 498         int     count;
 499 };
 500
 501 typedef struct hammer_holes *hammer_holes_t;
 502
 503 struct hammer_hole {
 504         TAILQ_ENTRY(hammer_hole) entry;
 505         hammer_off_t    offset;
 506         int             bytes;
 507 };
 508
 509 typedef struct hammer_hole *hammer_hole_t;
 510
 511 #include "hammer_cursor.h"
 512
 513 /*
 514  * Undo history tracking
 515  */
 516 #define HAMMER_MAX_UNDOS        256
 517
 518 struct hammer_undo {
 519         RB_ENTRY(hammer_undo)   rb_node;
 520         TAILQ_ENTRY(hammer_undo) lru_entry;
 521         hammer_off_t            offset;
 522         int                     bytes;
 523 };
 524
 525 typedef struct hammer_undo *hammer_undo_t;
 526
 527 /*
 528  * Internal hammer mount data structure
 529  */
 530 struct hammer_mount {
 531         struct mount *mp;
 532         /*struct vnode *rootvp;*/
 533         struct hammer_ino_rb_tree rb_inos_root;
 534         struct hammer_vol_rb_tree rb_vols_root;
 535         struct hammer_nod_rb_tree rb_nods_root;
 536         struct hammer_und_rb_tree rb_undo_root;
 537         struct hammer_volume *rootvol;
 538         struct hammer_base_elm root_btree_beg;
 539         struct hammer_base_elm root_btree_end;
 540         char    *zbuf;  /* HAMMER_BUFSIZE bytes worth of all-zeros */
 541         int     hflags;
 542         int     ronly;
 543         int     nvolumes;
 544         int     volume_iterator;
 545         int     rsv_inodes;     /* reserved space due to dirty inodes */
 546         int     rsv_databufs;   /* reserved space due to dirty buffers */
 547         int     rsv_databytes;  /* reserved space due to record data */
 548         int     rsv_recs;       /* reserved space due to dirty records */
 549         int     flusher_signal; /* flusher thread sequencer */
 550         int     flusher_act;    /* currently active flush group */
 551         int     flusher_done;   /* set to act when complete */
 552         int     flusher_next;   /* next flush group */
 553         int     flusher_lock;   /* lock sequencing of the next flush */
 554         int     flusher_exiting;
 555         hammer_tid_t flusher_tid; /* last flushed transaction id */
 556         hammer_off_t flusher_undo_start; /* UNDO window for flushes */
 557         int     reclaim_count;
 558         thread_t flusher_td;
 559         u_int   check_interrupt;
 560         uuid_t  fsid;
 561         udev_t  fsid_udev;
 562         struct hammer_io_list volu_list;        /* dirty undo buffers */
 563         struct hammer_io_list undo_list;        /* dirty undo buffers */
 564         struct hammer_io_list data_list;        /* dirty data buffers */
 565         struct hammer_io_list meta_list;        /* dirty meta bufs    */
 566         struct hammer_io_list lose_list;        /* loose buffers      */
 567         int     locked_dirty_count;             /* meta/volu count    */
 568         int     io_running_count;
 569         int     objid_cache_count;
 570         hammer_tid_t asof;
 571         hammer_off_t next_tid;
 572         int64_t copy_stat_freebigblocks;        /* number of free bigblocks */
 573
 574         u_int32_t namekey_iterator;
 575         hammer_off_t zone_limits[HAMMER_MAX_ZONES];
 576         struct netexport export;
 577         struct hammer_lock sync_lock;
 578         struct hammer_lock free_lock;
 579         struct lock blockmap_lock;
 580         struct hammer_blockmap  blockmap[HAMMER_MAX_ZONES];
 581         struct hammer_holes     holes[HAMMER_MAX_ZONES];
