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 1.93 2008/06/26 04:06:22 dillon Exp $
  35  */
  36 /*
  37  * This header file contains structures used internally by the HAMMERFS
  38  * implementation.  See hammer_disk.h for on-disk structures.
  39  */
  40
  41 #include <sys/param.h>
  42 #include <sys/types.h>
  43 #include <sys/kernel.h>
  44 #include <sys/conf.h>
  45 #include <sys/systm.h>
  46 #include <sys/tree.h>
  47 #include <sys/malloc.h>
  48 #include <sys/mount.h>
  49 #include <sys/mountctl.h>
  50 #include <sys/vnode.h>
  51 #include <sys/proc.h>
  52 #include <sys/stat.h>
  53 #include <sys/globaldata.h>
  54 #include <sys/lockf.h>
  55 #include <sys/buf.h>
  56 #include <sys/queue.h>
  57 #include <sys/ktr.h>
  58 #include <sys/globaldata.h>
  59
  60 #include <sys/buf2.h>
  61 #include <sys/signal2.h>
  62 #include "hammer_disk.h"
  63 #include "hammer_mount.h"
  64 #include "hammer_ioctl.h"
  65
  66 #if defined(_KERNEL) || defined(_KERNEL_STRUCTURES)
  67
  68 MALLOC_DECLARE(M_HAMMER);
  69
  70 /*
  71  * Kernel trace
  72  */
  73 #if !defined(KTR_HAMMER)
  74 #define KTR_HAMMER      KTR_ALL
  75 #endif
  76 KTR_INFO_MASTER_EXTERN(hammer);
  77
  78 /*
  79  * Misc structures
  80  */
  81 struct hammer_mount;
  82
  83 /*
  84  * Key structure used for custom RB tree inode lookups.  This prototypes
  85  * the function hammer_ino_rb_tree_RB_LOOKUP_INFO(root, info).
  86  */
  87 typedef struct hammer_inode_info {
  88         int64_t         obj_id;         /* (key) object identifier */
  89         hammer_tid_t    obj_asof;       /* (key) snapshot transid or 0 */
  90         u_int32_t       obj_localization; /* (key) pseudo-fs */
  91         union {
  92                 struct hammer_btree_leaf_elm *leaf;
  93         } u;
  94 } *hammer_inode_info_t;
  95
  96 typedef enum hammer_transaction_type {
  97         HAMMER_TRANS_RO,
  98         HAMMER_TRANS_STD,
  99         HAMMER_TRANS_FLS
 100 } hammer_transaction_type_t;
 101
 102 /*
 103  * HAMMER Transaction tracking
 104  */
 105 struct hammer_transaction {
 106         hammer_transaction_type_t type;
 107         struct hammer_mount *hmp;
 108         hammer_tid_t    tid;
 109         u_int64_t       time;
 110         u_int32_t       time32;
 111         int             sync_lock_refs;
 112         struct hammer_volume *rootvol;
 113 };
 114
 115 typedef struct hammer_transaction *hammer_transaction_t;
 116
 117 /*
 118  * HAMMER locks
 119  */
 120 struct hammer_lock {
 121         int     refs;           /* active references delay writes */
 122         int     lockcount;      /* lock count for exclusive/shared access */
 123         int     wanted;
 124         int     exwanted;       /* number of threads waiting for ex lock */
 125         struct thread *locktd;
 126 };
 127
 128 static __inline int
 129 hammer_islocked(struct hammer_lock *lock)
 130 {
 131         return(lock->lockcount != 0);
 132 }
 133
 134 static __inline int
 135 hammer_isactive(struct hammer_lock *lock)
 136 {
 137         return(lock->refs != 0);
 138 }
 139
 140 static __inline int
 141 hammer_islastref(struct hammer_lock *lock)
 142 {
 143         return(lock->refs == 1);
 144 }
 145
 146 /*
 147  * Return if we specifically own the lock exclusively.
 148  */
 149 static __inline int
 150 hammer_lock_excl_owned(struct hammer_lock *lock, thread_t td)
 151 {
 152         if (lock->lockcount > 0 && lock->locktd == td)
 153                 return(1);
 154         return(0);
 155 }
 156
 157 /*
 158  * Flush state, used by various structures
 159  */
 160 typedef enum hammer_inode_state {
 161         HAMMER_FST_IDLE,
 162         HAMMER_FST_SETUP,
 163         HAMMER_FST_FLUSH
 164 } hammer_inode_state_t;
 165
 166 TAILQ_HEAD(hammer_record_list, hammer_record);
 167
 168 /*
 169  * Cache object ids.  A fixed number of objid cache structures are
 170  * created to reserve object id's for newly created files in multiples
 171  * of 100,000, localized to a particular directory, and recycled as
 172  * needed.  This allows parallel create operations in different
 173  * directories to retain fairly localized object ids which in turn
 174  * improves reblocking performance and layout.
