sys/vfs/hammer/hammer.h

   1 /*
   2  * Copyright (c) 2007-2008 The DragonFly Project.  All rights reserved.
   3  *
   4  * This code is derived from software contributed to The DragonFly Project
   5  * by Matthew Dillon <dillon@backplane.com>
   6  *
   7  * Redistribution and use in source and binary forms, with or without
   8  * modification, are permitted provided that the following conditions
   9  * are met:
  10  *
  11  * 1. Redistributions of source code must retain the above copyright
  12  *    notice, this list of conditions and the following disclaimer.
  13  * 2. Redistributions in binary form must reproduce the above copyright
  14  *    notice, this list of conditions and the following disclaimer in
  15  *    the documentation and/or other materials provided with the
  16  *    distribution.
  17  * 3. Neither the name of The DragonFly Project nor the names of its
  18  *    contributors may be used to endorse or promote products derived
  19  *    from this software without specific, prior written permission.
  20  *
  21  * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
  22  * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
  23  * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS
  24  * FOR A PARTICULAR PURPOSE ARE DISCLAIMED.  IN NO EVENT SHALL THE
  25  * COPYRIGHT HOLDERS OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT,
  26  * INCIDENTAL, SPECIAL, EXEMPLARY OR CONSEQUENTIAL DAMAGES (INCLUDING,
  27  * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
  28  * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED
  29  * AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
  30  * OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT
  31  * OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
  32  * SUCH DAMAGE.
  33  *
  34  * $DragonFly: src/sys/vfs/hammer/hammer.h,v 1.130 2008/11/13 02:18:43 dillon Exp $
  35  */
  36 /*
  37  * This header file contains structures used internally by the HAMMERFS
  38  * implementation.  See hammer_disk.h for on-disk structures.
  39  */
  40
  41 #include <sys/param.h>
  42 #include <sys/types.h>
  43 #include <sys/kernel.h>
  44 #include <sys/conf.h>
  45 #include <sys/systm.h>
  46 #include <sys/tree.h>
  47 #include <sys/malloc.h>
  48 #include <sys/mount.h>
  49 #include <sys/mountctl.h>
  50 #include <sys/vnode.h>
  51 #include <sys/proc.h>
  52 #include <sys/priv.h>
  53 #include <sys/stat.h>
  54 #include <sys/globaldata.h>
  55 #include <sys/lockf.h>
  56 #include <sys/buf.h>
  57 #include <sys/queue.h>
  58 #include <sys/ktr.h>
  59 #include <sys/globaldata.h>
  60
  61 #include <sys/buf2.h>
  62 #include <sys/signal2.h>
  63 #include "hammer_disk.h"
  64 #include "hammer_mount.h"
  65 #include "hammer_ioctl.h"
  66
  67 #if defined(_KERNEL) || defined(_KERNEL_STRUCTURES)
  68
  69 MALLOC_DECLARE(M_HAMMER);
  70
  71 /*
  72  * Kernel trace
  73  */
  74 #if !defined(KTR_HAMMER)
  75 #define KTR_HAMMER      KTR_ALL
  76 #endif
  77 KTR_INFO_MASTER_EXTERN(hammer);
  78
  79 /*
  80  * Misc structures
  81  */
  82 struct hammer_mount;
  83
  84 /*
  85  * Key structure used for custom RB tree inode lookups.  This prototypes
  86  * the function hammer_ino_rb_tree_RB_LOOKUP_INFO(root, info).
  87  */
  88 typedef struct hammer_inode_info {
  89         int64_t         obj_id;         /* (key) object identifier */
  90         hammer_tid_t    obj_asof;       /* (key) snapshot transid or 0 */
  91         u_int32_t       obj_localization; /* (key) pseudo-fs */
  92         union {
  93                 struct hammer_btree_leaf_elm *leaf;
  94         } u;
  95 } *hammer_inode_info_t;
  96
  97 typedef enum hammer_transaction_type {
  98         HAMMER_TRANS_RO,
  99         HAMMER_TRANS_STD,
 100         HAMMER_TRANS_FLS
 101 } hammer_transaction_type_t;
 102
 103 /*
 104  * HAMMER Transaction tracking
 105  */
 106 struct hammer_transaction {
 107         hammer_transaction_type_t type;
 108         struct hammer_mount *hmp;
 109         hammer_tid_t    tid;
 110         u_int64_t       time;
 111         u_int32_t       time32;
 112         int             sync_lock_refs;
 113         int             flags;
 114         struct hammer_volume *rootvol;
 115 };
 116
 117 typedef struct hammer_transaction *hammer_transaction_t;
 118
 119 #define HAMMER_TRANSF_NEWINODE  0x0001
 120 #define HAMMER_TRANSF_DIDIO     0x0002
 121 #define HAMMER_TRANSF_CRCDOM    0x0004  /* EDOM on CRC error, less critical */
 122
 123 /*
 124  * HAMMER locks
 125  */
 126 struct hammer_lock {
 127         int             refs;           /* active references delay writes */
 128         volatile u_int  lockval;        /* lock count and control bits */
 129         struct thread   *owner;         /* owner if exclusively held */
 130 };
 131
 132 #define HAMMER_LOCKF_EXCLUSIVE  0x40000000
 133 #define HAMMER_LOCKF_WANTED     0x80000000
 134
 135 static __inline int
 136 hammer_notlocked(struct hammer_lock *lock)
 137 {
 138         return(lock->lockval == 0);
 139 }
 140
 141 static __inline int
 142 hammer_islocked(struct hammer_lock *lock)
 143 {
 144         return(lock->lockval != 0);
 145 }
 146
 147 static __inline int
 148 hammer_isactive(struct hammer_lock *lock)
 149 {
 150         return(lock->refs != 0);
 151 }
 152
 153 static __inline int
 154 hammer_islastref(struct hammer_lock *lock)
 155 {
 156         return(lock->refs == 1);
 157 }
 158
 159 /*
 160  * Return if we specifically own the lock exclusively.
 161  */
 162 static __inline int
 163 hammer_lock_excl_owned(struct hammer_lock *lock, thread_t td)
 164 {
 165         if ((lock->lockval & HAMMER_LOCKF_EXCLUSIVE) &&
 166             lock->owner == td) {
 167                 return(1);
 168         }
 169         return(0);
 170 }
 171
 172 /*
 173  * Flush state, used by various structures
 174  */
 175 typedef enum hammer_inode_state {
 176         HAMMER_FST_IDLE,
 177         HAMMER_FST_SETUP,
 178         HAMMER_FST_FLUSH
 179 } hammer_inode_state_t;
 180
 181 TAILQ_HEAD(hammer_record_list, hammer_record);
 182
 183 /*
 184  * Pseudo-filesystem extended data tracking
 185  */
 186 struct hammer_pfs_rb_tree;
 187 struct hammer_pseudofs_inmem;
 188 RB_HEAD(hammer_pfs_rb_tree, hammer_pseudofs_inmem);
 189 RB_PROTOTYPE2(hammer_pfs_rb_tree, hammer_pseudofs_inmem, rb_node,
 190               hammer_pfs_rb_compare, u_int32_t);
 191
 192 struct hammer_pseudofs_inmem {
 193         RB_ENTRY(hammer_pseudofs_inmem) rb_node;
 194         struct hammer_lock      lock;
 195         u_int32_t               localization;
 196         hammer_tid_t            create_tid;
 197         int                     flags;
 198         udev_t                  fsid_udev;
 199         struct hammer_pseudofs_data pfsd;
 200 };
 201
 202 typedef struct hammer_pseudofs_inmem *hammer_pseudofs_inmem_t;
 203
 204 #define HAMMER_PFSM_DELETED     0x0001
 205
 206 /*
 207  * Cache object ids.  A fixed number of objid cache structures are
 208  * created to reserve object id's for newly created files in multiples
 209  * of 100,000, localized to a particular directory, and recycled as
 210  * needed.  This allows parallel create operations in different
 211  * directories to retain fairly localized object ids which in turn
 212  * improves reblocking performance and layout.
 213  */
 214 #define OBJID_CACHE_SIZE        1024
 215 #define OBJID_CACHE_BULK_BITS   10              /* 10 bits (1024)       */
 216 #define OBJID_CACHE_BULK        (32 * 32)       /* two level (1024)     */
 217 #define OBJID_CACHE_BULK_MASK   (OBJID_CACHE_BULK - 1)
 218 #define OBJID_CACHE_BULK_MASK64 ((u_int64_t)(OBJID_CACHE_BULK - 1))
 219
 220 typedef struct hammer_objid_cache {
 221         TAILQ_ENTRY(hammer_objid_cache) entry;
 222         struct hammer_inode             *dip;
 223         hammer_tid_t                    base_tid;
 224         int                             count;
 225         u_int32_t                       bm0;
 226         u_int32_t                       bm1[32];
 227 } *hammer_objid_cache_t;
 228
 229 /*
 230  * Associate an inode with a B-Tree node to cache search start positions
 231  */
 232 typedef struct hammer_node_cache {
 233         TAILQ_ENTRY(hammer_node_cache)  entry;
 234         struct hammer_node              *node;
 235         struct hammer_inode             *ip;
 236 } *hammer_node_cache_t;
 237
 238 TAILQ_HEAD(hammer_node_cache_list, hammer_node_cache);
 239
 240 /*
 241  * Structure used to organize flush groups.  Flush groups must be
 242  * organized into chunks in order to avoid blowing out the UNDO FIFO.