 582         struct hammer_undo      undos[HAMMER_MAX_UNDOS];
 583         int                     undo_alloc;
 584         TAILQ_HEAD(, hammer_undo)  undo_lru_list;
 585         TAILQ_HEAD(, hammer_inode) flush_list;
 586         TAILQ_HEAD(, hammer_objid_cache) objid_cache_list;
 587 };
 588
 589 typedef struct hammer_mount     *hammer_mount_t;
 590
 591 struct hammer_sync_info {
 592         int error;
 593         int waitfor;
 594 };
 595
 596 #endif
 597
 598 #if defined(_KERNEL)
 599
 600 extern struct vop_ops hammer_vnode_vops;
 601 extern struct vop_ops hammer_spec_vops;
 602 extern struct vop_ops hammer_fifo_vops;
 603 extern struct bio_ops hammer_bioops;
 604
 605 extern int hammer_debug_io;
 606 extern int hammer_debug_general;
 607 extern int hammer_debug_debug;
 608 extern int hammer_debug_inode;
 609 extern int hammer_debug_locks;
 610 extern int hammer_debug_btree;
 611 extern int hammer_debug_tid;
 612 extern int hammer_debug_recover;
 613 extern int hammer_debug_recover_faults;
 614 extern int hammer_count_inodes;
 615 extern int hammer_count_records;
 616 extern int hammer_count_record_datas;
 617 extern int hammer_count_volumes;
 618 extern int hammer_count_buffers;
 619 extern int hammer_count_nodes;
 620 extern int hammer_count_dirtybufs;
 621 extern int hammer_limit_dirtybufs;
 622 extern int hammer_bio_count;
 623 extern int64_t hammer_contention_count;
 624
 625 int     hammer_vop_inactive(struct vop_inactive_args *);
 626 int     hammer_vop_reclaim(struct vop_reclaim_args *);
 627 int     hammer_get_vnode(struct hammer_inode *ip, struct vnode **vpp);
 628 struct hammer_inode *hammer_get_inode(hammer_transaction_t trans,
 629                         struct hammer_node **cache,
 630                         u_int64_t obj_id, hammer_tid_t asof, int flags,
 631                         int *errorp);
 632 void    hammer_put_inode(struct hammer_inode *ip);
 633 void    hammer_put_inode_ref(struct hammer_inode *ip);
 634
 635 int     hammer_unload_volume(hammer_volume_t volume, void *data __unused);
 636 int     hammer_unload_buffer(hammer_buffer_t buffer, void *data __unused);
 637 int     hammer_install_volume(hammer_mount_t hmp, const char *volname);
 638
 639 int     hammer_ip_lookup(hammer_cursor_t cursor);
 640 int     hammer_ip_first(hammer_cursor_t cursor);
 641 int     hammer_ip_next(hammer_cursor_t cursor);
 642 int     hammer_ip_resolve_data(hammer_cursor_t cursor);
 643 int     hammer_ip_delete_record(hammer_cursor_t cursor, hammer_inode_t ip,
 644                         hammer_tid_t tid);
 645 int     hammer_delete_at_cursor(hammer_cursor_t cursor, int64_t *stat_bytes);
 646 int     hammer_ip_check_directory_empty(hammer_transaction_t trans,
 647                         hammer_inode_t ip);
 648 int     hammer_sync_hmp(hammer_mount_t hmp, int waitfor);
 649 int     hammer_queue_inodes_flusher(hammer_mount_t hmp, int waitfor);
 650
 651
 652 hammer_record_t
 653         hammer_alloc_mem_record(hammer_inode_t ip, int data_len);
 654 void    hammer_flush_record_done(hammer_record_t record, int error);
 655 void    hammer_wait_mem_record(hammer_record_t record);
 656 void    hammer_rel_mem_record(hammer_record_t record);
 657
 658 int     hammer_cursor_up(hammer_cursor_t cursor);
 659 int     hammer_cursor_up_locked(hammer_cursor_t cursor);
 660 int     hammer_cursor_down(hammer_cursor_t cursor);
 661 int     hammer_cursor_upgrade(hammer_cursor_t cursor);
 662 void    hammer_cursor_downgrade(hammer_cursor_t cursor);
 663 int     hammer_cursor_seek(hammer_cursor_t cursor, hammer_node_t node,