 175  */
 176 #define OBJID_CACHE_SIZE        1024
 177 #define OBJID_CACHE_BULK        100000
 178
 179 typedef struct hammer_objid_cache {
 180         TAILQ_ENTRY(hammer_objid_cache) entry;
 181         struct hammer_inode             *dip;
 182         hammer_tid_t                    next_tid;
 183         int                             count;
 184 } *hammer_objid_cache_t;
 185
 186 /*
 187  * Associate an inode with a B-Tree node to cache search start positions
 188  */
 189 typedef struct hammer_node_cache {
 190         TAILQ_ENTRY(hammer_node_cache) entry;
 191         struct hammer_node              *node;
 192         struct hammer_inode             *ip;
 193 } *hammer_node_cache_t;
 194
 195 TAILQ_HEAD(hammer_node_cache_list, hammer_node_cache);
 196
 197 /*
 198  * Structure used to represent an inode in-memory.
 199  *
 200  * The record and data associated with an inode may be out of sync with
 201  * the disk (xDIRTY flags), or not even on the disk at all (ONDISK flag
 202  * clear).
 203  *
 204  * An inode may also hold a cache of unsynchronized records, used for
 205  * database and directories only.  Unsynchronized regular file data is
 206  * stored in the buffer cache.
 207  *
 208  * NOTE: A file which is created and destroyed within the initial
 209  * synchronization period can wind up not doing any disk I/O at all.
 210  *
 211  * Finally, an inode may cache numerous disk-referencing B-Tree cursors.
 212  */
 213 struct hammer_ino_rb_tree;
 214 struct hammer_inode;
 215 RB_HEAD(hammer_ino_rb_tree, hammer_inode);
 216 RB_PROTOTYPEX(hammer_ino_rb_tree, INFO, hammer_inode, rb_node,
 217               hammer_ino_rb_compare, hammer_inode_info_t);
 218
 219 struct hammer_rec_rb_tree;
 220 struct hammer_record;
 221 RB_HEAD(hammer_rec_rb_tree, hammer_record);
 222 RB_PROTOTYPEX(hammer_rec_rb_tree, INFO, hammer_record, rb_node,
 223               hammer_rec_rb_compare, hammer_btree_leaf_elm_t);
 224
 225 TAILQ_HEAD(hammer_node_list, hammer_node);
 226
 227 struct hammer_inode {
 228         RB_ENTRY(hammer_inode)  rb_node;
 229         hammer_inode_state_t    flush_state;
 230         int                     flush_group;
 231         TAILQ_ENTRY(hammer_inode) flush_entry;
 232         struct hammer_record_list target_list;  /* target of dependant recs */
 233         u_int64_t               obj_id;         /* (key) object identifier */
 234         hammer_tid_t            obj_asof;       /* (key) snapshot or 0 */
 235         u_int32_t               obj_localization; /* (key) pseudo-fs */
 236         struct hammer_mount     *hmp;
 237         hammer_objid_cache_t    objid_cache;
 238         int                     flags;
 239         int                     error;          /* flush error */
 240         int                     cursor_ip_refs; /* sanity */
 241         int                     rsv_databufs;
 242         int                     rsv_recs;
 243         struct vnode            *vp;
 244         struct lockf            advlock;
 245         struct hammer_lock      lock;           /* sync copy interlock */
 246         off_t                   trunc_off;
 247         struct hammer_btree_leaf_elm ino_leaf;  /* in-memory cache */
 248         struct hammer_inode_data ino_data;      /* in-memory cache */
 249         struct hammer_rec_rb_tree rec_tree;     /* in-memory cache */
 250         struct hammer_node_cache cache[2];      /* search initiate cache */
 251
 252         /*
 253          * When a demark is created to synchronize an inode to
 254          * disk, certain fields are copied so the front-end VOPs
 255          * can continue to run in parallel with the synchronization
 256          * occuring in the background.