 243  * Without this a 'sync' could end up flushing 50,000 inodes in a single
 244  * transaction.
 245  */
 246 struct hammer_fls_rb_tree;
 247 RB_HEAD(hammer_fls_rb_tree, hammer_inode);
 248 RB_PROTOTYPE(hammer_fls_rb_tree, hammer_inode, rb_flsnode,
 249               hammer_ino_rb_compare);
 250
 251 struct hammer_flush_group {
 252         TAILQ_ENTRY(hammer_flush_group) flush_entry;
 253         struct hammer_fls_rb_tree       flush_tree;
 254         int                             unused01;
 255         int                             total_count;    /* record load */
 256         int                             running;        /* group is running */
 257         int                             closed;
 258         int                             refs;
 259 };
 260
 261 typedef struct hammer_flush_group *hammer_flush_group_t;
 262
 263 TAILQ_HEAD(hammer_flush_group_list, hammer_flush_group);
 264
 265 /*
 266  * Structure used to represent an inode in-memory.
 267  *
 268  * The record and data associated with an inode may be out of sync with
 269  * the disk (xDIRTY flags), or not even on the disk at all (ONDISK flag
 270  * clear).
 271  *
 272  * An inode may also hold a cache of unsynchronized records, used for
 273  * database and directories only.  Unsynchronized regular file data is
 274  * stored in the buffer cache.
 275  *
 276  * NOTE: A file which is created and destroyed within the initial
 277  * synchronization period can wind up not doing any disk I/O at all.
 278  *
 279  * Finally, an inode may cache numerous disk-referencing B-Tree cursors.
 280  */
 281 struct hammer_ino_rb_tree;
 282 struct hammer_inode;
 283 RB_HEAD(hammer_ino_rb_tree, hammer_inode);
 284 RB_PROTOTYPEX(hammer_ino_rb_tree, INFO, hammer_inode, rb_node,
 285               hammer_ino_rb_compare, hammer_inode_info_t);
 286
 287 struct hammer_rec_rb_tree;
 288 struct hammer_record;
 289 RB_HEAD(hammer_rec_rb_tree, hammer_record);
 290 RB_PROTOTYPEX(hammer_rec_rb_tree, INFO, hammer_record, rb_node,
 291               hammer_rec_rb_compare, hammer_btree_leaf_elm_t);
 292
 293 TAILQ_HEAD(hammer_node_list, hammer_node);
 294
 295 struct hammer_inode {
 296         RB_ENTRY(hammer_inode)  rb_node;
 297         hammer_inode_state_t    flush_state;
 298         hammer_flush_group_t    flush_group;
 299         RB_ENTRY(hammer_inode)  rb_flsnode;     /* when on flush list */
 300         struct hammer_record_list target_list;  /* target of dependant recs */
 301         int64_t                 obj_id;         /* (key) object identifier */
 302         hammer_tid_t            obj_asof;       /* (key) snapshot or 0 */
 303         u_int32_t               obj_localization; /* (key) pseudo-fs */
 304         struct hammer_mount     *hmp;
 305         hammer_objid_cache_t    objid_cache;
 306         int                     flags;
 307         int                     error;          /* flush error */
 308         int                     cursor_ip_refs; /* sanity */
 309         int                     rsv_recs;
 310         struct vnode            *vp;
 311         hammer_pseudofs_inmem_t pfsm;
 312         struct lockf            advlock;
 313         struct hammer_lock      lock;           /* sync copy interlock */
 314         off_t                   trunc_off;
 315         struct hammer_btree_leaf_elm ino_leaf;  /* in-memory cache */
 316         struct hammer_inode_data ino_data;      /* in-memory cache */
 317         struct hammer_rec_rb_tree rec_tree;     /* in-memory cache */
 318         int                     rec_generation;
 319         struct hammer_node_cache cache[4];      /* search initiate cache */
 320
 321         /*
 322          * When a demark is created to synchronize an inode to
 323          * disk, certain fields are copied so the front-end VOPs
 324          * can continue to run in parallel with the synchronization
 325          * occuring in the background.
 326          */
 327         int             sync_flags;             /* to-sync flags cache */
 328         off_t           sync_trunc_off;         /* to-sync truncation */
 329         off_t           save_trunc_off;         /* write optimization */
 330         struct hammer_btree_leaf_elm sync_ino_leaf; /* to-sync cache */
 331         struct hammer_inode_data sync_ino_data; /* to-sync cache */
 332 };
 333
 334 typedef struct hammer_inode *hammer_inode_t;
 335
 336 #define VTOI(vp)        ((struct hammer_inode *)(vp)->v_data)
 337
 338 #define HAMMER_INODE_DDIRTY     0x0001  /* in-memory ino_data is dirty */
 339                                         /* (not including atime/mtime) */
 340 #define HAMMER_INODE_RSV_INODES 0x0002  /* hmp->rsv_inodes bumped */
 341 #define HAMMER_INODE_CONN_DOWN  0x0004  /* include in downward recursion */
 342 #define HAMMER_INODE_XDIRTY     0x0008  /* in-memory records */
 343 #define HAMMER_INODE_ONDISK     0x0010  /* inode is on-disk (else not yet) */
 344 #define HAMMER_INODE_FLUSH      0x0020  /* flush on last ref */
 345 #define HAMMER_INODE_DELETED    0x0080  /* inode delete (backend) */
 346 #define HAMMER_INODE_DELONDISK  0x0100  /* delete synchronized to disk */
 347 #define HAMMER_INODE_RO         0x0200  /* read-only (because of as-of) */
 348 #define HAMMER_INODE_VHELD      0x0400  /* vnode held on sync */
 349 #define HAMMER_INODE_DONDISK    0x0800  /* data records may be on disk */
 350 #define HAMMER_INODE_BUFS       0x1000  /* dirty high level bps present */
 351 #define HAMMER_INODE_REFLUSH    0x2000  /* flush on dependancy / reflush */
 352 #define HAMMER_INODE_RECLAIM    0x4000  /* trying to reclaim */
 353 #define HAMMER_INODE_FLUSHW     0x8000  /* Someone waiting for flush */
 354
 355 #define HAMMER_INODE_TRUNCATED  0x00010000
 356 #define HAMMER_INODE_DELETING   0x00020000 /* inode delete request (frontend)*/
 357 #define HAMMER_INODE_RESIGNAL   0x00040000 /* re-signal on re-flush */
 358 #define HAMMER_INODE_ATIME      0x00100000 /* in-memory atime modified */
 359 #define HAMMER_INODE_MTIME      0x00200000 /* in-memory mtime modified */
 360 #define HAMMER_INODE_WOULDBLOCK 0x00400000 /* re-issue to new flush group */
 361 #define HAMMER_INODE_DUMMY      0x00800000 /* dummy inode covering bad file */
 362 #define HAMMER_INODE_CLOSESYNC  0x01000000 /* synchronously fsync on close */
 363 #define HAMMER_INODE_CLOSEASYNC 0x02000000 /* asynchronously fsync on close */
 364
 365 #define HAMMER_INODE_MODMASK    (HAMMER_INODE_DDIRTY|                       \
 366                                  HAMMER_INODE_XDIRTY|HAMMER_INODE_BUFS|     \
 367                                  HAMMER_INODE_ATIME|HAMMER_INODE_MTIME|     \
 368                                  HAMMER_INODE_TRUNCATED|HAMMER_INODE_DELETING)
 369
 370 #define HAMMER_INODE_MODMASK_NOXDIRTY \
 371                                 (HAMMER_INODE_MODMASK & ~HAMMER_INODE_XDIRTY)
 372
 373 #define HAMMER_FLUSH_SIGNAL     0x0001
 374 #define HAMMER_FLUSH_RECURSION  0x0002
 375
 376 /*
 377  * Used by the inode reclaim code to pipeline reclaims and avoid
 378  * blowing out kernel memory or letting the flusher get too far
 379  * behind.  The reclaim wakes up when count reaches 0 or the
 380  * timer expires.
 381  */
 382 struct hammer_reclaim {
 383         TAILQ_ENTRY(hammer_reclaim) entry;
 384         int     count;
 385 };
 386
 387 #define HAMMER_RECLAIM_WAIT     4000    /* default vfs.hammer.limit_reclaim */
 388
 389 /*
 390  * Structure used to represent an unsynchronized record in-memory.  These
 391  * records typically represent directory entries.  Only non-historical
 392  * records are kept in-memory.
 393  *
 394  * Records are organized as a per-inode RB-Tree.  If the inode is not
 395  * on disk then neither are any records and the in-memory record tree
 396  * represents the entire contents of the inode.  If the inode is on disk
 397  * then the on-disk B-Tree is scanned in parallel with the in-memory
 398  * RB-Tree to synthesize the current state of the file.