 664                         int index);
 665 void    hammer_lock_ex(struct hammer_lock *lock);
 666 int     hammer_lock_ex_try(struct hammer_lock *lock);
 667 void    hammer_lock_sh(struct hammer_lock *lock);
 668 int     hammer_lock_upgrade(struct hammer_lock *lock);
 669 void    hammer_lock_downgrade(struct hammer_lock *lock);
 670 void    hammer_unlock(struct hammer_lock *lock);
 671 void    hammer_ref(struct hammer_lock *lock);
 672 void    hammer_unref(struct hammer_lock *lock);
 673
 674 void    hammer_sync_lock_ex(hammer_transaction_t trans);
 675 void    hammer_sync_lock_sh(hammer_transaction_t trans);
 676 void    hammer_sync_unlock(hammer_transaction_t trans);
 677
 678 u_int32_t hammer_to_unix_xid(uuid_t *uuid);
 679 void hammer_guid_to_uuid(uuid_t *uuid, u_int32_t guid);
 680 void    hammer_to_timespec(hammer_tid_t tid, struct timespec *ts);
 681 hammer_tid_t hammer_timespec_to_transid(struct timespec *ts);
 682 hammer_tid_t hammer_now_tid(void);
 683 hammer_tid_t hammer_str_to_tid(const char *str);
 684 hammer_tid_t hammer_alloc_objid(hammer_transaction_t trans, hammer_inode_t dip);
 685 void hammer_clear_objid(hammer_inode_t dip);
 686 void hammer_destroy_objid_cache(hammer_mount_t hmp);
 687
 688 int hammer_enter_undo_history(hammer_mount_t hmp, hammer_off_t offset,
 689                               int bytes);
 690 void hammer_clear_undo_history(hammer_mount_t hmp);
 691 enum vtype hammer_get_vnode_type(u_int8_t obj_type);
 692 int hammer_get_dtype(u_int8_t obj_type);
 693 u_int8_t hammer_get_obj_type(enum vtype vtype);
 694 int64_t hammer_directory_namekey(void *name, int len);
 695 int     hammer_nohistory(hammer_inode_t ip);
 696
 697 int     hammer_init_cursor(hammer_transaction_t trans, hammer_cursor_t cursor,
 698                            struct hammer_node **cache, hammer_inode_t ip);
 699 int     hammer_reinit_cursor(hammer_cursor_t cursor);
 700 void    hammer_normalize_cursor(hammer_cursor_t cursor);
 701 void    hammer_done_cursor(hammer_cursor_t cursor);
 702 void    hammer_mem_done(hammer_cursor_t cursor);
 703
 704 int     hammer_btree_lookup(hammer_cursor_t cursor);
 705 int     hammer_btree_first(hammer_cursor_t cursor);
 706 int     hammer_btree_last(hammer_cursor_t cursor);
 707 int     hammer_btree_extract(hammer_cursor_t cursor, int flags);
 708 int     hammer_btree_iterate(hammer_cursor_t cursor);
 709 int     hammer_btree_iterate_reverse(hammer_cursor_t cursor);
 710 int     hammer_btree_insert(hammer_cursor_t cursor,
 711                             hammer_btree_leaf_elm_t elm);
 712 int     hammer_btree_delete(hammer_cursor_t cursor);
 713 int     hammer_btree_cmp(hammer_base_elm_t key1, hammer_base_elm_t key2);
 714 int     hammer_btree_chkts(hammer_tid_t ts, hammer_base_elm_t key);
 715 int     hammer_btree_correct_rhb(hammer_cursor_t cursor, hammer_tid_t tid);
 716 int     hammer_btree_correct_lhb(hammer_cursor_t cursor, hammer_tid_t tid);
 717
 718 int     btree_set_parent(hammer_transaction_t trans, hammer_node_t node,
 719                         hammer_btree_elm_t elm);
 720 int     hammer_btree_lock_children(hammer_cursor_t cursor,
 721                         struct hammer_node_locklist **locklistp);
 722 void    hammer_btree_unlock_children(struct hammer_node_locklist **locklistp);
 723
 724
 725 void    hammer_print_btree_node(hammer_node_ondisk_t ondisk);
 726 void    hammer_print_btree_elm(hammer_btree_elm_t elm, u_int8_t type, int i);
 727
 728 void    *hammer_bread(struct hammer_mount *hmp, hammer_off_t off,