 257          */
 258         int             sync_flags;             /* to-sync flags cache */
 259         off_t           sync_trunc_off;         /* to-sync truncation */
 260         struct hammer_btree_leaf_elm sync_ino_leaf; /* to-sync cache */
 261         struct hammer_inode_data sync_ino_data; /* to-sync cache */
 262 };
 263
 264 typedef struct hammer_inode *hammer_inode_t;
 265
 266 #define VTOI(vp)        ((struct hammer_inode *)(vp)->v_data)
 267
 268 #define HAMMER_INODE_DDIRTY     0x0001  /* in-memory ino_data is dirty */
 269                                         /* (not including atime/mtime) */
 270 #define HAMMER_INODE_RSV_INODES 0x0002  /* hmp->rsv_inodes bumped */
 271 #define HAMMER_INODE_UNUSED0004 0x0004
 272 #define HAMMER_INODE_XDIRTY     0x0008  /* in-memory records */
 273 #define HAMMER_INODE_ONDISK     0x0010  /* inode is on-disk (else not yet) */
 274 #define HAMMER_INODE_FLUSH      0x0020  /* flush on last ref */
 275 #define HAMMER_INODE_DELETED    0x0080  /* inode delete (backend) */
 276 #define HAMMER_INODE_DELONDISK  0x0100  /* delete synchronized to disk */
 277 #define HAMMER_INODE_RO         0x0200  /* read-only (because of as-of) */
 278 #define HAMMER_INODE_VHELD      0x0400  /* vnode held on sync */
 279 #define HAMMER_INODE_DONDISK    0x0800  /* data records may be on disk */
 280 #define HAMMER_INODE_BUFS       0x1000  /* dirty high level bps present */
 281 #define HAMMER_INODE_REFLUSH    0x2000  /* pipelined flush during flush */
 282 #define HAMMER_INODE_RECLAIM    0x4000  /* trying to reclaim */
 283 #define HAMMER_INODE_FLUSHW     0x8000  /* Someone waiting for flush */
 284
 285 #define HAMMER_INODE_TRUNCATED  0x00010000
 286 #define HAMMER_INODE_DELETING   0x00020000 /* inode delete request (frontend)*/
 287 #define HAMMER_INODE_RESIGNAL   0x00040000 /* re-signal on re-flush */
 288 #define HAMMER_INODE_ATIME      0x00100000 /* in-memory atime modified */
 289 #define HAMMER_INODE_MTIME      0x00100000 /* in-memory mtime modified */
 290
 291 #define HAMMER_INODE_MODMASK    (HAMMER_INODE_DDIRTY|                       \
 292                                  HAMMER_INODE_XDIRTY|HAMMER_INODE_BUFS|     \
 293                                  HAMMER_INODE_ATIME|HAMMER_INODE_MTIME|     \
 294                                  HAMMER_INODE_TRUNCATED|HAMMER_INODE_DELETING)
 295
 296 #define HAMMER_INODE_MODMASK_NOXDIRTY \
 297                                 (HAMMER_INODE_MODMASK & ~HAMMER_INODE_XDIRTY)
 298
 299 #define HAMMER_FLUSH_GROUP_SIZE 64
 300
 301 #define HAMMER_FLUSH_SIGNAL     0x0001
 302 #define HAMMER_FLUSH_RECURSION  0x0002
 303
 304 /*
 305  * Used by the inode reclaim code to pipeline reclaims and avoid
 306  * blowing out kernel memory or letting the flusher get too far
 307  * behind.
 308  */
 309 struct hammer_reclaim {
 310         TAILQ_ENTRY(hammer_reclaim) entry;
 311         int     okydoky;
 312 };
 313
 314 #define HAMMER_RECLAIM_FLUSH    2000
 315 #define HAMMER_RECLAIM_WAIT     4000
 316
 317 /*
 318  * Structure used to represent an unsynchronized record in-memory.  These
 319  * records typically represent directory entries.  Only non-historical
 320  * records are kept in-memory.
 321  *
 322  * Records are organized as a per-inode RB-Tree.  If the inode is not
 323  * on disk then neither are any records and the in-memory record tree
 324  * represents the entire contents of the inode.  If the inode is on disk
 325  * then the on-disk B-Tree is scanned in parallel with the in-memory
 326  * RB-Tree to synthesize the current state of the file.
 327  *
 328  * Records are also used to enforce the ordering of directory create/delete
 329  * operations.  A new inode will not be flushed to disk unless its related
 330  * directory entry is also being flushed at the same time.  A directory entry
 331  * will not be removed unless its related inode is also being removed at the
 332  * same time.