 399  *
 400  * Records are also used to enforce the ordering of directory create/delete
 401  * operations.  A new inode will not be flushed to disk unless its related
 402  * directory entry is also being flushed at the same time.  A directory entry
 403  * will not be removed unless its related inode is also being removed at the
 404  * same time.
 405  */
 406 typedef enum hammer_record_type {
 407         HAMMER_MEM_RECORD_GENERAL,      /* misc record */
 408         HAMMER_MEM_RECORD_INODE,        /* inode record */
 409         HAMMER_MEM_RECORD_ADD,          /* positive memory cache record */
 410         HAMMER_MEM_RECORD_DEL,          /* negative delete-on-disk record */
 411         HAMMER_MEM_RECORD_DATA          /* bulk-data record w/on-disk ref */
 412 } hammer_record_type_t;
 413
 414 struct hammer_record {
 415         RB_ENTRY(hammer_record)         rb_node;
 416         TAILQ_ENTRY(hammer_record)      target_entry;
 417         hammer_inode_state_t            flush_state;
 418         hammer_flush_group_t            flush_group;
 419         hammer_record_type_t            type;
 420         struct hammer_lock              lock;
 421         struct hammer_reserve           *resv;
 422         struct hammer_inode             *ip;
 423         struct hammer_inode             *target_ip;
 424         struct hammer_btree_leaf_elm    leaf;
 425         union hammer_data_ondisk        *data;
 426         int                             flags;
 427         hammer_off_t                    zone2_offset;   /* direct-write only */
 428 };
 429
 430 typedef struct hammer_record *hammer_record_t;
 431
 432 /*
 433  * Record flags.  Note that FE can only be set by the frontend if the
 434  * record has not been interlocked by the backend w/ BE.
 435  */
 436 #define HAMMER_RECF_ALLOCDATA           0x0001
 437 #define HAMMER_RECF_ONRBTREE            0x0002
 438 #define HAMMER_RECF_DELETED_FE          0x0004  /* deleted (frontend) */
 439 #define HAMMER_RECF_DELETED_BE          0x0008  /* deleted (backend) */
 440 #define HAMMER_RECF_COMMITTED           0x0010  /* committed to the B-Tree */
 441 #define HAMMER_RECF_INTERLOCK_BE        0x0020  /* backend interlock */
 442 #define HAMMER_RECF_WANTED              0x0040  /* wanted by the frontend */
 443 #define HAMMER_RECF_CONVERT_DELETE      0x0100  /* special case */
 444 #define HAMMER_RECF_DIRECT_IO           0x0200  /* related direct I/O running*/
 445 #define HAMMER_RECF_DIRECT_WAIT         0x0400  /* related direct I/O running*/
 446 #define HAMMER_RECF_DIRECT_INVAL        0x0800  /* buffer alias invalidation */
 447
 448 /*
 449  * hammer_create_at_cursor() and hammer_delete_at_cursor() flags.
 450  */
 451 #define HAMMER_CREATE_MODE_UMIRROR      0x0001
 452 #define HAMMER_CREATE_MODE_SYS          0x0002
 453
 454 #define HAMMER_DELETE_ADJUST            0x0001
 455 #define HAMMER_DELETE_DESTROY           0x0002
 456
 457 /*
 458  * In-memory structures representing on-disk structures.
 459  */
 460 struct hammer_volume;
 461 struct hammer_buffer;
 462 struct hammer_node;
 463 struct hammer_undo;
 464 struct hammer_reserve;
 465
 466 RB_HEAD(hammer_vol_rb_tree, hammer_volume);
 467 RB_HEAD(hammer_buf_rb_tree, hammer_buffer);
 468 RB_HEAD(hammer_nod_rb_tree, hammer_node);
 469 RB_HEAD(hammer_und_rb_tree, hammer_undo);
 470 RB_HEAD(hammer_res_rb_tree, hammer_reserve);
 471
 472 RB_PROTOTYPE2(hammer_vol_rb_tree, hammer_volume, rb_node,
 473               hammer_vol_rb_compare, int32_t);
 474 RB_PROTOTYPE2(hammer_buf_rb_tree, hammer_buffer, rb_node,
 475               hammer_buf_rb_compare, hammer_off_t);
 476 RB_PROTOTYPE2(hammer_nod_rb_tree, hammer_node, rb_node,
 477               hammer_nod_rb_compare, hammer_off_t);
 478 RB_PROTOTYPE2(hammer_und_rb_tree, hammer_undo, rb_node,
 479               hammer_und_rb_compare, hammer_off_t);
 480 RB_PROTOTYPE2(hammer_res_rb_tree, hammer_reserve, rb_node,
 481               hammer_res_rb_compare, hammer_off_t);
 482
 483 /*
 484  * IO management - embedded at the head of various in-memory structures
 485  *
 486  * VOLUME       - hammer_volume containing meta-data
 487  * META_BUFFER  - hammer_buffer containing meta-data
 488  * DATA_BUFFER  - hammer_buffer containing pure-data
 489  *
 490  * Dirty volume headers and dirty meta-data buffers are locked until the
 491  * flusher can sequence them out.  Dirty pure-data buffers can be written.
 492  * Clean buffers can be passively released.
 493  */
 494 typedef enum hammer_io_type {
 495         HAMMER_STRUCTURE_VOLUME,
 496         HAMMER_STRUCTURE_META_BUFFER,
 497         HAMMER_STRUCTURE_UNDO_BUFFER,
 498         HAMMER_STRUCTURE_DATA_BUFFER
 499 } hammer_io_type_t;
 500
 501 union hammer_io_structure;
 502 struct hammer_io;
 503
 504 struct worklist {
 505         LIST_ENTRY(worklist) node;
 506 };
 507
 508 TAILQ_HEAD(hammer_io_list, hammer_io);
 509 typedef struct hammer_io_list *hammer_io_list_t;
 510
 511 struct hammer_io {
 512         struct worklist         worklist;
 513         struct hammer_lock      lock;
 514         enum hammer_io_type     type;
 515         struct hammer_mount     *hmp;
 516         struct hammer_volume    *volume;
 517         TAILQ_ENTRY(hammer_io)  mod_entry; /* list entry if modified */
 518         hammer_io_list_t        mod_list;
 519         struct buf              *bp;
 520         int64_t                 offset;    /* zone-2 offset */
 521         int                     bytes;     /* buffer cache buffer size */
 522         int                     loading;   /* loading/unloading interlock */
 523         int                     modify_refs;
 524
 525         u_int           modified : 1;   /* bp's data was modified */
 526         u_int           released : 1;   /* bp released (w/ B_LOCKED set) */
 527         u_int           running : 1;    /* bp write IO in progress */
 528         u_int           waiting : 1;    /* someone is waiting on us */
 529         u_int           validated : 1;  /* ondisk has been validated */
 530         u_int           waitdep : 1;    /* flush waits for dependancies */
 531         u_int           recovered : 1;  /* has recovery ref */
 532         u_int           waitmod : 1;    /* waiting for modify_refs */
 533         u_int           reclaim : 1;    /* reclaim requested */
 534         u_int           gencrc : 1;     /* crc needs to be generated */
 535         u_int           ioerror : 1;    /* abort on io-error */
 536 };
 537
 538 typedef struct hammer_io *hammer_io_t;
 539
 540 #define HAMMER_CLUSTER_SIZE     (64 * 1024)
 541 #if HAMMER_CLUSTER_SIZE > MAXBSIZE
 542 #undef  HAMMER_CLUSTER_SIZE
 543 #define HAMMER_CLUSTER_SIZE     MAXBSIZE
 544 #endif
 545 #define HAMMER_CLUSTER_BUFS     (HAMMER_CLUSTER_SIZE / HAMMER_BUFSIZE)
 546
 547 /*
 548  * In-memory volume representing on-disk buffer
 549  */
 550 struct hammer_volume {
 551         struct hammer_io io;
 552         RB_ENTRY(hammer_volume) rb_node;
 553         struct hammer_volume_ondisk *ondisk;
 554         int32_t vol_no;
 555         int64_t nblocks;        /* note: special calculation for statfs */
 556         int64_t buffer_base;    /* base offset of buffer 0 */
 557         hammer_off_t maxbuf_off; /* Maximum buffer offset (zone-2) */
 558         hammer_off_t maxraw_off; /* Maximum raw offset for device */
 559         char    *vol_name;
 560         struct vnode *devvp;
 561         int     vol_flags;
 562 };
 563
 564 typedef struct hammer_volume *hammer_volume_t;
 565
 566 /*
 567  * In-memory buffer (other then volume, super-cluster, or cluster),
 568  * representing an on-disk buffer.
 569  */
 570 struct hammer_buffer {
 571         struct hammer_io io;
 572         RB_ENTRY(hammer_buffer) rb_node;
 573         void *ondisk;
 574         hammer_off_t zoneX_offset;
 575         hammer_off_t zone2_offset;
 576         struct hammer_reserve *resv;
 577         struct hammer_node_list clist;
 578 };
 579
 580 typedef struct hammer_buffer *hammer_buffer_t;
 581
 582 /*
 583  * In-memory B-Tree node, representing an on-disk B-Tree node.