 729                         int *errorp, struct hammer_buffer **bufferp);
 730 void    *hammer_bnew(struct hammer_mount *hmp, hammer_off_t off,
 731                         int *errorp, struct hammer_buffer **bufferp);
 732
 733 hammer_volume_t hammer_get_root_volume(hammer_mount_t hmp, int *errorp);
 734 int     hammer_dowrite(hammer_cursor_t cursor, hammer_inode_t ip,
 735                         struct bio *bio);
 736
 737 hammer_volume_t hammer_get_volume(hammer_mount_t hmp,
 738                         int32_t vol_no, int *errorp);
 739 hammer_buffer_t hammer_get_buffer(hammer_mount_t hmp,
 740                         hammer_off_t buf_offset, int isnew, int *errorp);
 741 void            hammer_clrxlate_buffer(hammer_mount_t hmp,
 742                         hammer_off_t buf_offset);
 743 void    hammer_uncache_buffer(struct hammer_mount *hmp, hammer_off_t off);
 744
 745 int             hammer_ref_volume(hammer_volume_t volume);
 746 int             hammer_ref_buffer(hammer_buffer_t buffer);
 747 void            hammer_flush_buffer_nodes(hammer_buffer_t buffer);
 748
 749 void            hammer_rel_volume(hammer_volume_t volume, int flush);
 750 void            hammer_rel_buffer(hammer_buffer_t buffer, int flush);
 751
 752 int             hammer_vfs_export(struct mount *mp, int op,
 753                         const struct export_args *export);
 754 hammer_node_t   hammer_get_node(hammer_mount_t hmp, hammer_off_t node_offset,
 755                         int isnew, int *errorp);
 756 void            hammer_ref_node(hammer_node_t node);
 757 hammer_node_t   hammer_ref_node_safe(struct hammer_mount *hmp,
 758                         struct hammer_node **cache, int *errorp);
 759 void            hammer_rel_node(hammer_node_t node);
 760 void            hammer_delete_node(hammer_transaction_t trans,
 761                         hammer_node_t node);
 762 void            hammer_cache_node(hammer_node_t node,
 763                         struct hammer_node **cache);
 764 void            hammer_uncache_node(struct hammer_node **cache);
 765 void            hammer_flush_node(hammer_node_t node);
 766
 767 void hammer_dup_buffer(struct hammer_buffer **bufferp,
 768                         struct hammer_buffer *buffer);
 769 hammer_node_t hammer_alloc_btree(hammer_transaction_t trans, int *errorp);
 770 void *hammer_alloc_record(hammer_transaction_t trans,
 771                         hammer_off_t *rec_offp, u_int16_t rec_type,
 772                         struct hammer_buffer **rec_bufferp,
 773                         int32_t data_len, void **datap,
 774                         hammer_off_t *data_offp,
 775                         struct hammer_buffer **data_bufferp, int *errorp);
 776 void *hammer_alloc_data(hammer_transaction_t trans, int32_t data_len,
 777                         hammer_off_t *data_offsetp,
 778                         struct hammer_buffer **data_bufferp, int *errorp);
 779
 780 int hammer_generate_undo(hammer_transaction_t trans, hammer_io_t io,
 781                         hammer_off_t zone1_offset, void *base, int len);
 782
 783 void hammer_put_volume(struct hammer_volume *volume, int flush);
 784 void hammer_put_buffer(struct hammer_buffer *buffer, int flush);
 785
 786 hammer_off_t hammer_freemap_alloc(hammer_transaction_t trans,
 787                         hammer_off_t owner, int *errorp);
 788 void hammer_freemap_free(hammer_transaction_t trans, hammer_off_t phys_offset,
 789                         hammer_off_t owner, int *errorp);
 790 int hammer_checkspace(hammer_mount_t hmp);
