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 #include <sys/limits.h>
  61
  62 #include <sys/buf2.h>
  63 #include <sys/signal2.h>
  64 #include <sys/mplock2.h>
  65 #include "hammer_disk.h"
  66 #include "hammer_mount.h"
  67 #include "hammer_ioctl.h"
  68
  69 #if defined(_KERNEL) || defined(_KERNEL_STRUCTURES)
  70
  71 MALLOC_DECLARE(M_HAMMER);
  72
  73 /*
  74  * Kernel trace
  75  */
  76 #if !defined(KTR_HAMMER)
  77 #define KTR_HAMMER      KTR_ALL
  78 #endif
  79 KTR_INFO_MASTER_EXTERN(hammer);
  80
  81 /*
  82  * Misc structures
  83  */
  84 struct hammer_mount;
  85
  86 /*
  87  * Key structure used for custom RB tree inode lookups.  This prototypes
  88  * the function hammer_ino_rb_tree_RB_LOOKUP_INFO(root, info).
  89  */
  90 typedef struct hammer_inode_info {
  91         int64_t         obj_id;         /* (key) object identifier */
  92         hammer_tid_t    obj_asof;       /* (key) snapshot transid or 0 */
  93         u_int32_t       obj_localization; /* (key) pseudo-fs */
  94         union {
  95                 struct hammer_btree_leaf_elm *leaf;
  96         } u;
  97 } *hammer_inode_info_t;
  98
  99 typedef enum hammer_transaction_type {
 100         HAMMER_TRANS_RO,
 101         HAMMER_TRANS_STD,
 102         HAMMER_TRANS_FLS
 103 } hammer_transaction_type_t;
 104
 105 /*
 106  * HAMMER Transaction tracking
 107  */
 108 struct hammer_transaction {
 109         hammer_transaction_type_t type;
 110         struct hammer_mount *hmp;
 111         hammer_tid_t    tid;
 112         u_int64_t       time;
 113         u_int32_t       time32;
 114         int             sync_lock_refs;
 115         int             flags;
 116         struct hammer_volume *rootvol;
 117 };
 118
 119 typedef struct hammer_transaction *hammer_transaction_t;
 120
 121 #define HAMMER_TRANSF_NEWINODE  0x0001
 122 #define HAMMER_TRANSF_DIDIO     0x0002
 123 #define HAMMER_TRANSF_CRCDOM    0x0004  /* EDOM on CRC error, less critical */
 124
 125 /*
 126  * HAMMER locks
 127  */
 128 struct hammer_lock {
 129         int             refs;           /* active references delay writes */
 130         volatile u_int  lockval;        /* lock count and control bits */
 131         struct thread   *owner;         /* owner if exclusively held */
 132 };
 133
 134 #define HAMMER_LOCKF_EXCLUSIVE  0x40000000
 135 #define HAMMER_LOCKF_WANTED     0x80000000
 136
 137 static __inline int
 138 hammer_notlocked(struct hammer_lock *lock)
 139 {
 140         return(lock->lockval == 0);
 141 }
 142
 143 static __inline int
 144 hammer_islocked(struct hammer_lock *lock)
 145 {
 146         return(lock->lockval != 0);
 147 }
 148
 149 static __inline int
 150 hammer_isactive(struct hammer_lock *lock)
 151 {
 152         return(lock->refs != 0);
 153 }
 154
 155 static __inline int
 156 hammer_islastref(struct hammer_lock *lock)
 157 {
 158         return(lock->refs == 1);
 159 }
 160
 161 /*
 162  * Return if we specifically own the lock exclusively.
 163  */
 164 static __inline int
 165 hammer_lock_excl_owned(struct hammer_lock *lock, thread_t td)
 166 {
 167         if ((lock->lockval & HAMMER_LOCKF_EXCLUSIVE) &&
 168             lock->owner == td) {
 169                 return(1);
 170         }
 171         return(0);
 172 }
 173
 174 /*
 175  * Flush state, used by various structures
 176  */
 177 typedef enum hammer_inode_state {
 178         HAMMER_FST_IDLE,
 179         HAMMER_FST_SETUP,
 180         HAMMER_FST_FLUSH
 181 } hammer_inode_state_t;
 182
 183 TAILQ_HEAD(hammer_record_list, hammer_record);
 184
 185 /*
 186  * Pseudo-filesystem extended data tracking
 187  */
 188 struct hammer_pfs_rb_tree;
 189 struct hammer_pseudofs_inmem;
 190 RB_HEAD(hammer_pfs_rb_tree, hammer_pseudofs_inmem);
 191 RB_PROTOTYPE2(hammer_pfs_rb_tree, hammer_pseudofs_inmem, rb_node,
 192               hammer_pfs_rb_compare, u_int32_t);
 193
 194 struct hammer_pseudofs_inmem {
 195         RB_ENTRY(hammer_pseudofs_inmem) rb_node;
 196         struct hammer_lock      lock;
 197         u_int32_t               localization;
 198         hammer_tid_t            create_tid;
 199         int                     flags;
 200         udev_t                  fsid_udev;
 201         struct hammer_pseudofs_data pfsd;
 202 };
 203
 204 typedef struct hammer_pseudofs_inmem *hammer_pseudofs_inmem_t;
 205
 206 #define HAMMER_PFSM_DELETED     0x0001
 207
 208 /*
 209  * Cache object ids.  A fixed number of objid cache structures are
 210  * created to reserve object id's for newly created files in multiples
 211  * of 100,000, localized to a particular directory, and recycled as
 212  * needed.  This allows parallel create operations in different
 213  * directories to retain fairly localized object ids which in turn
 214  * improves reblocking performance and layout.
 215  */
 216 #define OBJID_CACHE_SIZE        1024
 217 #define OBJID_CACHE_BULK_BITS   10              /* 10 bits (1024)       */
 218 #define OBJID_CACHE_BULK        (32 * 32)       /* two level (1024)     */
 219 #define OBJID_CACHE_BULK_MASK   (OBJID_CACHE_BULK - 1)
 220 #define OBJID_CACHE_BULK_MASK64 ((u_int64_t)(OBJID_CACHE_BULK - 1))
 221
 222 typedef struct hammer_objid_cache {
 223         TAILQ_ENTRY(hammer_objid_cache) entry;
 224         struct hammer_inode             *dip;
 225         hammer_tid_t                    base_tid;
 226         int                             count;
 227         u_int32_t                       bm0;
 228         u_int32_t                       bm1[32];
 229 } *hammer_objid_cache_t;
 230
 231 /*
 232  * Associate an inode with a B-Tree node to cache search start positions
 233  */
 234 typedef struct hammer_node_cache {
 235         TAILQ_ENTRY(hammer_node_cache)  entry;
 236         struct hammer_node              *node;
 237         struct hammer_inode             *ip;
 238 } *hammer_node_cache_t;
 239
 240 TAILQ_HEAD(hammer_node_cache_list, hammer_node_cache);
 241
 242 /*
 243  * Structure used to organize flush groups.  Flush groups must be
 244  * organized into chunks in order to avoid blowing out the UNDO FIFO.
 245  * Without this a 'sync' could end up flushing 50,000 inodes in a single
 246  * transaction.
 247  */
 248 struct hammer_fls_rb_tree;
 249 RB_HEAD(hammer_fls_rb_tree, hammer_inode);
 250 RB_PROTOTYPE(hammer_fls_rb_tree, hammer_inode, rb_flsnode,
 251               hammer_ino_rb_compare);
 252
 253 struct hammer_flush_group {
 254         TAILQ_ENTRY(hammer_flush_group) flush_entry;
 255         struct hammer_fls_rb_tree       flush_tree;
 256         int                             unused01;
 257         int                             total_count;    /* record load */
 258         int                             running;        /* group is running */
 259         int                             closed;
 260         int                             refs;
 261 };
 262
 263 typedef struct hammer_flush_group *hammer_flush_group_t;
 264
 265 TAILQ_HEAD(hammer_flush_group_list, hammer_flush_group);
 266
 267 /*
 268  * Structure used to represent an inode in-memory.
 269  *
 270  * The record and data associated with an inode may be out of sync with
 271  * the disk (xDIRTY flags), or not even on the disk at all (ONDISK flag
 272  * clear).
 273  *
 274  * An inode may also hold a cache of unsynchronized records, used for
 275  * database and directories only.  Unsynchronized regular file data is
 276  * stored in the buffer cache.
 277  *
 278  * NOTE: A file which is created and destroyed within the initial
 279  * synchronization period can wind up not doing any disk I/O at all.
 280  *
 281  * Finally, an inode may cache numerous disk-referencing B-Tree cursors.