 584  *
 585  * This is a hang-on structure which is backed by a hammer_buffer,
 586  * indexed by a hammer_cluster, and used for fine-grained locking of
 587  * B-Tree nodes in order to properly control lock ordering.  A hammer_buffer
 588  * can contain multiple nodes representing wildly disassociated portions
 589  * of the B-Tree so locking cannot be done on a buffer-by-buffer basis.
 590  *
 591  * This structure uses a cluster-relative index to reduce the number
 592  * of layers required to access it, and also because all on-disk B-Tree
 593  * references are cluster-relative offsets.
 594  */
 595 struct hammer_node {
 596         struct hammer_lock      lock;           /* node-by-node lock */
 597         TAILQ_ENTRY(hammer_node) entry;         /* per-buffer linkage */
 598         RB_ENTRY(hammer_node)   rb_node;        /* per-cluster linkage */
 599         hammer_off_t            node_offset;    /* full offset spec */
 600         struct hammer_mount     *hmp;
 601         struct hammer_buffer    *buffer;        /* backing buffer */
 602         hammer_node_ondisk_t    ondisk;         /* ptr to on-disk structure */
 603         TAILQ_HEAD(, hammer_cursor) cursor_list;  /* deadlock recovery */
 604         struct hammer_node_cache_list cache_list; /* passive caches */
 605         int                     flags;
 606         int                     loading;        /* load interlock */
 607 };
 608
 609 #define HAMMER_NODE_DELETED     0x0001
 610 #define HAMMER_NODE_FLUSH       0x0002
 611 #define HAMMER_NODE_CRCGOOD     0x0004
 612 #define HAMMER_NODE_NEEDSCRC    0x0008
 613 #define HAMMER_NODE_NEEDSMIRROR 0x0010
 614 #define HAMMER_NODE_CRCBAD      0x0020
 615
 616 #define HAMMER_NODE_CRCANY      (HAMMER_NODE_CRCGOOD | HAMMER_NODE_CRCBAD)
 617
 618 typedef struct hammer_node      *hammer_node_t;
 619
 620 /*
 621  * List of locked nodes.  This structure is used to lock potentially large
 622  * numbers of nodes as an aid for complex B-Tree operations.
 623  */
 624 struct hammer_node_lock;
 625 TAILQ_HEAD(hammer_node_lock_list, hammer_node_lock);
 626
 627 struct hammer_node_lock {
 628         TAILQ_ENTRY(hammer_node_lock) entry;
 629         struct hammer_node_lock_list  list;
 630         struct hammer_node_lock       *parent;
 631         hammer_node_t   node;
 632         hammer_node_ondisk_t copy;      /* copy of on-disk data */
 633         int             index;          /* index of this node in parent */
 634         int             count;          /* count children */
 635         int             flags;
 636 };
 637
 638 typedef struct hammer_node_lock *hammer_node_lock_t;
 639
 640 #define HAMMER_NODE_LOCK_UPDATED        0x0001
 641
 642 /*
 643  * Common I/O management structure - embedded in in-memory structures
 644  * which are backed by filesystem buffers.
 645  */
 646 union hammer_io_structure {
 647         struct hammer_io        io;
 648         struct hammer_volume    volume;
 649         struct hammer_buffer    buffer;
 650 };
 651
 652 typedef union hammer_io_structure *hammer_io_structure_t;
 653
 654 /*
 655  * The reserve structure prevents the blockmap from allocating
 656  * out of a reserved bigblock.  Such reservations are used by
 657  * the direct-write mechanism.
 658  *
 659  * The structure is also used to hold off on reallocations of
 660  * big blocks from the freemap until flush dependancies have
 661  * been dealt with.
 662  */
 663 struct hammer_reserve {
 664         RB_ENTRY(hammer_reserve) rb_node;
 665         TAILQ_ENTRY(hammer_reserve) delay_entry;
 666         int             flush_group;
 667         int             flags;
 668         int             refs;
 669         int             zone;
 670         int             append_off;
 671         hammer_off_t    zone_offset;
 672 };
 673
 674 typedef struct hammer_reserve *hammer_reserve_t;
 675
 676 #define HAMMER_RESF_ONDELAY     0x0001
 677 #define HAMMER_RESF_LAYER2FREE  0x0002
 678
 679 #include "hammer_cursor.h"
 680
 681 /*
 682  * The undo structure tracks recent undos to avoid laying down duplicate
 683  * undos within a flush group, saving us a significant amount of overhead.
 684  *
 685  * This is strictly a heuristic.
 686  */
 687 #define HAMMER_MAX_UNDOS                1024
 688 #define HAMMER_MAX_FLUSHERS             4
 689
 690 struct hammer_undo {
 691         RB_ENTRY(hammer_undo)   rb_node;
 692         TAILQ_ENTRY(hammer_undo) lru_entry;
 693         hammer_off_t            offset;
 694         int                     bytes;
 695 };
 696
 697 typedef struct hammer_undo *hammer_undo_t;
 698
 699 struct hammer_flusher_info;
 700 TAILQ_HEAD(hammer_flusher_info_list, hammer_flusher_info);
 701
 702 struct hammer_flusher {
 703         int             signal;         /* flusher thread sequencer */
 704         int             act;            /* currently active flush group */
 705         int             done;           /* set to act when complete */
 706         int             next;           /* next flush group */
 707         int             group_lock;     /* lock sequencing of the next flush */
 708         int             exiting;        /* request master exit */
 709         thread_t        td;             /* master flusher thread */
 710         hammer_tid_t    tid;            /* last flushed transaction id */
 711         int             finalize_want;          /* serialize finalization */
 712         struct hammer_lock finalize_lock;       /* serialize finalization */
 713         struct hammer_transaction trans;        /* shared transaction */
 714         struct hammer_flusher_info_list run_list;
 715         struct hammer_flusher_info_list ready_list;
 716 };
 717
 718 /*
 719  * Internal hammer mount data structure
 720  */
 721 struct hammer_mount {
 722         struct mount *mp;
 723         /*struct vnode *rootvp;*/
 724         struct hammer_ino_rb_tree rb_inos_root;
 725         struct hammer_vol_rb_tree rb_vols_root;
 726         struct hammer_nod_rb_tree rb_nods_root;
 727         struct hammer_und_rb_tree rb_undo_root;
 728         struct hammer_res_rb_tree rb_resv_root;
 729         struct hammer_buf_rb_tree rb_bufs_root;
 730         struct hammer_pfs_rb_tree rb_pfsm_root;
 731         struct hammer_volume *rootvol;
 732         struct hammer_base_elm root_btree_beg;
 733         struct hammer_base_elm root_btree_end;
 734
 735         struct malloc_type      *m_misc;
 736         struct malloc_type      *m_inodes;
 737
 738         int     flags;          /* HAMMER_MOUNT_xxx flags */
 739         int     hflags;
 740         int     ronly;
 741         int     nvolumes;
 742         int     volume_iterator;
 743         int     master_id;      /* -1 or 0-15 - clustering and mirroring */
 744         int     version;        /* hammer filesystem version to use */
 745         int     rsv_inodes;     /* reserved space due to dirty inodes */
 746         int64_t rsv_databytes;  /* reserved space due to record data */
 747         int     rsv_recs;       /* reserved space due to dirty records */
 748         int     rsv_fromdelay;  /* bigblocks reserved due to flush delay */
 749         int     undo_rec_limit; /* based on size of undo area */
 750         int     last_newrecords;
 751         int     count_newrecords;
 752
 753         int     volume_to_remove; /* volume that is currently being removed */
 754
 755         int     inode_reclaims; /* inodes pending reclaim by flusher */
 756         int     count_inodes;   /* total number of inodes */
 757         int     count_iqueued;  /* inodes queued to flusher */
 758
 759         struct hammer_flusher flusher;
 760
 761         u_int   check_interrupt;
 762         u_int   check_yield;
 763         uuid_t  fsid;
 764         struct hammer_io_list volu_list;        /* dirty undo buffers */
 765         struct hammer_io_list undo_list;        /* dirty undo buffers */
 766         struct hammer_io_list data_list;        /* dirty data buffers */
 767         struct hammer_io_list alt_data_list;    /* dirty data buffers */
 768         struct hammer_io_list meta_list;        /* dirty meta bufs    */
 769         struct hammer_io_list lose_list;        /* loose buffers      */
 770         int     locked_dirty_space;             /* meta/volu count    */
 771         int     io_running_space;
 772         int     objid_cache_count;
 773         int     error;                          /* critical I/O error */
 774         struct krate    krate;                  /* rate limited kprintf */
 775         hammer_tid_t    asof;                   /* snapshot mount */
 776         hammer_tid_t    next_tid;
 777         hammer_tid_t    flush_tid1;             /* flusher tid sequencing */
 778         hammer_tid_t    flush_tid2;             /* flusher tid sequencing */
 779         int64_t copy_stat_freebigblocks;        /* number of free bigblocks */
 780         u_int32_t       undo_seqno;             /* UNDO/REDO FIFO seqno */
 781
 782         struct netexport export;
 783         struct hammer_lock sync_lock;
 784         struct hammer_lock free_lock;
 785         struct hammer_lock undo_lock;
 786         struct hammer_lock blkmap_lock;
 787         struct hammer_lock snapshot_lock;
 788         struct hammer_blockmap  blockmap[HAMMER_MAX_ZONES];
 789         struct hammer_undo      undos[HAMMER_MAX_UNDOS];
 790         int                     undo_alloc;
 791         TAILQ_HEAD(, hammer_undo)  undo_lru_list;
 792         TAILQ_HEAD(, hammer_reserve) delay_list;
 793         struct hammer_flush_group_list  flush_group_list;
 794         hammer_flush_group_t    next_flush_group;
 795         TAILQ_HEAD(, hammer_objid_cache) objid_cache_list;
 796         TAILQ_HEAD(, hammer_reclaim) reclaim_list;
 797 };
 798
 799 typedef struct hammer_mount     *hammer_mount_t;
 800
 801 #define HAMMER_MOUNT_CRITICAL_ERROR     0x0001
 802 #define HAMMER_MOUNT_FLUSH_RECOVERY     0x0002
 803
 804 struct hammer_sync_info {
 805         int error;
 806         int waitfor;
 807 };
 808
 809 #endif
 810
 811 /*
 812  * checkspace slop (8MB chunks), higher numbers are more conservative.