 791 hammer_off_t hammer_blockmap_alloc(hammer_transaction_t trans, int zone,
 792                         int bytes, int *errorp);
 793 void hammer_blockmap_free(hammer_transaction_t trans,
 794                         hammer_off_t bmap_off, int bytes);
 795 int hammer_blockmap_getfree(hammer_mount_t hmp, hammer_off_t bmap_off,
 796                         int *curp, int *errorp);
 797 hammer_off_t hammer_blockmap_lookup(hammer_mount_t hmp, hammer_off_t bmap_off,
 798                         int *errorp);
 799 hammer_off_t hammer_undo_lookup(hammer_mount_t hmp, hammer_off_t bmap_off,
 800                         int *errorp);
 801 int64_t hammer_undo_used(hammer_mount_t hmp);
 802 int64_t hammer_undo_space(hammer_mount_t hmp);
 803 int64_t hammer_undo_max(hammer_mount_t hmp);
 804
 805
 806 void hammer_start_transaction(struct hammer_transaction *trans,
 807                               struct hammer_mount *hmp);
 808 void hammer_simple_transaction(struct hammer_transaction *trans,
 809                               struct hammer_mount *hmp);
 810 void hammer_start_transaction_fls(struct hammer_transaction *trans,
 811                                   struct hammer_mount *hmp);
 812 void hammer_done_transaction(struct hammer_transaction *trans);
 813
 814 void hammer_modify_inode(struct hammer_transaction *trans,
 815                         hammer_inode_t ip, int flags);
 816 void hammer_flush_inode(hammer_inode_t ip, int flags);
 817 void hammer_flush_inode_done(hammer_inode_t ip);
 818 void hammer_wait_inode(hammer_inode_t ip);
 819
 820 int  hammer_create_inode(struct hammer_transaction *trans, struct vattr *vap,
 821                         struct ucred *cred, struct hammer_inode *dip,
 822                         struct hammer_inode **ipp);
 823 void hammer_rel_inode(hammer_inode_t ip, int flush);
 824 int hammer_sync_inode(hammer_inode_t ip);
 825 void hammer_test_inode(hammer_inode_t ip);
 826 void hammer_inode_unloadable_check(hammer_inode_t ip, int getvp);
 827
 828 int  hammer_ip_add_directory(struct hammer_transaction *trans,
 829                         hammer_inode_t dip, struct namecache *ncp,
 830                         hammer_inode_t nip);
 831 int  hammer_ip_del_directory(struct hammer_transaction *trans,
 832                         hammer_cursor_t cursor, hammer_inode_t dip,
 833                         hammer_inode_t ip);
 834 int  hammer_ip_add_record(struct hammer_transaction *trans,
 835                         hammer_record_t record);
 836 int  hammer_ip_delete_range(hammer_cursor_t cursor, hammer_inode_t ip,
 837                         int64_t ran_beg, int64_t ran_end);
 838 int  hammer_ip_delete_range_all(hammer_cursor_t cursor, hammer_inode_t ip,
 839                         int *countp);
 840 int  hammer_ip_sync_data(hammer_cursor_t cursor, hammer_inode_t ip,
 841                         int64_t offset, void *data, int bytes);
 842 int  hammer_ip_sync_record(hammer_transaction_t trans, hammer_record_t rec);
 843 int  hammer_ip_sync_record_cursor(hammer_cursor_t cursor, hammer_record_t rec);
 844
 845 int hammer_ioctl(hammer_inode_t ip, u_long com, caddr_t data, int fflag,
 846                         struct ucred *cred);
 847
 848 void hammer_io_init(hammer_io_t io, hammer_mount_t hmp,
 849                         enum hammer_io_type type);
 850 void hammer_io_reinit(hammer_io_t io, enum hammer_io_type type);
 851 int hammer_io_read(struct vnode *devvp, struct hammer_io *io,
 852                         hammer_off_t limit);