 282  */
 283 struct hammer_ino_rb_tree;
 284 struct hammer_inode;
 285 RB_HEAD(hammer_ino_rb_tree, hammer_inode);
 286 RB_PROTOTYPEX(hammer_ino_rb_tree, INFO, hammer_inode, rb_node,
 287               hammer_ino_rb_compare, hammer_inode_info_t);
 288
 289 struct hammer_rec_rb_tree;
 290 struct hammer_record;
 291 RB_HEAD(hammer_rec_rb_tree, hammer_record);
 292 RB_PROTOTYPEX(hammer_rec_rb_tree, INFO, hammer_record, rb_node,
 293               hammer_rec_rb_compare, hammer_btree_leaf_elm_t);
 294
 295 TAILQ_HEAD(hammer_node_list, hammer_node);
 296
 297 struct hammer_inode {
 298         RB_ENTRY(hammer_inode)  rb_node;
 299         hammer_inode_state_t    flush_state;
 300         hammer_flush_group_t    flush_group;
 301         RB_ENTRY(hammer_inode)  rb_flsnode;     /* when on flush list */
 302         struct hammer_record_list target_list;  /* target of dependant recs */
 303         int64_t                 obj_id;         /* (key) object identifier */
 304         hammer_tid_t            obj_asof;       /* (key) snapshot or 0 */
 305         u_int32_t               obj_localization; /* (key) pseudo-fs */
 306         struct hammer_mount     *hmp;
 307         hammer_objid_cache_t    objid_cache;
 308         int                     flags;
 309         int                     error;          /* flush error */
 310         int                     cursor_ip_refs; /* sanity */
 311         int                     rsv_recs;
 312         struct vnode            *vp;
 313         hammer_pseudofs_inmem_t pfsm;
 314         struct lockf            advlock;
 315         struct hammer_lock      lock;           /* sync copy interlock */
 316         off_t                   trunc_off;
 317         struct hammer_btree_leaf_elm ino_leaf;  /* in-memory cache */
 318         struct hammer_inode_data ino_data;      /* in-memory cache */
 319         struct hammer_rec_rb_tree rec_tree;     /* in-memory cache */
 320         int                     rec_generation;
 321         struct hammer_node_cache cache[4];      /* search initiate cache */
 322
 323         /*
 324          * When a demark is created to synchronize an inode to
 325          * disk, certain fields are copied so the front-end VOPs
 326          * can continue to run in parallel with the synchronization
 327          * occuring in the background.
 328          */
 329         int             sync_flags;             /* to-sync flags cache */
 330         off_t           sync_trunc_off;         /* to-sync truncation */
 331         off_t           save_trunc_off;         /* write optimization */
 332         struct hammer_btree_leaf_elm sync_ino_leaf; /* to-sync cache */
 333         struct hammer_inode_data sync_ino_data; /* to-sync cache */
 334         size_t          redo_count;
 335 };
 336
 337 typedef struct hammer_inode *hammer_inode_t;
 338
 339 #define VTOI(vp)        ((struct hammer_inode *)(vp)->v_data)
 340
 341 /*
 342  * NOTE: DDIRTY does not include atime or mtime and does not include
 343  *       write-append size changes.  SDIRTY handles write-append size
 344  *       changes.
 345  */
 346                                         /* (not including atime/mtime) */
 347 #define HAMMER_INODE_DDIRTY     0x0001  /* in-memory ino_data is dirty */
 348 #define HAMMER_INODE_RSV_INODES 0x0002  /* hmp->rsv_inodes bumped */
 349 #define HAMMER_INODE_CONN_DOWN  0x0004  /* include in downward recursion */
 350 #define HAMMER_INODE_XDIRTY     0x0008  /* in-memory records */
 351 #define HAMMER_INODE_ONDISK     0x0010  /* inode is on-disk (else not yet) */
 352 #define HAMMER_INODE_FLUSH      0x0020  /* flush on last ref */
 353 #define HAMMER_INODE_DELETED    0x0080  /* inode delete (backend) */
 354 #define HAMMER_INODE_DELONDISK  0x0100  /* delete synchronized to disk */
 355 #define HAMMER_INODE_RO         0x0200  /* read-only (because of as-of) */
 356 #define HAMMER_INODE_VHELD      0x0400  /* vnode held on sync */
 357 #define HAMMER_INODE_DONDISK    0x0800  /* data records may be on disk */
 358 #define HAMMER_INODE_BUFS       0x1000  /* dirty high level bps present */
 359 #define HAMMER_INODE_REFLUSH    0x2000  /* flush on dependancy / reflush */
 360 #define HAMMER_INODE_RECLAIM    0x4000  /* trying to reclaim */
 361 #define HAMMER_INODE_FLUSHW     0x8000  /* Someone waiting for flush */
 362
 363 #define HAMMER_INODE_TRUNCATED  0x00010000
 364 #define HAMMER_INODE_DELETING   0x00020000 /* inode delete request (frontend)*/
 365 #define HAMMER_INODE_RESIGNAL   0x00040000 /* re-signal on re-flush */
 366 #define HAMMER_INODE_ATIME      0x00100000 /* in-memory atime modified */
 367 #define HAMMER_INODE_MTIME      0x00200000 /* in-memory mtime modified */
 368 #define HAMMER_INODE_WOULDBLOCK 0x00400000 /* re-issue to new flush group */
 369 #define HAMMER_INODE_DUMMY      0x00800000 /* dummy inode covering bad file */
 370 #define HAMMER_INODE_SDIRTY     0x01000000 /* in-memory ino_data.size is dirty*/
 371
 372 #define HAMMER_INODE_MODMASK    (HAMMER_INODE_DDIRTY|HAMMER_INODE_SDIRTY|   \
 373                                  HAMMER_INODE_XDIRTY|HAMMER_INODE_BUFS|     \
 374                                  HAMMER_INODE_ATIME|HAMMER_INODE_MTIME|     \
 375                                  HAMMER_INODE_TRUNCATED|HAMMER_INODE_DELETING)
 376
 377 #define HAMMER_INODE_MODMASK_NOXDIRTY   \
 378                                 (HAMMER_INODE_MODMASK & ~HAMMER_INODE_XDIRTY)
 379
 380 #define HAMMER_INODE_MODMASK_NOREDO     \
 381                                 (HAMMER_INODE_DDIRTY|                       \
 382                                  HAMMER_INODE_XDIRTY|                       \
 383                                  HAMMER_INODE_TRUNCATED|HAMMER_INODE_DELETING)
 384
 385 #define HAMMER_FLUSH_SIGNAL     0x0001
 386 #define HAMMER_FLUSH_RECURSION  0x0002
 387
 388 /*
 389  * Used by the inode reclaim code to pipeline reclaims and avoid
 390  * blowing out kernel memory or letting the flusher get too far
 391  * behind.  The reclaim wakes up when count reaches 0 or the
 392  * timer expires.
 393  */
 394 struct hammer_reclaim {
 395         TAILQ_ENTRY(hammer_reclaim) entry;
 396         int     count;
 397 };
 398
 399 #define HAMMER_RECLAIM_WAIT     4000    /* default vfs.hammer.limit_reclaim */
 400
 401 /*
 402  * Structure used to represent an unsynchronized record in-memory.  These
 403  * records typically represent directory entries.  Only non-historical
 404  * records are kept in-memory.
 405  *
 406  * Records are organized as a per-inode RB-Tree.  If the inode is not
 407  * on disk then neither are any records and the in-memory record tree
 408  * represents the entire contents of the inode.  If the inode is on disk
 409  * then the on-disk B-Tree is scanned in parallel with the in-memory
 410  * RB-Tree to synthesize the current state of the file.
 411  *
 412  * Records are also used to enforce the ordering of directory create/delete
 413  * operations.  A new inode will not be flushed to disk unless its related
 414  * directory entry is also being flushed at the same time.  A directory entry
 415  * will not be removed unless its related inode is also being removed at the
 416  * same time.
 417  */
 418 typedef enum hammer_record_type {
 419         HAMMER_MEM_RECORD_GENERAL,      /* misc record */
 420         HAMMER_MEM_RECORD_INODE,        /* inode record */
 421         HAMMER_MEM_RECORD_ADD,          /* positive memory cache record */
 422         HAMMER_MEM_RECORD_DEL,          /* negative delete-on-disk record */
 423         HAMMER_MEM_RECORD_DATA          /* bulk-data record w/on-disk ref */
 424 } hammer_record_type_t;
 425
 426 struct hammer_record {
 427         RB_ENTRY(hammer_record)         rb_node;
 428         TAILQ_ENTRY(hammer_record)      target_entry;
 429         hammer_inode_state_t            flush_state;
 430         hammer_flush_group_t            flush_group;
 431         hammer_record_type_t            type;
 432         struct hammer_lock              lock;
 433         struct hammer_reserve           *resv;
 434         struct hammer_inode             *ip;
 435         struct hammer_inode             *target_ip;
 436         struct hammer_btree_leaf_elm    leaf;
 437         union hammer_data_ondisk        *data;
 438         int                             flags;
 439         hammer_off_t                    zone2_offset;   /* direct-write only */
 440 };
 441
 442 typedef struct hammer_record *hammer_record_t;
 443
 444 /*
 445  * Record flags.  Note that FE can only be set by the frontend if the
 446  * record has not been interlocked by the backend w/ BE.