 333  */
 334 typedef enum hammer_record_type {
 335         HAMMER_MEM_RECORD_GENERAL,      /* misc record */
 336         HAMMER_MEM_RECORD_INODE,        /* inode record */
 337         HAMMER_MEM_RECORD_ADD,          /* positive memory cache record */
 338         HAMMER_MEM_RECORD_DEL,          /* negative delete-on-disk record */
 339         HAMMER_MEM_RECORD_DATA          /* bulk-data record w/on-disk ref */
 340 } hammer_record_type_t;
 341
 342 struct hammer_record {
 343         RB_ENTRY(hammer_record)         rb_node;
 344         TAILQ_ENTRY(hammer_record)      target_entry;
 345         hammer_inode_state_t            flush_state;
 346         int                             flush_group;
 347         hammer_record_type_t            type;
 348         struct hammer_lock              lock;
 349         struct hammer_reserve           *resv;
 350         struct hammer_inode             *ip;
 351         struct hammer_inode             *target_ip;
 352         struct hammer_btree_leaf_elm    leaf;
 353         union hammer_data_ondisk        *data;
 354         int                             flags;
 355 };
 356
 357 typedef struct hammer_record *hammer_record_t;
 358
 359 /*
 360  * Record flags.  Note that FE can only be set by the frontend if the
 361  * record has not been interlocked by the backend w/ BE.
 362  */
 363 #define HAMMER_RECF_ALLOCDATA           0x0001
 364 #define HAMMER_RECF_ONRBTREE            0x0002
 365 #define HAMMER_RECF_DELETED_FE          0x0004  /* deleted (frontend) */
 366 #define HAMMER_RECF_DELETED_BE          0x0008  /* deleted (backend) */
 367 #define HAMMER_RECF_UNUSED0010          0x0010
 368 #define HAMMER_RECF_INTERLOCK_BE        0x0020  /* backend interlock */
 369 #define HAMMER_RECF_WANTED              0x0040  /* wanted by the frontend */
 370 #define HAMMER_RECF_CONVERT_DELETE      0x0100 /* special case */
 371
 372 /*
 373  * In-memory structures representing on-disk structures.
 374  */
 375 struct hammer_volume;
 376 struct hammer_buffer;
 377 struct hammer_node;
 378 struct hammer_undo;
 379 struct hammer_reserve;
 380
 381 RB_HEAD(hammer_vol_rb_tree, hammer_volume);
 382 RB_HEAD(hammer_buf_rb_tree, hammer_buffer);
 383 RB_HEAD(hammer_nod_rb_tree, hammer_node);
 384 RB_HEAD(hammer_und_rb_tree, hammer_undo);
 385 RB_HEAD(hammer_res_rb_tree, hammer_reserve);
 386
 387 RB_PROTOTYPE2(hammer_vol_rb_tree, hammer_volume, rb_node,
 388               hammer_vol_rb_compare, int32_t);
 389 RB_PROTOTYPE2(hammer_buf_rb_tree, hammer_buffer, rb_node,
 390               hammer_buf_rb_compare, hammer_off_t);
 391 RB_PROTOTYPE2(hammer_nod_rb_tree, hammer_node, rb_node,
 392               hammer_nod_rb_compare, hammer_off_t);
 393 RB_PROTOTYPE2(hammer_und_rb_tree, hammer_undo, rb_node,
 394               hammer_und_rb_compare, hammer_off_t);
 395 RB_PROTOTYPE2(hammer_res_rb_tree, hammer_reserve, rb_node,
 396               hammer_res_rb_compare, hammer_off_t);
 397
 398 /*
 399  * IO management - embedded at the head of various in-memory structures
 400  *
 401  * VOLUME       - hammer_volume containing meta-data
 402  * META_BUFFER  - hammer_buffer containing meta-data
 403  * DATA_BUFFER  - hammer_buffer containing pure-data
 404  *
 405  * Dirty volume headers and dirty meta-data buffers are locked until the
 406  * flusher can sequence them out.  Dirty pure-data buffers can be written.
 407  * Clean buffers can be passively released.