 813  */
 814 #define HAMMER_CHKSPC_REBLOCK   25
 815 #define HAMMER_CHKSPC_MIRROR    20
 816 #define HAMMER_CHKSPC_WRITE     20
 817 #define HAMMER_CHKSPC_CREATE    20
 818 #define HAMMER_CHKSPC_REMOVE    10
 819 #define HAMMER_CHKSPC_EMERGENCY 0
 820
 821 #if defined(_KERNEL)
 822
 823 extern struct vop_ops hammer_vnode_vops;
 824 extern struct vop_ops hammer_spec_vops;
 825 extern struct vop_ops hammer_fifo_vops;
 826 extern struct bio_ops hammer_bioops;
 827
 828 extern int hammer_debug_io;
 829 extern int hammer_debug_general;
 830 extern int hammer_debug_debug;
 831 extern int hammer_debug_inode;
 832 extern int hammer_debug_locks;
 833 extern int hammer_debug_btree;
 834 extern int hammer_debug_tid;
 835 extern int hammer_debug_recover;
 836 extern int hammer_debug_recover_faults;
 837 extern int hammer_debug_critical;
 838 extern int hammer_cluster_enable;
 839 extern int hammer_count_fsyncs;
 840 extern int hammer_count_inodes;
 841 extern int hammer_count_iqueued;
 842 extern int hammer_count_reclaiming;
 843 extern int hammer_count_records;
 844 extern int hammer_count_record_datas;
 845 extern int hammer_count_volumes;
 846 extern int hammer_count_buffers;
 847 extern int hammer_count_nodes;
 848 extern int64_t hammer_count_extra_space_used;
 849 extern int64_t hammer_stats_btree_lookups;
 850 extern int64_t hammer_stats_btree_searches;
 851 extern int64_t hammer_stats_btree_inserts;
 852 extern int64_t hammer_stats_btree_deletes;
 853 extern int64_t hammer_stats_btree_elements;
 854 extern int64_t hammer_stats_btree_splits;
 855 extern int64_t hammer_stats_btree_iterations;
 856 extern int64_t hammer_stats_btree_root_iterations;
 857 extern int64_t hammer_stats_record_iterations;
 858 extern int64_t hammer_stats_file_read;
 859 extern int64_t hammer_stats_file_write;
 860 extern int64_t hammer_stats_file_iopsr;
 861 extern int64_t hammer_stats_file_iopsw;
 862 extern int64_t hammer_stats_disk_read;
 863 extern int64_t hammer_stats_disk_write;
 864 extern int64_t hammer_stats_inode_flushes;
 865 extern int64_t hammer_stats_commits;
 866 extern int64_t hammer_stats_undo;
 867 extern int hammer_count_dirtybufspace;
 868 extern int hammer_count_refedbufs;
 869 extern int hammer_count_reservations;
 870 extern int hammer_count_io_running_read;
 871 extern int hammer_count_io_running_write;
 872 extern int hammer_count_io_locked;
 873 extern int hammer_limit_dirtybufspace;
 874 extern int hammer_limit_recs;
 875 extern int hammer_limit_inode_recs;
 876 extern int hammer_limit_reclaim;
 877 extern int hammer_bio_count;
 878 extern int hammer_verify_zone;
 879 extern int hammer_verify_data;
 880 extern int hammer_write_mode;
 881 extern int hammer_yield_check;
 882 extern int hammer_fsync_mode;
 883 extern int hammer_autoflush;
 884 extern int64_t hammer_contention_count;
 885
 886 void    hammer_critical_error(hammer_mount_t hmp, hammer_inode_t ip,
 887                         int error, const char *msg);
 888 int     hammer_vop_inactive(struct vop_inactive_args *);
 889 int     hammer_vop_reclaim(struct vop_reclaim_args *);
 890 int     hammer_get_vnode(struct hammer_inode *ip, struct vnode **vpp);
 891 struct hammer_inode *hammer_get_inode(hammer_transaction_t trans,
 892                         hammer_inode_t dip, int64_t obj_id,
 893                         hammer_tid_t asof, u_int32_t localization,
 894                         int flags, int *errorp);
 895 struct hammer_inode *hammer_get_dummy_inode(hammer_transaction_t trans,
 896                         hammer_inode_t dip, int64_t obj_id,
 897                         hammer_tid_t asof, u_int32_t localization,
 898                         int flags, int *errorp);
 899 struct hammer_inode *hammer_find_inode(hammer_transaction_t trans,
 900                         int64_t obj_id, hammer_tid_t asof,
 901                         u_int32_t localization);
 902 void    hammer_scan_inode_snapshots(hammer_mount_t hmp,
 903                         hammer_inode_info_t iinfo,
 904                         int (*callback)(hammer_inode_t ip, void *data),
 905                         void *data);
 906 void    hammer_put_inode(struct hammer_inode *ip);
 907 void    hammer_put_inode_ref(struct hammer_inode *ip);
 908 void    hammer_inode_waitreclaims(hammer_mount_t hmp);
 909 void    hammer_inode_waithard(hammer_mount_t hmp);
 910
 911 int     hammer_unload_volume(hammer_volume_t volume, void *data __unused);
 912 int     hammer_adjust_volume_mode(hammer_volume_t volume, void *data __unused);
 913
 914 int     hammer_unload_buffer(hammer_buffer_t buffer, void *data);
 915 int     hammer_install_volume(hammer_mount_t hmp, const char *volname,
 916                         struct vnode *devvp);
 917 int     hammer_mountcheck_volumes(hammer_mount_t hmp);
 918
 919 int     hammer_mem_add(hammer_record_t record);
 920 int     hammer_ip_lookup(hammer_cursor_t cursor);
 921 int     hammer_ip_first(hammer_cursor_t cursor);
 922 int     hammer_ip_next(hammer_cursor_t cursor);
 923 int     hammer_ip_resolve_data(hammer_cursor_t cursor);
 924 int     hammer_ip_delete_record(hammer_cursor_t cursor, hammer_inode_t ip,
 925                         hammer_tid_t tid);
 926 int     hammer_create_at_cursor(hammer_cursor_t cursor,
 927                         hammer_btree_leaf_elm_t leaf, void *udata, int mode);
 928 int     hammer_delete_at_cursor(hammer_cursor_t cursor, int delete_flags,
 929                         hammer_tid_t delete_tid, u_int32_t delete_ts,
 930                         int track, int64_t *stat_bytes);
 931 int     hammer_ip_check_directory_empty(hammer_transaction_t trans,
 932                         hammer_inode_t ip);
 933 int     hammer_sync_hmp(hammer_mount_t hmp, int waitfor);
 934 int     hammer_queue_inodes_flusher(hammer_mount_t hmp, int waitfor);
 935
 936 hammer_record_t
 937         hammer_alloc_mem_record(hammer_inode_t ip, int data_len);
 938 void    hammer_flush_record_done(hammer_record_t record, int error);
 939 void    hammer_wait_mem_record_ident(hammer_record_t record, const char *ident);
 940 void    hammer_rel_mem_record(hammer_record_t record);
 941
 942 int     hammer_cursor_up(hammer_cursor_t cursor);
 943 int     hammer_cursor_up_locked(hammer_cursor_t cursor);
 944 int     hammer_cursor_down(hammer_cursor_t cursor);
 945 int     hammer_cursor_upgrade(hammer_cursor_t cursor);
 946 int     hammer_cursor_upgrade_node(hammer_cursor_t cursor);
 947 void    hammer_cursor_downgrade(hammer_cursor_t cursor);
 948 int     hammer_cursor_seek(hammer_cursor_t cursor, hammer_node_t node,
 949                         int index);
 950 void    hammer_lock_ex_ident(struct hammer_lock *lock, const char *ident);
 951 int     hammer_lock_ex_try(struct hammer_lock *lock);
 952 void    hammer_lock_sh(struct hammer_lock *lock);
 953 int     hammer_lock_sh_try(struct hammer_lock *lock);
 954 int     hammer_lock_upgrade(struct hammer_lock *lock);
 955 void    hammer_lock_downgrade(struct hammer_lock *lock);
 956 int     hammer_lock_status(struct hammer_lock *lock);
 957 void    hammer_unlock(struct hammer_lock *lock);
 958 void    hammer_ref(struct hammer_lock *lock);
 959 void    hammer_unref(struct hammer_lock *lock);
 960
 961 void    hammer_sync_lock_ex(hammer_transaction_t trans);
 962 void    hammer_sync_lock_sh(hammer_transaction_t trans);
 963 int     hammer_sync_lock_sh_try(hammer_transaction_t trans);
 964 void    hammer_sync_unlock(hammer_transaction_t trans);
 965
 966 u_int32_t hammer_to_unix_xid(uuid_t *uuid);
 967 void hammer_guid_to_uuid(uuid_t *uuid, u_int32_t guid);