 447  */
 448 #define HAMMER_RECF_ALLOCDATA           0x0001
 449 #define HAMMER_RECF_ONRBTREE            0x0002
 450 #define HAMMER_RECF_DELETED_FE          0x0004  /* deleted (frontend) */
 451 #define HAMMER_RECF_DELETED_BE          0x0008  /* deleted (backend) */
 452 #define HAMMER_RECF_COMMITTED           0x0010  /* committed to the B-Tree */
 453 #define HAMMER_RECF_INTERLOCK_BE        0x0020  /* backend interlock */
 454 #define HAMMER_RECF_WANTED              0x0040  /* wanted by the frontend */
 455 #define HAMMER_RECF_CONVERT_DELETE      0x0100  /* special case */
 456 #define HAMMER_RECF_DIRECT_IO           0x0200  /* related direct I/O running*/
 457 #define HAMMER_RECF_DIRECT_WAIT         0x0400  /* related direct I/O running*/
 458 #define HAMMER_RECF_DIRECT_INVAL        0x0800  /* buffer alias invalidation */
 459
 460 /*
 461  * hammer_create_at_cursor() and hammer_delete_at_cursor() flags.
 462  */
 463 #define HAMMER_CREATE_MODE_UMIRROR      0x0001
 464 #define HAMMER_CREATE_MODE_SYS          0x0002
 465
 466 #define HAMMER_DELETE_ADJUST            0x0001
 467 #define HAMMER_DELETE_DESTROY           0x0002
 468
 469 /*
 470  * In-memory structures representing on-disk structures.
 471  */
 472 struct hammer_volume;
 473 struct hammer_buffer;
 474 struct hammer_node;
 475 struct hammer_undo;
 476 struct hammer_reserve;
 477
 478 RB_HEAD(hammer_vol_rb_tree, hammer_volume);
 479 RB_HEAD(hammer_buf_rb_tree, hammer_buffer);
 480 RB_HEAD(hammer_nod_rb_tree, hammer_node);
 481 RB_HEAD(hammer_und_rb_tree, hammer_undo);
 482 RB_HEAD(hammer_res_rb_tree, hammer_reserve);
 483
 484 RB_PROTOTYPE2(hammer_vol_rb_tree, hammer_volume, rb_node,
 485               hammer_vol_rb_compare, int32_t);
 486 RB_PROTOTYPE2(hammer_buf_rb_tree, hammer_buffer, rb_node,
 487               hammer_buf_rb_compare, hammer_off_t);
 488 RB_PROTOTYPE2(hammer_nod_rb_tree, hammer_node, rb_node,
 489               hammer_nod_rb_compare, hammer_off_t);
 490 RB_PROTOTYPE2(hammer_und_rb_tree, hammer_undo, rb_node,
 491               hammer_und_rb_compare, hammer_off_t);
 492 RB_PROTOTYPE2(hammer_res_rb_tree, hammer_reserve, rb_node,
 493               hammer_res_rb_compare, hammer_off_t);
 494
 495 /*
 496  * IO management - embedded at the head of various in-memory structures
 497  *
 498  * VOLUME       - hammer_volume containing meta-data
 499  * META_BUFFER  - hammer_buffer containing meta-data
 500  * DATA_BUFFER  - hammer_buffer containing pure-data
 501  *
 502  * Dirty volume headers and dirty meta-data buffers are locked until the
 503  * flusher can sequence them out.  Dirty pure-data buffers can be written.
 504  * Clean buffers can be passively released.
 505  */
 506 typedef enum hammer_io_type {
 507         HAMMER_STRUCTURE_VOLUME,
 508         HAMMER_STRUCTURE_META_BUFFER,
 509         HAMMER_STRUCTURE_UNDO_BUFFER,
 510         HAMMER_STRUCTURE_DATA_BUFFER
 511 } hammer_io_type_t;
 512
 513 union hammer_io_structure;
 514 struct hammer_io;
 515
 516 struct worklist {
 517         LIST_ENTRY(worklist) node;
 518 };
 519
 520 TAILQ_HEAD(hammer_io_list, hammer_io);
 521 typedef struct hammer_io_list *hammer_io_list_t;
 522
 523 struct hammer_io {
 524         struct worklist         worklist;
 525         struct hammer_lock      lock;
 526         enum hammer_io_type     type;
 527         struct hammer_mount     *hmp;
 528         struct hammer_volume    *volume;
 529         TAILQ_ENTRY(hammer_io)  mod_entry; /* list entry if modified */
 530         hammer_io_list_t        mod_list;
 531         struct buf              *bp;
 532         int64_t                 offset;    /* zone-2 offset */
 533         int                     bytes;     /* buffer cache buffer size */
 534         int                     loading;   /* loading/unloading interlock */
 535         int                     modify_refs;
 536
 537         u_int           modified : 1;   /* bp's data was modified */
 538         u_int           released : 1;   /* bp released (w/ B_LOCKED set) */
 539         u_int           running : 1;    /* bp write IO in progress */
 540         u_int           waiting : 1;    /* someone is waiting on us */
 541         u_int           validated : 1;  /* ondisk has been validated */
 542         u_int           waitdep : 1;    /* flush waits for dependancies */
 543         u_int           recovered : 1;  /* has recovery ref */
 544         u_int           waitmod : 1;    /* waiting for modify_refs */
 545         u_int           reclaim : 1;    /* reclaim requested */
 546         u_int           gencrc : 1;     /* crc needs to be generated */
 547         u_int           ioerror : 1;    /* abort on io-error */
 548 };
 549
 550 typedef struct hammer_io *hammer_io_t;
 551
 552 #define HAMMER_CLUSTER_SIZE     (64 * 1024)
 553 #if HAMMER_CLUSTER_SIZE > MAXBSIZE
 554 #undef  HAMMER_CLUSTER_SIZE
 555 #define HAMMER_CLUSTER_SIZE     MAXBSIZE
 556 #endif
 557 #define HAMMER_CLUSTER_BUFS     (HAMMER_CLUSTER_SIZE / HAMMER_BUFSIZE)
 558
 559 /*
 560  * In-memory volume representing on-disk buffer
 561  */
 562 struct hammer_volume {
 563         struct hammer_io io;
 564         RB_ENTRY(hammer_volume) rb_node;
 565         struct hammer_volume_ondisk *ondisk;
 566         int32_t vol_no;
 567         int64_t nblocks;        /* note: special calculation for statfs */
 568         int64_t buffer_base;    /* base offset of buffer 0 */
 569         hammer_off_t maxbuf_off; /* Maximum buffer offset (zone-2) */
 570         hammer_off_t maxraw_off; /* Maximum raw offset for device */
 571         char    *vol_name;
 572         struct vnode *devvp;
 573         int     vol_flags;
 574 };
 575
 576 typedef struct hammer_volume *hammer_volume_t;
 577
 578 /*
 579  * In-memory buffer (other then volume, super-cluster, or cluster),
 580  * representing an on-disk buffer.
 581  */
 582 struct hammer_buffer {
 583         struct hammer_io io;
 584         RB_ENTRY(hammer_buffer) rb_node;
 585         void *ondisk;
 586         hammer_off_t zoneX_offset;
 587         hammer_off_t zone2_offset;
 588         struct hammer_reserve *resv;
 589         struct hammer_node_list clist;
 590 };
 591
 592 typedef struct hammer_buffer *hammer_buffer_t;
 593
 594 /*
 595  * In-memory B-Tree node, representing an on-disk B-Tree node.
 596  *
 597  * This is a hang-on structure which is backed by a hammer_buffer,
 598  * indexed by a hammer_cluster, and used for fine-grained locking of
 599  * B-Tree nodes in order to properly control lock ordering.  A hammer_buffer
 600  * can contain multiple nodes representing wildly disassociated portions
 601  * of the B-Tree so locking cannot be done on a buffer-by-buffer basis.
 602  *
 603  * This structure uses a cluster-relative index to reduce the number
 604  * of layers required to access it, and also because all on-disk B-Tree
 605  * references are cluster-relative offsets.