 408  */
 409 typedef enum hammer_io_type {
 410         HAMMER_STRUCTURE_VOLUME,
 411         HAMMER_STRUCTURE_META_BUFFER,
 412         HAMMER_STRUCTURE_UNDO_BUFFER,
 413         HAMMER_STRUCTURE_DATA_BUFFER
 414 } hammer_io_type_t;
 415
 416 union hammer_io_structure;
 417 struct hammer_io;
 418
 419 struct worklist {
 420         LIST_ENTRY(worklist) node;
 421 };
 422
 423 TAILQ_HEAD(hammer_io_list, hammer_io);
 424 typedef struct hammer_io_list *hammer_io_list_t;
 425
 426 struct hammer_io {
 427         struct worklist         worklist;
 428         struct hammer_lock      lock;
 429         enum hammer_io_type     type;
 430         struct hammer_mount     *hmp;
 431         TAILQ_ENTRY(hammer_io)  mod_entry; /* list entry if modified */
 432         hammer_io_list_t        mod_list;
 433         struct buf              *bp;
 434         int64_t                 offset;    /* zone-2 offset */
 435         int                     bytes;     /* buffer cache buffer size */
 436         int                     loading;   /* loading/unloading interlock */
 437         int                     modify_refs;
 438
 439         u_int           modified : 1;   /* bp's data was modified */
 440         u_int           released : 1;   /* bp released (w/ B_LOCKED set) */
 441         u_int           running : 1;    /* bp write IO in progress */
 442         u_int           waiting : 1;    /* someone is waiting on us */
 443         u_int           validated : 1;  /* ondisk has been validated */
 444         u_int           waitdep : 1;    /* flush waits for dependancies */
 445         u_int           recovered : 1;  /* has recovery ref */
 446         u_int           waitmod : 1;    /* waiting for modify_refs */
 447         u_int           reclaim : 1;    /* reclaim requested */
 448         u_int           gencrc : 1;     /* crc needs to be generated */
 449 };
 450
 451 typedef struct hammer_io *hammer_io_t;
 452
 453 #define HAMMER_CLUSTER_SIZE     (64 * 1024)
 454 #if HAMMER_CLUSTER_SIZE > MAXBSIZE
 455 #undef  HAMMER_CLUSTER_SIZE
 456 #define HAMMER_CLUSTER_SIZE     MAXBSIZE
 457 #endif
 458 #define HAMMER_CLUSTER_BUFS     (HAMMER_CLUSTER_SIZE / HAMMER_BUFSIZE)
 459
 460 /*
 461  * In-memory volume representing on-disk buffer
 462  */
 463 struct hammer_volume {
 464         struct hammer_io io;
 465         RB_ENTRY(hammer_volume) rb_node;
 466         struct hammer_volume_ondisk *ondisk;
 467         int32_t vol_no;
 468         int64_t nblocks;        /* note: special calculation for statfs */
 469         int64_t buffer_base;    /* base offset of buffer 0 */
 470         hammer_off_t maxbuf_off; /* Maximum buffer offset (zone-2) */
 471         hammer_off_t maxraw_off; /* Maximum raw offset for device */
 472         char    *vol_name;
 473         struct vnode *devvp;
 474         int     vol_flags;
 475 };
 476
 477 typedef struct hammer_volume *hammer_volume_t;
 478
 479 /*
 480  * In-memory buffer (other then volume, super-cluster, or cluster),
 481  * representing an on-disk buffer.
 482  */
 483 struct hammer_buffer {
 484         struct hammer_io io;
 485         RB_ENTRY(hammer_buffer) rb_node;
 486         void *ondisk;
 487         struct hammer_volume *volume;
 488         hammer_off_t zoneX_offset;
 489         hammer_off_t zone2_offset;
 490         struct hammer_reserve *resv;
 491         struct hammer_node_list clist;
 492 };
 493
 494 typedef struct hammer_buffer *hammer_buffer_t;
 495
 496 /*
 497  * In-memory B-Tree node, representing an on-disk B-Tree node.
 498  *
 499  * This is a hang-on structure which is backed by a hammer_buffer,
 500  * indexed by a hammer_cluster, and used for fine-grained locking of
 501  * B-Tree nodes in order to properly control lock ordering.  A hammer_buffer
 502  * can contain multiple nodes representing wildly disassociated portions
 503  * of the B-Tree so locking cannot be done on a buffer-by-buffer basis.
 504  *
 505  * This structure uses a cluster-relative index to reduce the number
 506  * of layers required to access it, and also because all on-disk B-Tree
 507  * references are cluster-relative offsets.