 968 void    hammer_time_to_timespec(u_int64_t xtime, struct timespec *ts);
 969 u_int64_t hammer_timespec_to_time(struct timespec *ts);
 970 int     hammer_str_to_tid(const char *str, int *ispfsp,
 971                         hammer_tid_t *tidp, u_int32_t *localizationp);
 972 int     hammer_is_atatext(const char *name, int len);
 973 hammer_tid_t hammer_alloc_objid(hammer_mount_t hmp, hammer_inode_t dip,
 974                         int64_t namekey);
 975 void hammer_clear_objid(hammer_inode_t dip);
 976 void hammer_destroy_objid_cache(hammer_mount_t hmp);
 977
 978 int hammer_enter_undo_history(hammer_mount_t hmp, hammer_off_t offset,
 979                         int bytes);
 980 void hammer_clear_undo_history(hammer_mount_t hmp);
 981 enum vtype hammer_get_vnode_type(u_int8_t obj_type);
 982 int hammer_get_dtype(u_int8_t obj_type);
 983 u_int8_t hammer_get_obj_type(enum vtype vtype);
 984 int64_t hammer_directory_namekey(hammer_inode_t dip, const void *name, int len,
 985                         u_int32_t *max_iterationsp);
 986 int     hammer_nohistory(hammer_inode_t ip);
 987
 988 int     hammer_init_cursor(hammer_transaction_t trans, hammer_cursor_t cursor,
 989                         hammer_node_cache_t cache, hammer_inode_t ip);
 990 void    hammer_normalize_cursor(hammer_cursor_t cursor);
 991 void    hammer_done_cursor(hammer_cursor_t cursor);
 992 int     hammer_recover_cursor(hammer_cursor_t cursor);
 993 void    hammer_unlock_cursor(hammer_cursor_t cursor);
 994 int     hammer_lock_cursor(hammer_cursor_t cursor);
 995 hammer_cursor_t hammer_push_cursor(hammer_cursor_t ocursor);
 996 void    hammer_pop_cursor(hammer_cursor_t ocursor, hammer_cursor_t ncursor);
 997
 998 void    hammer_cursor_replaced_node(hammer_node_t onode, hammer_node_t nnode);
 999 void    hammer_cursor_removed_node(hammer_node_t onode, hammer_node_t parent,
1000                         int index);
1001 void    hammer_cursor_split_node(hammer_node_t onode, hammer_node_t nnode,
1002                         int index);
1003 void    hammer_cursor_moved_element(hammer_node_t onode, hammer_node_t nnode,
1004                         int oindex, int nindex);
1005 void    hammer_cursor_parent_changed(hammer_node_t node, hammer_node_t oparent,
1006                         hammer_node_t nparent, int nindex);
1007 void    hammer_cursor_inserted_element(hammer_node_t node, int index);
1008 void    hammer_cursor_deleted_element(hammer_node_t node, int index);
1009
1010 int     hammer_btree_lookup(hammer_cursor_t cursor);
1011 int     hammer_btree_first(hammer_cursor_t cursor);
1012 int     hammer_btree_last(hammer_cursor_t cursor);
1013 int     hammer_btree_extract(hammer_cursor_t cursor, int flags);
1014 int     hammer_btree_iterate(hammer_cursor_t cursor);
1015 int     hammer_btree_iterate_reverse(hammer_cursor_t cursor);
1016 int     hammer_btree_insert(hammer_cursor_t cursor,
1017                             hammer_btree_leaf_elm_t elm, int *doprop);
1018 int     hammer_btree_delete(hammer_cursor_t cursor);
1019 void    hammer_btree_do_propagation(hammer_cursor_t cursor,
1020                             hammer_pseudofs_inmem_t pfsm,
1021                             hammer_btree_leaf_elm_t leaf);
1022 int     hammer_btree_cmp(hammer_base_elm_t key1, hammer_base_elm_t key2);
1023 int     hammer_btree_chkts(hammer_tid_t ts, hammer_base_elm_t key);
1024 int     hammer_btree_correct_rhb(hammer_cursor_t cursor, hammer_tid_t tid);
1025 int     hammer_btree_correct_lhb(hammer_cursor_t cursor, hammer_tid_t tid);
1026
1027 int     btree_set_parent(hammer_transaction_t trans, hammer_node_t node,
1028                         hammer_btree_elm_t elm);
1029 void    hammer_node_lock_init(hammer_node_lock_t parent, hammer_node_t node);
1030 int     hammer_btree_lock_children(hammer_cursor_t cursor, int depth,
1031                         hammer_node_lock_t parent);
1032 void    hammer_btree_lock_copy(hammer_cursor_t cursor,
1033                         hammer_node_lock_t parent);
1034 int     hammer_btree_sync_copy(hammer_cursor_t cursor,
1035                         hammer_node_lock_t parent);
1036 void    hammer_btree_unlock_children(hammer_cursor_t cursor,
1037                         hammer_node_lock_t parent);
1038 int     hammer_btree_search_node(hammer_base_elm_t elm, hammer_node_ondisk_t node);
1039 hammer_node_t hammer_btree_get_parent(hammer_transaction_t trans,
1040                         hammer_node_t node, int *parent_indexp,
1041                         int *errorp, int try_exclusive);
1042
1043 void    hammer_print_btree_node(hammer_node_ondisk_t ondisk);
1044 void    hammer_print_btree_elm(hammer_btree_elm_t elm, u_int8_t type, int i);
1045
1046 void    *hammer_bread(struct hammer_mount *hmp, hammer_off_t off,
1047                         int *errorp, struct hammer_buffer **bufferp);
1048 void    *hammer_bnew(struct hammer_mount *hmp, hammer_off_t off,
1049                         int *errorp, struct hammer_buffer **bufferp);
1050 void    *hammer_bread_ext(struct hammer_mount *hmp, hammer_off_t off, int bytes,
1051                         int *errorp, struct hammer_buffer **bufferp);
1052 void    *hammer_bnew_ext(struct hammer_mount *hmp, hammer_off_t off, int bytes,
1053                         int *errorp, struct hammer_buffer **bufferp);
1054
1055 hammer_volume_t hammer_get_root_volume(hammer_mount_t hmp, int *errorp);
1056
1057 hammer_volume_t hammer_get_volume(hammer_mount_t hmp,
1058                         int32_t vol_no, int *errorp);
1059 hammer_buffer_t hammer_get_buffer(hammer_mount_t hmp, hammer_off_t buf_offset,
1060                         int bytes, int isnew, int *errorp);
1061 void            hammer_sync_buffers(hammer_mount_t hmp,
1062                         hammer_off_t base_offset, int bytes);
1063 int             hammer_del_buffers(hammer_mount_t hmp,
1064                         hammer_off_t base_offset,
1065                         hammer_off_t zone2_offset, int bytes,
1066                         int report_conflicts);
1067
1068 int             hammer_ref_volume(hammer_volume_t volume);
1069 int             hammer_ref_buffer(hammer_buffer_t buffer);
1070 void            hammer_flush_buffer_nodes(hammer_buffer_t buffer);
1071
1072 void            hammer_rel_volume(hammer_volume_t volume, int flush);
1073 void            hammer_rel_buffer(hammer_buffer_t buffer, int flush);
1074
1075 int             hammer_vfs_export(struct mount *mp, int op,
1076                         const struct export_args *export);
1077 hammer_node_t   hammer_get_node(hammer_transaction_t trans,
1078                         hammer_off_t node_offset, int isnew, int *errorp);
1079 void            hammer_ref_node(hammer_node_t node);
1080 hammer_node_t   hammer_ref_node_safe(hammer_transaction_t trans,
1081                         hammer_node_cache_t cache, int *errorp);
1082 void            hammer_rel_node(hammer_node_t node);
1083 void            hammer_delete_node(hammer_transaction_t trans,
1084                         hammer_node_t node);
1085 void            hammer_cache_node(hammer_node_cache_t cache,
1086                         hammer_node_t node);
1087 void            hammer_uncache_node(hammer_node_cache_t cache);
1088 void            hammer_flush_node(hammer_node_t node);
1089
1090 void hammer_dup_buffer(struct hammer_buffer **bufferp,
1091                         struct hammer_buffer *buffer);
1092 hammer_node_t hammer_alloc_btree(hammer_transaction_t trans,
1093                         hammer_off_t hint, int *errorp);
1094 void *hammer_alloc_data(hammer_transaction_t trans, int32_t data_len,