 606  */
 607 struct hammer_node {
 608         struct hammer_lock      lock;           /* node-by-node lock */
 609         TAILQ_ENTRY(hammer_node) entry;         /* per-buffer linkage */
 610         RB_ENTRY(hammer_node)   rb_node;        /* per-cluster linkage */
 611         hammer_off_t            node_offset;    /* full offset spec */
 612         struct hammer_mount     *hmp;
 613         struct hammer_buffer    *buffer;        /* backing buffer */
 614         hammer_node_ondisk_t    ondisk;         /* ptr to on-disk structure */
 615         TAILQ_HEAD(, hammer_cursor) cursor_list;  /* deadlock recovery */
 616         struct hammer_node_cache_list cache_list; /* passive caches */
 617         int                     flags;
 618         int                     loading;        /* load interlock */
 619 };
 620
 621 #define HAMMER_NODE_DELETED     0x0001
 622 #define HAMMER_NODE_FLUSH       0x0002
 623 #define HAMMER_NODE_CRCGOOD     0x0004
 624 #define HAMMER_NODE_NEEDSCRC    0x0008
 625 #define HAMMER_NODE_NEEDSMIRROR 0x0010
 626 #define HAMMER_NODE_CRCBAD      0x0020
 627
 628 #define HAMMER_NODE_CRCANY      (HAMMER_NODE_CRCGOOD | HAMMER_NODE_CRCBAD)
 629
 630 typedef struct hammer_node      *hammer_node_t;
 631
 632 /*
 633  * List of locked nodes.  This structure is used to lock potentially large
 634  * numbers of nodes as an aid for complex B-Tree operations.
 635  */
 636 struct hammer_node_lock;
 637 TAILQ_HEAD(hammer_node_lock_list, hammer_node_lock);
 638
 639 struct hammer_node_lock {
 640         TAILQ_ENTRY(hammer_node_lock) entry;
 641         struct hammer_node_lock_list  list;
 642         struct hammer_node_lock       *parent;
 643         hammer_node_t   node;
 644         hammer_node_ondisk_t copy;      /* copy of on-disk data */
 645         int             index;          /* index of this node in parent */
 646         int             count;          /* count children */
 647         int             flags;
 648 };
 649
 650 typedef struct hammer_node_lock *hammer_node_lock_t;
 651
 652 #define HAMMER_NODE_LOCK_UPDATED        0x0001
 653
 654 /*
 655  * Common I/O management structure - embedded in in-memory structures
 656  * which are backed by filesystem buffers.
 657  */
 658 union hammer_io_structure {
 659         struct hammer_io        io;
 660         struct hammer_volume    volume;
 661         struct hammer_buffer    buffer;
 662 };
 663
 664 typedef union hammer_io_structure *hammer_io_structure_t;
 665
 666 /*
 667  * The reserve structure prevents the blockmap from allocating
 668  * out of a reserved bigblock.  Such reservations are used by
 669  * the direct-write mechanism.
 670  *
 671  * The structure is also used to hold off on reallocations of
 672  * big blocks from the freemap until flush dependancies have
 673  * been dealt with.
 674  */
 675 struct hammer_reserve {
 676         RB_ENTRY(hammer_reserve) rb_node;
 677         TAILQ_ENTRY(hammer_reserve) delay_entry;
 678         int             flush_group;
 679         int             flags;
 680         int             refs;
 681         int             zone;
 682         int             append_off;
 683         hammer_off_t    zone_offset;
 684 };
 685
 686 typedef struct hammer_reserve *hammer_reserve_t;
 687
 688 #define HAMMER_RESF_ONDELAY     0x0001
 689 #define HAMMER_RESF_LAYER2FREE  0x0002
 690
 691 #include "hammer_cursor.h"
 692
 693 /*
 694  * The undo structure tracks recent undos to avoid laying down duplicate
 695  * undos within a flush group, saving us a significant amount of overhead.
 696  *
 697  * This is strictly a heuristic.
 698  */
 699 #define HAMMER_MAX_UNDOS                1024
 700 #define HAMMER_MAX_FLUSHERS             4
 701
 702 struct hammer_undo {
 703         RB_ENTRY(hammer_undo)   rb_node;
 704         TAILQ_ENTRY(hammer_undo) lru_entry;
 705         hammer_off_t            offset;
 706         int                     bytes;
 707 };
 708
 709 typedef struct hammer_undo *hammer_undo_t;
 710
 711 struct hammer_flusher_info;
 712 TAILQ_HEAD(hammer_flusher_info_list, hammer_flusher_info);
 713
 714 struct hammer_flusher {
 715         int             signal;         /* flusher thread sequencer */
 716         int             act;            /* currently active flush group */
 717         int             done;           /* set to act when complete */
 718         int             next;           /* next flush group */
 719         int             group_lock;     /* lock sequencing of the next flush */
 720         int             exiting;        /* request master exit */
 721         thread_t        td;             /* master flusher thread */
 722         hammer_tid_t    tid;            /* last flushed transaction id */
 723         int             finalize_want;          /* serialize finalization */
 724         struct hammer_lock finalize_lock;       /* serialize finalization */
 725         struct hammer_transaction trans;        /* shared transaction */
 726         struct hammer_flusher_info_list run_list;
 727         struct hammer_flusher_info_list ready_list;
 728 };
 729
 730 #define HAMMER_FLUSH_UNDOS_RELAXED      0
 731 #define HAMMER_FLUSH_UNDOS_FORCED       1
 732 #define HAMMER_FLUSH_UNDOS_AUTO         2
 733 /*
 734  * Internal hammer mount data structure
 735  */
 736 struct hammer_mount {
 737         struct mount *mp;
 738         /*struct vnode *rootvp;*/
 739         struct hammer_ino_rb_tree rb_inos_root;
 740         struct hammer_vol_rb_tree rb_vols_root;
 741         struct hammer_nod_rb_tree rb_nods_root;
 742         struct hammer_und_rb_tree rb_undo_root;
 743         struct hammer_res_rb_tree rb_resv_root;
 744         struct hammer_buf_rb_tree rb_bufs_root;
 745         struct hammer_pfs_rb_tree rb_pfsm_root;
 746         struct hammer_volume *rootvol;
 747         struct hammer_base_elm root_btree_beg;
 748         struct hammer_base_elm root_btree_end;
 749
 750         struct malloc_type      *m_misc;
 751         struct malloc_type      *m_inodes;
 752
 753         int     flags;          /* HAMMER_MOUNT_xxx flags */
 754         int     hflags;
 755         int     ronly;
 756         int     nvolumes;
 757         int     volume_iterator;
 758         int     master_id;      /* -1 or 0-15 - clustering and mirroring */
 759         int     version;        /* hammer filesystem version to use */
 760         int     rsv_inodes;     /* reserved space due to dirty inodes */
 761         int64_t rsv_databytes;  /* reserved space due to record data */
 762         int     rsv_recs;       /* reserved space due to dirty records */
 763         int     rsv_fromdelay;  /* bigblocks reserved due to flush delay */
 764         int     undo_rec_limit; /* based on size of undo area */
 765         int     last_newrecords;
 766         int     count_newrecords;
 767
 768         int     volume_to_remove; /* volume that is currently being removed */
 769
 770         int     inode_reclaims; /* inodes pending reclaim by flusher */
 771         int     count_inodes;   /* total number of inodes */
 772         int     count_iqueued;  /* inodes queued to flusher */
 773
 774         struct hammer_flusher flusher;
 775
 776         u_int   check_interrupt;
 777         u_int   check_yield;
 778         uuid_t  fsid;
 779         struct hammer_io_list volu_list;        /* dirty undo buffers */
 780         struct hammer_io_list undo_list;        /* dirty undo buffers */
 781         struct hammer_io_list data_list;        /* dirty data buffers */
 782         struct hammer_io_list alt_data_list;    /* dirty data buffers */
 783         struct hammer_io_list meta_list;        /* dirty meta bufs    */
 784         struct hammer_io_list lose_list;        /* loose buffers      */
 785         int     locked_dirty_space;             /* meta/volu count    */
 786         int     io_running_space;
 787         int     objid_cache_count;
 788         int     error;                          /* critical I/O error */
 789         struct krate    krate;                  /* rate limited kprintf */
 790         hammer_tid_t    asof;                   /* snapshot mount */
 791         hammer_tid_t    next_tid;
 792         hammer_tid_t    flush_tid1;             /* flusher tid sequencing */
 793         hammer_tid_t    flush_tid2;             /* flusher tid sequencing */
 794         int64_t copy_stat_freebigblocks;        /* number of free bigblocks */
 795         u_int32_t       undo_seqno;             /* UNDO/REDO FIFO seqno */
 796
 797         struct netexport export;
 798         struct hammer_lock sync_lock;
 799         struct hammer_lock free_lock;
 800         struct hammer_lock undo_lock;
 801         struct hammer_lock blkmap_lock;
 802         struct hammer_lock snapshot_lock;
 803         struct hammer_lock volume_lock;
 804         struct hammer_blockmap  blockmap[HAMMER_MAX_ZONES];
 805         struct hammer_undo      undos[HAMMER_MAX_UNDOS];
 806         int                     undo_alloc;
 807         TAILQ_HEAD(, hammer_undo)  undo_lru_list;
 808         TAILQ_HEAD(, hammer_reserve) delay_list;
 809         struct hammer_flush_group_list  flush_group_list;
 810         hammer_flush_group_t    next_flush_group;
 811         TAILQ_HEAD(, hammer_objid_cache) objid_cache_list;
 812         TAILQ_HEAD(, hammer_reclaim) reclaim_list;
 813 };
 814
 815 typedef struct hammer_mount     *hammer_mount_t;
 816
 817 #define HAMMER_MOUNT_CRITICAL_ERROR     0x0001
 818 #define HAMMER_MOUNT_FLUSH_RECOVERY     0x0002
 819
 820 struct hammer_sync_info {
 821         int error;
 822         int waitfor;
 823 };
 824
 825 #endif
 826
 827 /*
 828  * checkspace slop (8MB chunks), higher numbers are more conservative.