 508  */
 509 struct hammer_node {
 510         struct hammer_lock      lock;           /* node-by-node lock */
 511         TAILQ_ENTRY(hammer_node) entry;         /* per-buffer linkage */
 512         RB_ENTRY(hammer_node)   rb_node;        /* per-cluster linkage */
 513         hammer_off_t            node_offset;    /* full offset spec */
 514         struct hammer_mount     *hmp;
 515         struct hammer_buffer    *buffer;        /* backing buffer */
 516         hammer_node_ondisk_t    ondisk;         /* ptr to on-disk structure */
 517         struct hammer_node_cache_list cache_list; /* passive caches */
 518         int                     flags;
 519         int                     loading;        /* load interlock */
 520 };
 521
 522 #define HAMMER_NODE_DELETED     0x0001
 523 #define HAMMER_NODE_FLUSH       0x0002
 524 #define HAMMER_NODE_CRCGOOD     0x0004
 525 #define HAMMER_NODE_NEEDSCRC    0x0008
 526 #define HAMMER_NODE_NEEDSMIRROR 0x0010
 527
 528 typedef struct hammer_node      *hammer_node_t;
 529
 530 /*
 531  * List of locked nodes.
 532  */
 533 struct hammer_node_locklist {
 534         struct hammer_node_locklist *next;
 535         hammer_node_t   node;
 536 };
 537
 538 typedef struct hammer_node_locklist *hammer_node_locklist_t;
 539
 540
 541 /*
 542  * Common I/O management structure - embedded in in-memory structures
 543  * which are backed by filesystem buffers.
 544  */
 545 union hammer_io_structure {
 546         struct hammer_io        io;
 547         struct hammer_volume    volume;
 548         struct hammer_buffer    buffer;
 549 };
 550
 551 typedef union hammer_io_structure *hammer_io_structure_t;
 552
 553 /*
 554  * The reserve structure prevents the blockmap from allocating
 555  * out of a reserved bigblock.  Such reservations are used by
 556  * the direct-write mechanism.
 557  *
 558  * The structure is also used to hold off on reallocations of
 559  * big blocks from the freemap until flush dependancies have
 560  * been dealt with.
 561  */
 562 struct hammer_reserve {
 563         RB_ENTRY(hammer_reserve) rb_node;
 564         TAILQ_ENTRY(hammer_reserve) delay_entry;
 565         int             flush_group;
 566         int             flags;
 567         int             refs;
 568         int             zone;
 569         hammer_off_t    zone_offset;
 570 };
 571
 572 typedef struct hammer_reserve *hammer_reserve_t;
 573
 574 #define HAMMER_RESF_ONDELAY     0x0001
 575
 576 #include "hammer_cursor.h"
 577
 578 /*
 579  * The undo structure tracks recent undos to avoid laying down duplicate
 580  * undos within a flush group, saving us a significant amount of overhead.
 581  *
 582  * This is strictly a heuristic.
 583  */
 584 #define HAMMER_MAX_UNDOS        1024
 585 #define HAMMER_MAX_FLUSHERS     4
 586
 587 struct hammer_undo {
 588         RB_ENTRY(hammer_undo)   rb_node;
 589         TAILQ_ENTRY(hammer_undo) lru_entry;
 590         hammer_off_t            offset;
 591         int                     bytes;
 592 };
 593
 594 typedef struct hammer_undo *hammer_undo_t;
 595
 596 struct hammer_flusher_info;
 597
 598 struct hammer_flusher {
 599         int             signal;         /* flusher thread sequencer */
 600         int             act;            /* currently active flush group */
 601         int             done;           /* set to act when complete */
 602         int             next;           /* next flush group */
 603         int             group_lock;     /* lock sequencing of the next flush */
 604         int             exiting;        /* request master exit */
 605         int             count;          /* number of slave flushers */
 606         int             running;        /* number of slave flushers running */
 607         thread_t        td;             /* master flusher thread */
 608         hammer_tid_t    tid;            /* last flushed transaction id */
 609         int             finalize_want;          /* serialize finalization */
 610         struct hammer_lock finalize_lock;       /* serialize finalization */
 611         struct hammer_transaction trans;        /* shared transaction */
 612         struct hammer_flusher_info *info[HAMMER_MAX_FLUSHERS];
 613 };
 614
 615 /*