1095                         u_int16_t rec_type, hammer_off_t *data_offsetp,
1096                         struct hammer_buffer **data_bufferp,
1097                         hammer_off_t hint, int *errorp);
1098
1099 int hammer_generate_undo(hammer_transaction_t trans,
1100                         hammer_off_t zone1_offset, void *base, int len);
1101 int hammer_upgrade_undo_4(hammer_transaction_t trans);
1102
1103 void hammer_put_volume(struct hammer_volume *volume, int flush);
1104 void hammer_put_buffer(struct hammer_buffer *buffer, int flush);
1105
1106 hammer_off_t hammer_freemap_alloc(hammer_transaction_t trans,
1107                         hammer_off_t owner, int *errorp);
1108 void hammer_freemap_free(hammer_transaction_t trans, hammer_off_t phys_offset,
1109                         hammer_off_t owner, int *errorp);
1110 int _hammer_checkspace(hammer_mount_t hmp, int slop, int64_t *resp);
1111 hammer_off_t hammer_blockmap_alloc(hammer_transaction_t trans, int zone,
1112                         int bytes, hammer_off_t hint, int *errorp);
1113 hammer_reserve_t hammer_blockmap_reserve(hammer_mount_t hmp, int zone,
1114                         int bytes, hammer_off_t *zone_offp, int *errorp);
1115 void hammer_blockmap_reserve_complete(hammer_mount_t hmp,
1116                         hammer_reserve_t resv);
1117 void hammer_reserve_clrdelay(hammer_mount_t hmp, hammer_reserve_t resv);
1118 void hammer_blockmap_free(hammer_transaction_t trans,
1119                         hammer_off_t bmap_off, int bytes);
1120 int hammer_blockmap_finalize(hammer_transaction_t trans,
1121                         hammer_reserve_t resv,
1122                         hammer_off_t bmap_off, int bytes);
1123 int hammer_blockmap_getfree(hammer_mount_t hmp, hammer_off_t bmap_off,
1124                         int *curp, int *errorp);
1125 hammer_off_t hammer_blockmap_lookup(hammer_mount_t hmp, hammer_off_t bmap_off,
1126                         int *errorp);
1127 hammer_off_t hammer_undo_lookup(hammer_mount_t hmp, hammer_off_t bmap_off,
1128                         int *errorp);
1129 int64_t hammer_undo_used(hammer_transaction_t trans);
1130 int64_t hammer_undo_space(hammer_transaction_t trans);
1131 int64_t hammer_undo_max(hammer_mount_t hmp);
1132 int hammer_undo_reclaim(hammer_io_t io);
1133
1134 void hammer_start_transaction(struct hammer_transaction *trans,
1135                               struct hammer_mount *hmp);
1136 void hammer_simple_transaction(struct hammer_transaction *trans,
1137                               struct hammer_mount *hmp);
1138 void hammer_start_transaction_fls(struct hammer_transaction *trans,
1139                                   struct hammer_mount *hmp);
1140 void hammer_done_transaction(struct hammer_transaction *trans);
1141 hammer_tid_t hammer_alloc_tid(hammer_mount_t hmp, int count);
1142
1143 void hammer_modify_inode(hammer_inode_t ip, int flags);
1144 void hammer_flush_inode(hammer_inode_t ip, int flags);
1145 void hammer_flush_inode_done(hammer_inode_t ip, int error);
1146 void hammer_wait_inode(hammer_inode_t ip);
1147
1148 int  hammer_create_inode(struct hammer_transaction *trans, struct vattr *vap,
1149                         struct ucred *cred, struct hammer_inode *dip,
1150                         const char *name, int namelen,
1151                         hammer_pseudofs_inmem_t pfsm,
1152                         struct hammer_inode **ipp);
1153 void hammer_rel_inode(hammer_inode_t ip, int flush);
1154 int hammer_reload_inode(hammer_inode_t ip, void *arg __unused);
1155 int hammer_ino_rb_compare(hammer_inode_t ip1, hammer_inode_t ip2);
1156 int hammer_destroy_inode_callback(hammer_inode_t ip, void *data __unused);
1157
1158 int hammer_sync_inode(hammer_transaction_t trans, hammer_inode_t ip);
1159 void hammer_test_inode(hammer_inode_t dip);
1160 void hammer_inode_unloadable_check(hammer_inode_t ip, int getvp);
1161
1162 int  hammer_ip_add_directory(struct hammer_transaction *trans,
1163                         hammer_inode_t dip, const char *name, int bytes,
1164                         hammer_inode_t nip);
1165 int  hammer_ip_del_directory(struct hammer_transaction *trans,
1166                         hammer_cursor_t cursor, hammer_inode_t dip,
1167                         hammer_inode_t ip);
1168 hammer_record_t hammer_ip_add_bulk(hammer_inode_t ip, off_t file_offset,
1169                         void *data, int bytes, int *errorp);
1170 int  hammer_ip_frontend_trunc(struct hammer_inode *ip, off_t file_size);
1171 int  hammer_ip_add_record(struct hammer_transaction *trans,
1172                         hammer_record_t record);
1173 int  hammer_ip_delete_range(hammer_cursor_t cursor, hammer_inode_t ip,
1174                         int64_t ran_beg, int64_t ran_end, int truncating);
1175 int  hammer_ip_delete_clean(hammer_cursor_t cursor, hammer_inode_t ip,
1176                         int *countp);
1177 int  hammer_ip_sync_data(hammer_cursor_t cursor, hammer_inode_t ip,
1178                         int64_t offset, void *data, int bytes);
1179 int  hammer_ip_sync_record(hammer_transaction_t trans, hammer_record_t rec);
1180 int  hammer_ip_sync_record_cursor(hammer_cursor_t cursor, hammer_record_t rec);
1181 hammer_pseudofs_inmem_t  hammer_load_pseudofs(hammer_transaction_t trans,
1182                         u_int32_t localization, int *errorp);
1183 int  hammer_mkroot_pseudofs(hammer_transaction_t trans, struct ucred *cred,
1184                         hammer_pseudofs_inmem_t pfsm);
1185 int  hammer_save_pseudofs(hammer_transaction_t trans,
1186                         hammer_pseudofs_inmem_t pfsm);
1187 int  hammer_unload_pseudofs(hammer_transaction_t trans, u_int32_t localization);
1188 void hammer_rel_pseudofs(hammer_mount_t hmp, hammer_pseudofs_inmem_t pfsm);
1189 int hammer_ioctl(hammer_inode_t ip, u_long com, caddr_t data, int fflag,
1190                         struct ucred *cred);
1191
1192 void hammer_io_init(hammer_io_t io, hammer_volume_t volume,
1193                         enum hammer_io_type type);
1194 int hammer_io_read(struct vnode *devvp, struct hammer_io *io,
1195                         hammer_off_t limit);
1196 void hammer_io_advance(struct hammer_io *io);
1197 int hammer_io_new(struct vnode *devvp, struct hammer_io *io);
1198 int hammer_io_inval(hammer_volume_t volume, hammer_off_t zone2_offset);
1199 struct buf *hammer_io_release(struct hammer_io *io, int flush);
1200 void hammer_io_flush(struct hammer_io *io, int reclaim);
1201 void hammer_io_wait(struct hammer_io *io);
1202 void hammer_io_waitdep(struct hammer_io *io);
1203 void hammer_io_wait_all(hammer_mount_t hmp, const char *ident);
1204 int hammer_io_direct_read(hammer_mount_t hmp, struct bio *bio,
1205                         hammer_btree_leaf_elm_t leaf);
1206 int hammer_io_direct_write(hammer_mount_t hmp, hammer_record_t record,
1207                         struct bio *bio);
1208 void hammer_io_direct_wait(hammer_record_t record);
1209 void hammer_io_direct_uncache(hammer_mount_t hmp, hammer_btree_leaf_elm_t leaf);
1210 void hammer_io_write_interlock(hammer_io_t io);
1211 void hammer_io_done_interlock(hammer_io_t io);
1212 void hammer_io_clear_modify(struct hammer_io *io, int inval);
1213 void hammer_io_clear_modlist(struct hammer_io *io);
1214 void hammer_io_flush_sync(hammer_mount_t hmp);
1215 void hammer_io_clear_error(struct hammer_io *io);
1216
1217 void hammer_modify_volume(hammer_transaction_t trans, hammer_volume_t volume,
1218                         void *base, int len);
1219 void hammer_modify_buffer(hammer_transaction_t trans, hammer_buffer_t buffer,