 853 int hammer_io_new(struct vnode *devvp, struct hammer_io *io);
 854 void hammer_io_release(struct hammer_io *io, int flush);
 855 void hammer_io_flush(struct hammer_io *io);
 856 void hammer_io_clear_modify(struct hammer_io *io);
 857 void hammer_io_waitdep(struct hammer_io *io);
 858
 859 void hammer_modify_volume(hammer_transaction_t trans, hammer_volume_t volume,
 860                         void *base, int len);
 861 void hammer_modify_buffer(hammer_transaction_t trans, hammer_buffer_t buffer,
 862                         void *base, int len);
 863 void hammer_modify_volume_done(hammer_volume_t volume);
 864 void hammer_modify_buffer_done(hammer_buffer_t buffer);
 865
 866 int hammer_ioc_reblock(hammer_transaction_t trans, hammer_inode_t ip,
 867                         struct hammer_ioc_reblock *reblock);
 868 int hammer_ioc_prune(hammer_transaction_t trans, hammer_inode_t ip,
 869                         struct hammer_ioc_prune *prune);
 870
 871 void hammer_init_holes(hammer_mount_t hmp, hammer_holes_t holes);
 872 void hammer_free_holes(hammer_mount_t hmp, hammer_holes_t holes);
 873 int hammer_signal_check(hammer_mount_t hmp);
 874
 875 void hammer_flusher_create(hammer_mount_t hmp);
 876 void hammer_flusher_destroy(hammer_mount_t hmp);
 877 void hammer_flusher_sync(hammer_mount_t hmp);
 878 void hammer_flusher_async(hammer_mount_t hmp);
 879
 880 int hammer_recover(hammer_mount_t hmp, hammer_volume_t rootvol);
 881
 882 void hammer_crc_set_blockmap(hammer_blockmap_t blockmap);
 883 void hammer_crc_set_volume(hammer_volume_ondisk_t ondisk);
 884
 885 int hammer_crc_test_blockmap(hammer_blockmap_t blockmap);
 886 int hammer_crc_test_volume(hammer_volume_ondisk_t ondisk);
 887 int hammer_crc_test_btree(hammer_node_ondisk_t ondisk);
 888 void hkprintf(const char *ctl, ...);
 889
 890 #endif
 891
 892 static __inline void
 893 hammer_modify_node_noundo(hammer_transaction_t trans, hammer_node_t node)
 894 {
 895         hammer_modify_buffer(trans, node->buffer, NULL, 0);
 896 }
 897
 898 static __inline void
 899 hammer_modify_node_all(hammer_transaction_t trans, struct hammer_node *node)
 900 {
 901         hammer_modify_buffer(trans, node->buffer,
 902                              node->ondisk, sizeof(*node->ondisk));
 903 }
 904
 905 static __inline void
 906 hammer_modify_node(hammer_transaction_t trans, hammer_node_t node,
 907                    void *base, int len)
 908 {
 909         hammer_crc_t *crcptr;
 910
 911         KKASSERT((char *)base >= (char *)node->ondisk &&
 912                  (char *)base + len <=
 913                     (char *)node->ondisk + sizeof(*node->ondisk));
 914         hammer_modify_buffer(trans, node->buffer, base, len);
 915         crcptr = &node->ondisk->crc;
 916         hammer_modify_buffer(trans, node->buffer, crcptr, sizeof(hammer_crc_t));
 917         --node->buffer->io.modify_refs; /* only want one ref */
 918 }
 919
 920 static __inline void
 921 hammer_modify_node_done(hammer_node_t node)
 922 {
 923         node->ondisk->crc = crc32(&node->ondisk->crc + 1, HAMMER_BTREE_CRCSIZE);
 924         hammer_modify_buffer_done(node->buffer);
 925 }
 926
 927 #define hammer_modify_volume_field(trans, vol, field)           \
 928         hammer_modify_volume(trans, vol, &(vol)->ondisk->field, \
 929                              sizeof((vol)->ondisk->field))
 930
 931 #define hammer_modify_node_field(trans, node, field)            \
 932         hammer_modify_node(trans, node, &(node)->ondisk->field, \
 933                              sizeof((node)->ondisk->field))
 934