 616  * Internal hammer mount data structure
 617  */
 618 struct hammer_mount {
 619         struct mount *mp;
 620         /*struct vnode *rootvp;*/
 621         struct hammer_ino_rb_tree rb_inos_root;
 622         struct hammer_vol_rb_tree rb_vols_root;
 623         struct hammer_nod_rb_tree rb_nods_root;
 624         struct hammer_und_rb_tree rb_undo_root;
 625         struct hammer_res_rb_tree rb_resv_root;
 626         struct hammer_buf_rb_tree rb_bufs_root;
 627         struct hammer_volume *rootvol;
 628         struct hammer_base_elm root_btree_beg;
 629         struct hammer_base_elm root_btree_end;
 630         int     flags;
 631         int     hflags;
 632         int     ronly;
 633         int     nvolumes;
 634         int     volume_iterator;
 635         int     masterid;       /* -1 or 0-15 - clustering and mirroring */
 636         int     rsv_inodes;     /* reserved space due to dirty inodes */
 637         int     rsv_databufs;   /* reserved space due to dirty buffers */
 638         int     rsv_databytes;  /* reserved space due to record data */
 639         int     rsv_recs;       /* reserved space due to dirty records */
 640         int     last_newrecords;
 641         int     count_newrecords;
 642
 643         int     inode_reclaims; /* inodes pending reclaim by flusher */
 644         int     count_inodes;   /* total number of inodes */
 645         int     count_iqueued;  /* inodes queued to flusher */
 646
 647         struct hammer_flusher flusher;
 648
 649         u_int   check_interrupt;
 650         uuid_t  fsid;
 651         udev_t  fsid_udev;
 652         struct hammer_io_list volu_list;        /* dirty undo buffers */
 653         struct hammer_io_list undo_list;        /* dirty undo buffers */
 654         struct hammer_io_list data_list;        /* dirty data buffers */
 655         struct hammer_io_list alt_data_list;    /* dirty data buffers */
 656         struct hammer_io_list meta_list;        /* dirty meta bufs    */
 657         struct hammer_io_list lose_list;        /* loose buffers      */
 658         int     locked_dirty_count;             /* meta/volu count    */
 659         int     io_running_count;
 660         int     objid_cache_count;
 661         hammer_tid_t    asof;                   /* snapshot mount */
 662         hammer_off_t    next_tid;
 663         int64_t copy_stat_freebigblocks;        /* number of free bigblocks */
 664
 665         u_int32_t namekey_iterator;
 666         struct netexport export;
 667         struct hammer_lock sync_lock;
 668         struct hammer_lock free_lock;
 669         struct hammer_lock undo_lock;
 670         struct hammer_lock blkmap_lock;
 671         struct hammer_blockmap  blockmap[HAMMER_MAX_ZONES];
 672         struct hammer_undo      undos[HAMMER_MAX_UNDOS];
 673         int                     undo_alloc;
 674         TAILQ_HEAD(, hammer_undo)  undo_lru_list;
 675         TAILQ_HEAD(, hammer_inode) flush_list;
 676         TAILQ_HEAD(, hammer_reserve) delay_list;
 677         TAILQ_HEAD(, hammer_objid_cache) objid_cache_list;
 678         TAILQ_HEAD(, hammer_reclaim) reclaim_list;
 679 };
 680
 681 typedef struct hammer_mount     *hammer_mount_t;
 682
 683 #define HAMMER_MOUNT_UNUSED0001 0x0001
 684
 685 struct hammer_sync_info {
 686         int error;
 687         int waitfor;
 688 };
 689
 690 #endif
 691
 692 #if defined(_KERNEL)
 693
 694 extern struct vop_ops hammer_vnode_vops;
 695 extern struct vop_ops hammer_spec_vops;
 696 extern struct vop_ops hammer_fifo_vops;
 697 extern struct bio_ops hammer_bioops;
 698
 699 extern int hammer_debug_io;
 700 extern int hammer_debug_general;
 701 extern int hammer_debug_debug;
 702 extern int hammer_debug_inode;
 703 extern int hammer_debug_locks;
 704 extern int hammer_debug_btree;
 705 extern int hammer_debug_tid;
 706 extern int hammer_debug_recover;
 707 extern int hammer_debug_recover_faults;
 708 extern int hammer_debug_cluster_enable;
 709 extern int hammer_count_inodes;
 710 extern int hammer_count_iqueued;
 711 extern int hammer_count_reclaiming;
 712 extern int hammer_count_records;
 713 extern int hammer_count_record_datas;
 714 extern int hammer_count_volumes;
 715 extern int hammer_count_buffers;
 716 extern int hammer_count_nodes;