 829  */
 830 #define HAMMER_CHKSPC_REBLOCK   25
 831 #define HAMMER_CHKSPC_MIRROR    20
 832 #define HAMMER_CHKSPC_WRITE     20
 833 #define HAMMER_CHKSPC_CREATE    20
 834 #define HAMMER_CHKSPC_REMOVE    10
 835 #define HAMMER_CHKSPC_EMERGENCY 0
 836
 837 #if defined(_KERNEL)
 838
 839 extern struct vop_ops hammer_vnode_vops;
 840 extern struct vop_ops hammer_spec_vops;
 841 extern struct vop_ops hammer_fifo_vops;
 842 extern struct bio_ops hammer_bioops;
 843
 844 extern int hammer_debug_io;
 845 extern int hammer_debug_general;
 846 extern int hammer_debug_debug;
 847 extern int hammer_debug_inode;
 848 extern int hammer_debug_locks;
 849 extern int hammer_debug_btree;
 850 extern int hammer_debug_tid;
 851 extern int hammer_debug_recover;
 852 extern int hammer_debug_recover_faults;
 853 extern int hammer_debug_critical;
 854 extern int hammer_cluster_enable;
 855 extern int hammer_count_fsyncs;
 856 extern int hammer_count_inodes;
 857 extern int hammer_count_iqueued;
 858 extern int hammer_count_reclaiming;
 859 extern int hammer_count_records;
 860 extern int hammer_count_record_datas;
 861 extern int hammer_count_volumes;
 862 extern int hammer_count_buffers;
 863 extern int hammer_count_nodes;
 864 extern int64_t hammer_count_extra_space_used;
 865 extern int64_t hammer_stats_btree_lookups;
 866 extern int64_t hammer_stats_btree_searches;
 867 extern int64_t hammer_stats_btree_inserts;
 868 extern int64_t hammer_stats_btree_deletes;
 869 extern int64_t hammer_stats_btree_elements;
 870 extern int64_t hammer_stats_btree_splits;
 871 extern int64_t hammer_stats_btree_iterations;
 872 extern int64_t hammer_stats_btree_root_iterations;
 873 extern int64_t hammer_stats_record_iterations;
 874 extern int64_t hammer_stats_file_read;
 875 extern int64_t hammer_stats_file_write;
 876 extern int64_t hammer_stats_file_iopsr;
 877 extern int64_t hammer_stats_file_iopsw;
 878 extern int64_t hammer_stats_disk_read;
 879 extern int64_t hammer_stats_disk_write;
 880 extern int64_t hammer_stats_inode_flushes;
 881 extern int64_t hammer_stats_commits;
 882 extern int64_t hammer_stats_undo;
 883 extern int64_t hammer_stats_redo;
 884 extern int hammer_count_dirtybufspace;
 885 extern int hammer_count_refedbufs;
 886 extern int hammer_count_reservations;
 887 extern int hammer_count_io_running_read;
 888 extern int hammer_count_io_running_write;
 889 extern int hammer_count_io_locked;
 890 extern int hammer_limit_dirtybufspace;
 891 extern int hammer_limit_recs;
 892 extern int hammer_limit_inode_recs;
 893 extern int hammer_limit_reclaim;
 894 extern int hammer_limit_redo;
 895 extern int hammer_bio_count;
 896 extern int hammer_verify_zone;
 897 extern int hammer_verify_data;
 898 extern int hammer_write_mode;
 899 extern int hammer_yield_check;
 900 extern int hammer_fsync_mode;
 901 extern int hammer_autoflush;
 902 extern int64_t hammer_contention_count;
 903
 904 void    hammer_critical_error(hammer_mount_t hmp, hammer_inode_t ip,
 905                         int error, const char *msg);
 906 int     hammer_vop_inactive(struct vop_inactive_args *);
 907 int     hammer_vop_reclaim(struct vop_reclaim_args *);
 908 int     hammer_get_vnode(struct hammer_inode *ip, struct vnode **vpp);
 909 struct hammer_inode *hammer_get_inode(hammer_transaction_t trans,
 910                         hammer_inode_t dip, int64_t obj_id,
 911                         hammer_tid_t asof, u_int32_t localization,
 912                         int flags, int *errorp);
 913 struct hammer_inode *hammer_get_dummy_inode(hammer_transaction_t trans,
 914                         hammer_inode_t dip, int64_t obj_id,
 915                         hammer_tid_t asof, u_int32_t localization,
 916                         int flags, int *errorp);
 917 struct hammer_inode *hammer_find_inode(hammer_transaction_t trans,
 918                         int64_t obj_id, hammer_tid_t asof,
 919                         u_int32_t localization);
 920 void    hammer_scan_inode_snapshots(hammer_mount_t hmp,
 921                         hammer_inode_info_t iinfo,
 922                         int (*callback)(hammer_inode_t ip, void *data),
 923                         void *data);
 924 void    hammer_put_inode(struct hammer_inode *ip);
 925 void    hammer_put_inode_ref(struct hammer_inode *ip);
 926 void    hammer_inode_waitreclaims(hammer_mount_t hmp);
 927 void    hammer_inode_waithard(hammer_mount_t hmp);
 928
 929 int     hammer_unload_volume(hammer_volume_t volume, void *data __unused);
 930 int     hammer_adjust_volume_mode(hammer_volume_t volume, void *data __unused);
 931
 932 int     hammer_unload_buffer(hammer_buffer_t buffer, void *data);
 933 int     hammer_install_volume(hammer_mount_t hmp, const char *volname,
 934                         struct vnode *devvp);
 935 int     hammer_mountcheck_volumes(hammer_mount_t hmp);
 936
 937 int     hammer_mem_add(hammer_record_t record);
 938 int     hammer_ip_lookup(hammer_cursor_t cursor);
 939 int     hammer_ip_first(hammer_cursor_t cursor);
 940 int     hammer_ip_next(hammer_cursor_t cursor);
 941 int     hammer_ip_resolve_data(hammer_cursor_t cursor);
 942 int     hammer_ip_delete_record(hammer_cursor_t cursor, hammer_inode_t ip,
 943                         hammer_tid_t tid);
 944 int     hammer_create_at_cursor(hammer_cursor_t cursor,
 945                         hammer_btree_leaf_elm_t leaf, void *udata, int mode);
 946 int     hammer_delete_at_cursor(hammer_cursor_t cursor, int delete_flags,
 947                         hammer_tid_t delete_tid, u_int32_t delete_ts,
 948                         int track, int64_t *stat_bytes);
 949 int     hammer_ip_check_directory_empty(hammer_transaction_t trans,
 950                         hammer_inode_t ip);
 951 int     hammer_sync_hmp(hammer_mount_t hmp, int waitfor);
 952 int     hammer_queue_inodes_flusher(hammer_mount_t hmp, int waitfor);
 953
 954 hammer_record_t
 955         hammer_alloc_mem_record(hammer_inode_t ip, int data_len);
 956 void    hammer_flush_record_done(hammer_record_t record, int error);
 957 void    hammer_wait_mem_record_ident(hammer_record_t record, const char *ident);
 958 void    hammer_rel_mem_record(hammer_record_t record);
 959
 960 int     hammer_cursor_up(hammer_cursor_t cursor);
 961 int     hammer_cursor_up_locked(hammer_cursor_t cursor);
 962 int     hammer_cursor_down(hammer_cursor_t cursor);
 963 int     hammer_cursor_upgrade(hammer_cursor_t cursor);
 964 int     hammer_cursor_upgrade_node(hammer_cursor_t cursor);
 965 void    hammer_cursor_downgrade(hammer_cursor_t cursor);
 966 int     hammer_cursor_seek(hammer_cursor_t cursor, hammer_node_t node,
 967                         int index);
 968 void    hammer_lock_ex_ident(struct hammer_lock *lock, const char *ident);
 969 int     hammer_lock_ex_try(struct hammer_lock *lock);
 970 void    hammer_lock_sh(struct hammer_lock *lock);
 971 int     hammer_lock_sh_try(struct hammer_lock *lock);
 972 int     hammer_lock_upgrade(struct hammer_lock *lock);
 973 void    hammer_lock_downgrade(struct hammer_lock *lock);
 974 int     hammer_lock_status(struct hammer_lock *lock);
 975 void    hammer_unlock(struct hammer_lock *lock);
 976 void    hammer_ref(struct hammer_lock *lock);
 977 void    hammer_unref(struct hammer_lock *lock);
 978
 979 void    hammer_sync_lock_ex(hammer_transaction_t trans);
 980 void    hammer_sync_lock_sh(hammer_transaction_t trans);
 981 int     hammer_sync_lock_sh_try(hammer_transaction_t trans);
 982 void    hammer_sync_unlock(hammer_transaction_t trans);
 983
 984 u_int32_t hammer_to_unix_xid(uuid_t *uuid);
 985 void hammer_guid_to_uuid(uuid_t *uuid, u_int32_t guid);