1220                         void *base, int len);
1221 void hammer_modify_volume_done(hammer_volume_t volume);
1222 void hammer_modify_buffer_done(hammer_buffer_t buffer);
1223
1224 int hammer_ioc_reblock(hammer_transaction_t trans, hammer_inode_t ip,
1225                         struct hammer_ioc_reblock *reblock);
1226 int hammer_ioc_rebalance(hammer_transaction_t trans, hammer_inode_t ip,
1227                         struct hammer_ioc_rebalance *rebal);
1228 int hammer_ioc_prune(hammer_transaction_t trans, hammer_inode_t ip,
1229                         struct hammer_ioc_prune *prune);
1230 int hammer_ioc_mirror_read(hammer_transaction_t trans, hammer_inode_t ip,
1231                         struct hammer_ioc_mirror_rw *mirror);
1232 int hammer_ioc_mirror_write(hammer_transaction_t trans, hammer_inode_t ip,
1233                         struct hammer_ioc_mirror_rw *mirror);
1234 int hammer_ioc_set_pseudofs(hammer_transaction_t trans, hammer_inode_t ip,
1235                         struct ucred *cred, struct hammer_ioc_pseudofs_rw *pfs);
1236 int hammer_ioc_get_pseudofs(hammer_transaction_t trans, hammer_inode_t ip,
1237                         struct hammer_ioc_pseudofs_rw *pfs);
1238 int hammer_ioc_destroy_pseudofs(hammer_transaction_t trans, hammer_inode_t ip,
1239                         struct hammer_ioc_pseudofs_rw *pfs);
1240 int hammer_ioc_downgrade_pseudofs(hammer_transaction_t trans, hammer_inode_t ip,
1241                         struct hammer_ioc_pseudofs_rw *pfs);
1242 int hammer_ioc_upgrade_pseudofs(hammer_transaction_t trans, hammer_inode_t ip,
1243                         struct hammer_ioc_pseudofs_rw *pfs);
1244 int hammer_ioc_wait_pseudofs(hammer_transaction_t trans, hammer_inode_t ip,
1245                         struct hammer_ioc_pseudofs_rw *pfs);
1246 int hammer_ioc_volume_add(hammer_transaction_t trans, hammer_inode_t ip,
1247                         struct hammer_ioc_volume *ioc);
1248 int hammer_ioc_volume_del(hammer_transaction_t trans, hammer_inode_t ip,
1249                         struct hammer_ioc_volume *ioc);
1250
1251 int hammer_signal_check(hammer_mount_t hmp);
1252
1253 void hammer_flusher_create(hammer_mount_t hmp);
1254 void hammer_flusher_destroy(hammer_mount_t hmp);
1255 void hammer_flusher_sync(hammer_mount_t hmp);
1256 int  hammer_flusher_async(hammer_mount_t hmp, hammer_flush_group_t flg);
1257 int  hammer_flusher_async_one(hammer_mount_t hmp);
1258 void hammer_flusher_wait(hammer_mount_t hmp, int seq);
1259 void hammer_flusher_wait_next(hammer_mount_t hmp);
1260 int  hammer_flusher_meta_limit(hammer_mount_t hmp);
1261 int  hammer_flusher_meta_halflimit(hammer_mount_t hmp);
1262 int  hammer_flusher_undo_exhausted(hammer_transaction_t trans, int quarter);
1263 void hammer_flusher_clean_loose_ios(hammer_mount_t hmp);
1264 void hammer_flusher_finalize(hammer_transaction_t trans, int final);
1265 int  hammer_flusher_haswork(hammer_mount_t hmp);
1266
1267
1268 int hammer_recover_stage1(hammer_mount_t hmp, hammer_volume_t rootvol);
1269 int hammer_recover_stage2(hammer_mount_t hmp, hammer_volume_t rootvol);
1270 void hammer_recover_flush_buffers(hammer_mount_t hmp,
1271                         hammer_volume_t root_volume, int final);
1272
1273 void hammer_crc_set_blockmap(hammer_blockmap_t blockmap);
1274 void hammer_crc_set_volume(hammer_volume_ondisk_t ondisk);
1275 void hammer_crc_set_leaf(void *data, hammer_btree_leaf_elm_t leaf);
1276
1277 int hammer_crc_test_blockmap(hammer_blockmap_t blockmap);
1278 int hammer_crc_test_volume(hammer_volume_ondisk_t ondisk);
1279 int hammer_crc_test_btree(hammer_node_ondisk_t ondisk);
1280 int hammer_crc_test_leaf(void *data, hammer_btree_leaf_elm_t leaf);
1281 void hkprintf(const char *ctl, ...);
1282 udev_t hammer_fsid_to_udev(uuid_t *uuid);
1283
1284
1285 int hammer_blocksize(int64_t file_offset);
1286 int64_t hammer_blockdemarc(int64_t file_offset1, int64_t file_offset2);
1287
1288 /*
1289  * Shortcut for _hammer_checkspace(), used all over the code.
1290  */
1291 static __inline int
1292 hammer_checkspace(hammer_mount_t hmp, int slop)
1293 {
1294         return(_hammer_checkspace(hmp, slop, NULL));
1295 }
1296
1297 #endif
1298
1299 static __inline void
1300 hammer_wait_mem_record(hammer_record_t record)
1301 {
1302         hammer_wait_mem_record_ident(record, "hmmwai");
1303 }
1304
1305 static __inline void
1306 hammer_lock_ex(struct hammer_lock *lock)
1307 {
1308         hammer_lock_ex_ident(lock, "hmrlck");
1309 }
1310
1311 /*
1312  * Indicate that a B-Tree node is being modified.
1313  */
1314 static __inline void
1315 hammer_modify_node_noundo(hammer_transaction_t trans, hammer_node_t node)
1316 {
1317         KKASSERT((node->flags & HAMMER_NODE_CRCBAD) == 0);
1318         hammer_modify_buffer(trans, node->buffer, NULL, 0);
1319 }
1320
1321 static __inline void
1322 hammer_modify_node_all(hammer_transaction_t trans, struct hammer_node *node)
1323 {
1324         KKASSERT((node->flags & HAMMER_NODE_CRCBAD) == 0);
1325         hammer_modify_buffer(trans, node->buffer,
1326                              node->ondisk, sizeof(*node->ondisk));
1327 }
1328
1329 static __inline void
1330 hammer_modify_node(hammer_transaction_t trans, hammer_node_t node,
1331                    void *base, int len)
1332 {
1333         hammer_crc_t *crcptr;
1334
1335         KKASSERT((char *)base >= (char *)node->ondisk &&
1336                  (char *)base + len <=
1337                     (char *)node->ondisk + sizeof(*node->ondisk));
1338         KKASSERT((node->flags & HAMMER_NODE_CRCBAD) == 0);
1339         hammer_modify_buffer(trans, node->buffer, base, len);
1340         crcptr = &node->ondisk->crc;
1341         hammer_modify_buffer(trans, node->buffer, crcptr, sizeof(hammer_crc_t));
1342         --node->buffer->io.modify_refs; /* only want one ref */
1343 }
1344
1345 /*
1346  * Indicate that the specified modifications have been completed.
1347  *
1348  * Do not try to generate the crc here, it's very expensive to do and a
1349  * sequence of insertions or deletions can result in many calls to this
1350  * function on the same node.
1351  */
1352 static __inline void
1353 hammer_modify_node_done(hammer_node_t node)
1354 {
1355         node->flags |= HAMMER_NODE_CRCGOOD;
1356         if ((node->flags & HAMMER_NODE_NEEDSCRC) == 0) {
1357                 node->flags |= HAMMER_NODE_NEEDSCRC;
1358                 node->buffer->io.gencrc = 1;
1359                 hammer_ref_node(node);
1360         }
1361         hammer_modify_buffer_done(node->buffer);
1362 }
1363
1364 #define hammer_modify_volume_field(trans, vol, field)           \
1365         hammer_modify_volume(trans, vol, &(vol)->ondisk->field, \
1366                              sizeof((vol)->ondisk->field))
1367
1368 #define hammer_modify_node_field(trans, node, field)            \
1369         hammer_modify_node(trans, node, &(node)->ondisk->field, \
1370                              sizeof((node)->ondisk->field))
1371
1372 /*
1373  * The HAMMER_INODE_CAP_DIR_LOCAL_INO capability is set on newly
1374  * created directories for HAMMER version 2 or greater and causes
1375  * directory entries to be placed the inode localization zone in
1376  * the B-Tree instead of the misc zone.
1377  *
1378  * This greatly improves localization between directory entries and
1379  * inodes
1380  */
1381 static __inline u_int32_t
1382 hammer_dir_localization(hammer_inode_t dip)
1383 {
1384         if (dip->ino_data.cap_flags & HAMMER_INODE_CAP_DIR_LOCAL_INO)
1385                 return(HAMMER_LOCALIZE_INODE);
1386         else
1387                 return(HAMMER_LOCALIZE_MISC);
1388 }