 986 void    hammer_time_to_timespec(u_int64_t xtime, struct timespec *ts);
 987 u_int64_t hammer_timespec_to_time(struct timespec *ts);
 988 int     hammer_str_to_tid(const char *str, int *ispfsp,
 989                         hammer_tid_t *tidp, u_int32_t *localizationp);
 990 int     hammer_is_atatext(const char *name, int len);
 991 hammer_tid_t hammer_alloc_objid(hammer_mount_t hmp, hammer_inode_t dip,
 992                         int64_t namekey);
 993 void hammer_clear_objid(hammer_inode_t dip);
 994 void hammer_destroy_objid_cache(hammer_mount_t hmp);
 995
 996 int hammer_enter_undo_history(hammer_mount_t hmp, hammer_off_t offset,
 997                         int bytes);
 998 void hammer_clear_undo_history(hammer_mount_t hmp);
 999 enum vtype hammer_get_vnode_type(u_int8_t obj_type);
1000 int hammer_get_dtype(u_int8_t obj_type);
1001 u_int8_t hammer_get_obj_type(enum vtype vtype);
1002 int64_t hammer_directory_namekey(hammer_inode_t dip, const void *name, int len,
1003                         u_int32_t *max_iterationsp);
1004 int     hammer_nohistory(hammer_inode_t ip);
1005
1006 int     hammer_init_cursor(hammer_transaction_t trans, hammer_cursor_t cursor,
1007                         hammer_node_cache_t cache, hammer_inode_t ip);
1008 void    hammer_normalize_cursor(hammer_cursor_t cursor);
1009 void    hammer_done_cursor(hammer_cursor_t cursor);
1010 int     hammer_recover_cursor(hammer_cursor_t cursor);
1011 void    hammer_unlock_cursor(hammer_cursor_t cursor);
1012 int     hammer_lock_cursor(hammer_cursor_t cursor);
1013 hammer_cursor_t hammer_push_cursor(hammer_cursor_t ocursor);
1014 void    hammer_pop_cursor(hammer_cursor_t ocursor, hammer_cursor_t ncursor);
1015
1016 void    hammer_cursor_replaced_node(hammer_node_t onode, hammer_node_t nnode);
1017 void    hammer_cursor_removed_node(hammer_node_t onode, hammer_node_t parent,
1018                         int index);
1019 void    hammer_cursor_split_node(hammer_node_t onode, hammer_node_t nnode,
1020                         int index);
1021 void    hammer_cursor_moved_element(hammer_node_t onode, hammer_node_t nnode,
1022                         int oindex, int nindex);
1023 void    hammer_cursor_parent_changed(hammer_node_t node, hammer_node_t oparent,
1024                         hammer_node_t nparent, int nindex);
1025 void    hammer_cursor_inserted_element(hammer_node_t node, int index);
1026 void    hammer_cursor_deleted_element(hammer_node_t node, int index);
1027
1028 int     hammer_btree_lookup(hammer_cursor_t cursor);
1029 int     hammer_btree_first(hammer_cursor_t cursor);
1030 int     hammer_btree_last(hammer_cursor_t cursor);
1031 int     hammer_btree_extract(hammer_cursor_t cursor, int flags);
1032 int     hammer_btree_iterate(hammer_cursor_t cursor);
1033 int     hammer_btree_iterate_reverse(hammer_cursor_t cursor);
1034 int     hammer_btree_insert(hammer_cursor_t cursor,
1035                             hammer_btree_leaf_elm_t elm, int *doprop);
1036 int     hammer_btree_delete(hammer_cursor_t cursor);
1037 void    hammer_btree_do_propagation(hammer_cursor_t cursor,
1038                             hammer_pseudofs_inmem_t pfsm,
1039                             hammer_btree_leaf_elm_t leaf);
1040 int     hammer_btree_cmp(hammer_base_elm_t key1, hammer_base_elm_t key2);
1041 int     hammer_btree_chkts(hammer_tid_t ts, hammer_base_elm_t key);
1042 int     hammer_btree_correct_rhb(hammer_cursor_t cursor, hammer_tid_t tid);
1043 int     hammer_btree_correct_lhb(hammer_cursor_t cursor, hammer_tid_t tid);
1044
1045 int     btree_set_parent(hammer_transaction_t trans, hammer_node_t node,
1046                         hammer_btree_elm_t elm);
1047 void    hammer_node_lock_init(hammer_node_lock_t parent, hammer_node_t node);
1048 int     hammer_btree_lock_children(hammer_cursor_t cursor, int depth,
1049                         hammer_node_lock_t parent);
1050 void    hammer_btree_lock_copy(hammer_cursor_t cursor,
1051                         hammer_node_lock_t parent);
1052 int     hammer_btree_sync_copy(hammer_cursor_t cursor,
1053                         hammer_node_lock_t parent);
1054 void    hammer_btree_unlock_children(hammer_cursor_t cursor,
1055                         hammer_node_lock_t parent);
1056 int     hammer_btree_search_node(hammer_base_elm_t elm, hammer_node_ondisk_t node);
1057 hammer_node_t hammer_btree_get_parent(hammer_transaction_t trans,
1058                         hammer_node_t node, int *parent_indexp,
1059                         int *errorp, int try_exclusive);
1060
1061 void    hammer_print_btree_node(hammer_node_ondisk_t ondisk);
1062 void    hammer_print_btree_elm(hammer_btree_elm_t elm, u_int8_t type, int i);
1063
1064 void    *hammer_bread(struct hammer_mount *hmp, hammer_off_t off,
1065                         int *errorp, struct hammer_buffer **bufferp);
1066 void    *hammer_bnew(struct hammer_mount *hmp, hammer_off_t off,
1067                         int *errorp, struct hammer_buffer **bufferp);
1068 void    *hammer_bread_ext(struct hammer_mount *hmp, hammer_off_t off, int bytes,
1069                         int *errorp, struct hammer_buffer **bufferp);
1070 void    *hammer_bnew_ext(struct hammer_mount *hmp, hammer_off_t off, int bytes,
1071                         int *errorp, struct hammer_buffer **bufferp);
1072
1073 hammer_volume_t hammer_get_root_volume(hammer_mount_t hmp, int *errorp);
1074
1075 hammer_volume_t hammer_get_volume(hammer_mount_t hmp,
1076                         int32_t vol_no, int *errorp);
1077 hammer_buffer_t hammer_get_buffer(hammer_mount_t hmp, hammer_off_t buf_offset,
1078                         int bytes, int isnew, int *errorp);
1079 void            hammer_sync_buffers(hammer_mount_t hmp,
1080                         hammer_off_t base_offset, int bytes);
1081 int             hammer_del_buffers(hammer_mount_t hmp,
1082                         hammer_off_t base_offset,
1083                         hammer_off_t zone2_offset, int bytes,
1084                         int report_conflicts);
1085
1086 int             hammer_ref_volume(hammer_volume_t volume);
1087 int             hammer_ref_buffer(hammer_buffer_t buffer);
1088 void            hammer_flush_buffer_nodes(hammer_buffer_t buffer);
1089
1090 void            hammer_rel_volume(hammer_volume_t volume, int flush);
1091 void            hammer_rel_buffer(hammer_buffer_t buffer, int flush);
1092
1093 int             hammer_vfs_export(struct mount *mp, int op,
1094                         const struct export_args *export);
1095 hammer_node_t   hammer_get_node(hammer_transaction_t trans,
1096                         hammer_off_t node_offset, int isnew, int *errorp);
1097 void            hammer_ref_node(hammer_node_t node);
1098 hammer_node_t   hammer_ref_node_safe(hammer_transaction_t trans,
1099                         hammer_node_cache_t cache, int *errorp);
1100 void            hammer_rel_node(hammer_node_t node);
1101 void            hammer_delete_node(hammer_transaction_t trans,
1102                         hammer_node_t node);
1103 void            hammer_cache_node(hammer_node_cache_t cache,
1104                         hammer_node_t node);
1105 void            hammer_uncache_node(hammer_node_cache_t cache);
1106 void            hammer_flush_node(hammer_node_t node);
1107
1108 void hammer_dup_buffer(struct hammer_buffer **bufferp,
1109                         struct hammer_buffer *buffer);
1110 hammer_node_t hammer_alloc_btree(hammer_transaction_t trans,
1111                         hammer_off_t hint, int *errorp);
1112 void *hammer_alloc_data(hammer_transaction_t trans, int32_t data_len,
1113                         u_int16_t rec_type, hammer_off_t *data_offsetp,
1114                         struct hammer_buffer **data_bufferp,
1115                         hammer_off_t hint, int *errorp);
1116
1117 int hammer_generate_undo(hammer_transaction_t trans,
1118                         hammer_off_t zone1_offset, void *base, int len);
1119 int hammer_generate_redo(hammer_transaction_t trans, hammer_inode_t ip,
1120                         hammer_off_t file_offset, void *base, int len);
1121 int hammer_upgrade_undo_4(hammer_transaction_t trans);
1122
1123 void hammer_put_volume(struct hammer_volume *volume, int flush);
1124 void hammer_put_buffer(struct hammer_buffer *buffer, int flush);
1125
1126 hammer_off_t hammer_freemap_alloc(hammer_transaction_t trans,
1127                         hammer_off_t owner, int *errorp);
1128 void hammer_freemap_free(hammer_transaction_t trans, hammer_off_t phys_offset,
1129                         hammer_off_t owner, int *errorp);
1130 int _hammer_checkspace(hammer_mount_t hmp, int slop, int64_t *resp);
1131 hammer_off_t hammer_blockmap_alloc(hammer_transaction_t trans, int zone,
1132                         int bytes, hammer_off_t hint, int *errorp);
1133 hammer_reserve_t hammer_blockmap_reserve(hammer_mount_t hmp, int zone,
1134                         int bytes, hammer_off_t *zone_offp, int *errorp);
1135 void hammer_blockmap_reserve_complete(hammer_mount_t hmp,
1136                         hammer_reserve_t resv);
1137 void hammer_reserve_clrdelay(hammer_mount_t hmp, hammer_reserve_t resv);
1138 void hammer_blockmap_free(hammer_transaction_t trans,
1139                         hammer_off_t bmap_off, int bytes);
1140 int hammer_blockmap_finalize(hammer_transaction_t trans,
1141                         hammer_reserve_t resv,
1142                         hammer_off_t bmap_off, int bytes);
1143 int hammer_blockmap_getfree(hammer_mount_t hmp, hammer_off_t bmap_off,
1144                         int *curp, int *errorp);
1145 hammer_off_t hammer_blockmap_lookup(hammer_mount_t hmp, hammer_off_t bmap_off,
1146                         int *errorp);
1147 hammer_off_t hammer_undo_lookup(hammer_mount_t hmp, hammer_off_t bmap_off,
1148                         int *errorp);
1149 int64_t hammer_undo_used(hammer_transaction_t trans);
1150 int64_t hammer_undo_space(hammer_transaction_t trans);
1151 int64_t hammer_undo_max(hammer_mount_t hmp);
1152 int hammer_undo_reclaim(hammer_io_t io);
1153
1154 void hammer_start_transaction(struct hammer_transaction *trans,
1155                               struct hammer_mount *hmp);
1156 void hammer_simple_transaction(struct hammer_transaction *trans,
1157                               struct hammer_mount *hmp);
1158 void hammer_start_transaction_fls(struct hammer_transaction *trans,
1159                                   struct hammer_mount *hmp);
1160 void hammer_done_transaction(struct hammer_transaction *trans);
1161 hammer_tid_t hammer_alloc_tid(hammer_mount_t hmp, int count);
1162
1163 void hammer_modify_inode(hammer_inode_t ip, int flags);
1164 void hammer_flush_inode(hammer_inode_t ip, int flags);
1165 void hammer_flush_inode_done(hammer_inode_t ip, int error);
1166 void hammer_wait_inode(hammer_inode_t ip);
1167
1168 int  hammer_create_inode(struct hammer_transaction *trans, struct vattr *vap,
1169                         struct ucred *cred, struct hammer_inode *dip,
1170                         const char *name, int namelen,
1171                         hammer_pseudofs_inmem_t pfsm,
1172                         struct hammer_inode **ipp);
1173 void hammer_rel_inode(hammer_inode_t ip, int flush);
1174 int hammer_reload_inode(hammer_inode_t ip, void *arg __unused);
1175 int hammer_ino_rb_compare(hammer_inode_t ip1, hammer_inode_t ip2);
1176 int hammer_destroy_inode_callback(hammer_inode_t ip, void *data __unused);
1177
1178 int hammer_sync_inode(hammer_transaction_t trans, hammer_inode_t ip);
1179 void hammer_test_inode(hammer_inode_t dip);
1180 void hammer_inode_unloadable_check(hammer_inode_t ip, int getvp);
1181
1182 int  hammer_ip_add_directory(struct hammer_transaction *trans,
1183                         hammer_inode_t dip, const char *name, int bytes,
1184                         hammer_inode_t nip);
1185 int  hammer_ip_del_directory(struct hammer_transaction *trans,
1186                         hammer_cursor_t cursor, hammer_inode_t dip,
1187                         hammer_inode_t ip);
1188 hammer_record_t hammer_ip_add_bulk(hammer_inode_t ip, off_t file_offset,
1189                         void *data, int bytes, int *errorp);
1190 int  hammer_ip_frontend_trunc(struct hammer_inode *ip, off_t file_size);
1191 int  hammer_ip_add_record(struct hammer_transaction *trans,
1192                         hammer_record_t record);
1193 int  hammer_ip_delete_range(hammer_cursor_t cursor, hammer_inode_t ip,
1194                         int64_t ran_beg, int64_t ran_end, int truncating);
1195 int  hammer_ip_delete_clean(hammer_cursor_t cursor, hammer_inode_t ip,
1196                         int *countp);
1197 int  hammer_ip_sync_data(hammer_cursor_t cursor, hammer_inode_t ip,
1198                         int64_t offset, void *data, int bytes);
1199 int  hammer_ip_sync_record(hammer_transaction_t trans, hammer_record_t rec);
1200 int  hammer_ip_sync_record_cursor(hammer_cursor_t cursor, hammer_record_t rec);
1201 hammer_pseudofs_inmem_t  hammer_load_pseudofs(hammer_transaction_t trans,
1202                         u_int32_t localization, int *errorp);
1203 int  hammer_mkroot_pseudofs(hammer_transaction_t trans, struct ucred *cred,
1204                         hammer_pseudofs_inmem_t pfsm);
1205 int  hammer_save_pseudofs(hammer_transaction_t trans,
1206                         hammer_pseudofs_inmem_t pfsm);
1207 int  hammer_unload_pseudofs(hammer_transaction_t trans, u_int32_t localization);
1208 void hammer_rel_pseudofs(hammer_mount_t hmp, hammer_pseudofs_inmem_t pfsm);
1209 int hammer_ioctl(hammer_inode_t ip, u_long com, caddr_t data, int fflag,
1210                         struct ucred *cred);
1211
1212 void hammer_io_init(hammer_io_t io, hammer_volume_t volume,
1213                         enum hammer_io_type type);
1214 int hammer_io_read(struct vnode *devvp, struct hammer_io *io,
1215                         hammer_off_t limit);
1216 void hammer_io_advance(struct hammer_io *io);
1217 int hammer_io_new(struct vnode *devvp, struct hammer_io *io);
1218 int hammer_io_inval(hammer_volume_t volume, hammer_off_t zone2_offset);
1219 struct buf *hammer_io_release(struct hammer_io *io, int flush);
1220 void hammer_io_flush(struct hammer_io *io, int reclaim);
1221 void hammer_io_wait(struct hammer_io *io);
1222 void hammer_io_waitdep(struct hammer_io *io);
1223 void hammer_io_wait_all(hammer_mount_t hmp, const char *ident);
1224 int hammer_io_direct_read(hammer_mount_t hmp, struct bio *bio,
1225                         hammer_btree_leaf_elm_t leaf);
1226 int hammer_io_direct_write(hammer_mount_t hmp, hammer_record_t record,
1227                         struct bio *bio);
1228 void hammer_io_direct_wait(hammer_record_t record);
1229 void hammer_io_direct_uncache(hammer_mount_t hmp, hammer_btree_leaf_elm_t leaf);
1230 void hammer_io_write_interlock(hammer_io_t io);
1231 void hammer_io_done_interlock(hammer_io_t io);
1232 void hammer_io_clear_modify(struct hammer_io *io, int inval);
1233 void hammer_io_clear_modlist(struct hammer_io *io);
1234 void hammer_io_flush_sync(hammer_mount_t hmp);
1235 void hammer_io_clear_error(struct hammer_io *io);
1236
1237 void hammer_modify_volume(hammer_transaction_t trans, hammer_volume_t volume,
1238                         void *base, int len);
1239 void hammer_modify_buffer(hammer_transaction_t trans, hammer_buffer_t buffer,
1240                         void *base, int len);
1241 void hammer_modify_volume_done(hammer_volume_t volume);
1242 void hammer_modify_buffer_done(hammer_buffer_t buffer);
1243
1244 int hammer_ioc_reblock(hammer_transaction_t trans, hammer_inode_t ip,
1245                         struct hammer_ioc_reblock *reblock);
1246 int hammer_ioc_rebalance(hammer_transaction_t trans, hammer_inode_t ip,
1247                         struct hammer_ioc_rebalance *rebal);
1248 int hammer_ioc_prune(hammer_transaction_t trans, hammer_inode_t ip,
1249                         struct hammer_ioc_prune *prune);
1250 int hammer_ioc_mirror_read(hammer_transaction_t trans, hammer_inode_t ip,
1251                         struct hammer_ioc_mirror_rw *mirror);
1252 int hammer_ioc_mirror_write(hammer_transaction_t trans, hammer_inode_t ip,
1253                         struct hammer_ioc_mirror_rw *mirror);
1254 int hammer_ioc_set_pseudofs(hammer_transaction_t trans, hammer_inode_t ip,
1255                         struct ucred *cred, struct hammer_ioc_pseudofs_rw *pfs);
1256 int hammer_ioc_get_pseudofs(hammer_transaction_t trans, hammer_inode_t ip,
1257                         struct hammer_ioc_pseudofs_rw *pfs);
1258 int hammer_ioc_destroy_pseudofs(hammer_transaction_t trans, hammer_inode_t ip,
1259                         struct hammer_ioc_pseudofs_rw *pfs);
1260 int hammer_ioc_downgrade_pseudofs(hammer_transaction_t trans, hammer_inode_t ip,
1261                         struct hammer_ioc_pseudofs_rw *pfs);
1262 int hammer_ioc_upgrade_pseudofs(hammer_transaction_t trans, hammer_inode_t ip,
1263                         struct hammer_ioc_pseudofs_rw *pfs);
1264 int hammer_ioc_wait_pseudofs(hammer_transaction_t trans, hammer_inode_t ip,
1265                         struct hammer_ioc_pseudofs_rw *pfs);
1266 int hammer_ioc_volume_add(hammer_transaction_t trans, hammer_inode_t ip,
1267                         struct hammer_ioc_volume *ioc);
1268 int hammer_ioc_volume_del(hammer_transaction_t trans, hammer_inode_t ip,
1269                         struct hammer_ioc_volume *ioc);
1270
1271 int hammer_signal_check(hammer_mount_t hmp);
1272
1273 void hammer_flusher_create(hammer_mount_t hmp);
1274 void hammer_flusher_destroy(hammer_mount_t hmp);
1275 void hammer_flusher_sync(hammer_mount_t hmp);
1276 int  hammer_flusher_async(hammer_mount_t hmp, hammer_flush_group_t flg);
1277 int  hammer_flusher_async_one(hammer_mount_t hmp);
1278 void hammer_flusher_wait(hammer_mount_t hmp, int seq);
1279 void hammer_flusher_wait_next(hammer_mount_t hmp);
1280 int  hammer_flusher_meta_limit(hammer_mount_t hmp);
1281 int  hammer_flusher_meta_halflimit(hammer_mount_t hmp);
1282 int  hammer_flusher_undo_exhausted(hammer_transaction_t trans, int quarter);
1283 void hammer_flusher_clean_loose_ios(hammer_mount_t hmp);
1284 void hammer_flusher_finalize(hammer_transaction_t trans, int final);
1285 int  hammer_flusher_haswork(hammer_mount_t hmp);
1286 void hammer_flusher_flush_undos(hammer_mount_t hmp, int already_flushed);
1287
1288 int hammer_recover_stage1(hammer_mount_t hmp, hammer_volume_t rootvol);
1289 int hammer_recover_stage2(hammer_mount_t hmp, hammer_volume_t rootvol);
1290 void hammer_recover_flush_buffers(hammer_mount_t hmp,
1291                         hammer_volume_t root_volume, int final);
1292
1293 void hammer_crc_set_blockmap(hammer_blockmap_t blockmap);
1294 void hammer_crc_set_volume(hammer_volume_ondisk_t ondisk);
1295 void hammer_crc_set_leaf(void *data, hammer_btree_leaf_elm_t leaf);
1296
1297 int hammer_crc_test_blockmap(hammer_blockmap_t blockmap);
1298 int hammer_crc_test_volume(hammer_volume_ondisk_t ondisk);
1299 int hammer_crc_test_btree(hammer_node_ondisk_t ondisk);
1300 int hammer_crc_test_leaf(void *data, hammer_btree_leaf_elm_t leaf);
1301 void hkprintf(const char *ctl, ...);
1302 udev_t hammer_fsid_to_udev(uuid_t *uuid);
1303
1304
1305 int hammer_blocksize(int64_t file_offset);
1306 int64_t hammer_blockdemarc(int64_t file_offset1, int64_t file_offset2);
1307
1308 /*
1309  * Shortcut for _hammer_checkspace(), used all over the code.
1310  */
1311 static __inline int
1312 hammer_checkspace(hammer_mount_t hmp, int slop)
1313 {
1314         return(_hammer_checkspace(hmp, slop, NULL));
1315 }
1316
1317 #endif
1318
1319 static __inline void
1320 hammer_wait_mem_record(hammer_record_t record)
1321 {
1322         hammer_wait_mem_record_ident(record, "hmmwai");
1323 }
1324
1325 static __inline void
1326 hammer_lock_ex(struct hammer_lock *lock)
1327 {
1328         hammer_lock_ex_ident(lock, "hmrlck");
1329 }
1330
1331 /*
1332  * Indicate that a B-Tree node is being modified.
1333  */
1334 static __inline void
1335 hammer_modify_node_noundo(hammer_transaction_t trans, hammer_node_t node)
1336 {
1337         KKASSERT((node->flags & HAMMER_NODE_CRCBAD) == 0);
1338         hammer_modify_buffer(trans, node->buffer, NULL, 0);
1339 }
1340
1341 static __inline void
1342 hammer_modify_node_all(hammer_transaction_t trans, struct hammer_node *node)
1343 {
1344         KKASSERT((node->flags & HAMMER_NODE_CRCBAD) == 0);
1345         hammer_modify_buffer(trans, node->buffer,
1346                              node->ondisk, sizeof(*node->ondisk));
1347 }
1348
1349 static __inline void
1350 hammer_modify_node(hammer_transaction_t trans, hammer_node_t node,
1351                    void *base, int len)
1352 {
1353         hammer_crc_t *crcptr;
1354
1355         KKASSERT((char *)base >= (char *)node->ondisk &&
1356                  (char *)base + len <=
1357                     (char *)node->ondisk + sizeof(*node->ondisk));
1358         KKASSERT((node->flags & HAMMER_NODE_CRCBAD) == 0);
1359         hammer_modify_buffer(trans, node->buffer, base, len);
1360         crcptr = &node->ondisk->crc;
1361         hammer_modify_buffer(trans, node->buffer, crcptr, sizeof(hammer_crc_t));
1362         --node->buffer->io.modify_refs; /* only want one ref */
1363 }
1364
1365 /*
1366  * Indicate that the specified modifications have been completed.
1367  *
1368  * Do not try to generate the crc here, it's very expensive to do and a
1369  * sequence of insertions or deletions can result in many calls to this
1370  * function on the same node.
1371  */
1372 static __inline void
1373 hammer_modify_node_done(hammer_node_t node)
1374 {
1375         node->flags |= HAMMER_NODE_CRCGOOD;
1376         if ((node->flags & HAMMER_NODE_NEEDSCRC) == 0) {
1377                 node->flags |= HAMMER_NODE_NEEDSCRC;
1378                 node->buffer->io.gencrc = 1;
1379                 hammer_ref_node(node);
1380         }
1381         hammer_modify_buffer_done(node->buffer);
1382 }
1383
1384 #define hammer_modify_volume_field(trans, vol, field)           \
1385         hammer_modify_volume(trans, vol, &(vol)->ondisk->field, \
1386                              sizeof((vol)->ondisk->field))
1387
1388 #define hammer_modify_node_field(trans, node, field)            \
1389         hammer_modify_node(trans, node, &(node)->ondisk->field, \
1390                              sizeof((node)->ondisk->field))
1391
1392 /*
1393  * The HAMMER_INODE_CAP_DIR_LOCAL_INO capability is set on newly
1394  * created directories for HAMMER version 2 or greater and causes
1395  * directory entries to be placed the inode localization zone in
1396  * the B-Tree instead of the misc zone.
1397  *
1398  * This greatly improves localization between directory entries and
1399  * inodes
1400  */
1401 static __inline u_int32_t
1402 hammer_dir_localization(hammer_inode_t dip)
1403 {
1404         if (dip->ino_data.cap_flags & HAMMER_INODE_CAP_DIR_LOCAL_INO)
1405                 return(HAMMER_LOCALIZE_INODE);
1406         else
1407                 return(HAMMER_LOCALIZE_MISC);
1408 }