 717 extern int64_t hammer_stats_btree_lookups;
 718 extern int64_t hammer_stats_btree_searches;
 719 extern int64_t hammer_stats_btree_inserts;
 720 extern int64_t hammer_stats_btree_deletes;
 721 extern int64_t hammer_stats_btree_elements;
 722 extern int64_t hammer_stats_btree_splits;
 723 extern int64_t hammer_stats_btree_iterations;
 724 extern int64_t hammer_stats_record_iterations;
 725 extern int hammer_count_dirtybufs;
 726 extern int hammer_count_refedbufs;
 727 extern int hammer_count_reservations;
 728 extern int hammer_count_io_running_read;
 729 extern int hammer_count_io_running_write;
 730 extern int hammer_count_io_locked;
 731 extern int hammer_limit_dirtybufs;
 732 extern int hammer_limit_iqueued;
 733 extern int hammer_limit_recs;
 734 extern int hammer_bio_count;
 735 extern int hammer_verify_zone;
 736 extern int hammer_write_mode;
 737 extern int64_t hammer_contention_count;
 738
 739 int     hammer_vop_inactive(struct vop_inactive_args *);
 740 int     hammer_vop_reclaim(struct vop_reclaim_args *);
 741 int     hammer_get_vnode(struct hammer_inode *ip, struct vnode **vpp);
 742 struct hammer_inode *hammer_get_inode(hammer_transaction_t trans,
 743                         hammer_inode_t dip, u_int64_t obj_id,
 744                         hammer_tid_t asof, u_int32_t localization,
 745                         int flags, int *errorp);
 746 void    hammer_scan_inode_snapshots(hammer_mount_t hmp,
 747                         hammer_inode_info_t iinfo,
 748                         int (*callback)(hammer_inode_t ip, void *data),
 749                         void *data);
 750 void    hammer_put_inode(struct hammer_inode *ip);
 751 void    hammer_put_inode_ref(struct hammer_inode *ip);
 752 void    hammer_inode_waitreclaims(hammer_mount_t hmp);
 753
 754 int     hammer_unload_volume(hammer_volume_t volume, void *data __unused);
 755 int     hammer_adjust_volume_mode(hammer_volume_t volume, void *data __unused);
 756
 757 int     hammer_unload_buffer(hammer_buffer_t buffer, void *data __unused);
 758 int     hammer_install_volume(hammer_mount_t hmp, const char *volname);
 759
 760 int     hammer_ip_lookup(hammer_cursor_t cursor);
 761 int     hammer_ip_first(hammer_cursor_t cursor);
 762 int     hammer_ip_next(hammer_cursor_t cursor);
 763 int     hammer_ip_resolve_data(hammer_cursor_t cursor);
 764 int     hammer_ip_delete_record(hammer_cursor_t cursor, hammer_inode_t ip,
 765                         hammer_tid_t tid);
 766 int     hammer_delete_at_cursor(hammer_cursor_t cursor, int64_t *stat_bytes);
 767 int     hammer_ip_check_directory_empty(hammer_transaction_t trans,
 768                         hammer_inode_t ip);
 769 int     hammer_sync_hmp(hammer_mount_t hmp, int waitfor);
 770 int     hammer_queue_inodes_flusher(hammer_mount_t hmp, int waitfor);
 771
 772
 773 hammer_record_t
 774         hammer_alloc_mem_record(hammer_inode_t ip, int data_len);
 775 void    hammer_flush_record_done(hammer_record_t record, int error);
 776 void    hammer_wait_mem_record_ident(hammer_record_t record, const char *ident);
 777 void    hammer_rel_mem_record(hammer_record_t record);
 778
 779 int     hammer_cursor_up(hammer_cursor_t cursor);
 780 int     hammer_cursor_up_locked(hammer_cursor_t cursor);
 781 int     hammer_cursor_down(hammer_cursor_t cursor);
 782 int     hammer_cursor_upgrade(hammer_cursor_t cursor);
 783 int     hammer_cursor_upgrade_node(hammer_cursor_t cursor);
 784 void    hammer_cursor_downgrade(hammer_cursor_t cursor);
 785 int     hammer_cursor_seek(hammer_cursor_t cursor, hammer_node_t node,
 786                         int index);
 787 void    hammer_lock_ex_ident(struct hammer_lock *lock, const char *ident);
 788 int     hammer_lock_ex_try(struct hammer_lock *lock);
 789 void    hammer_lock_sh(struct hammer_lock *lock);
 790 void    hammer_lock_sh_lowpri(struct hammer_lock *lock);
 791 int     hammer_lock_sh_try(struct hammer_lock *lock);
 792 int     hammer_lock_upgrade(struct hammer_lock *lock);
 793 void    hammer_lock_downgrade(struct hammer_lock *lock);
 794 void    hammer_unlock(struct hammer_lock *lock);
 795 void    hammer_ref(struct hammer_lock *lock);
 796 void    hammer_unref(struct hammer_lock *lock);
 797
 798 void    hammer_sync_lock_ex(hammer_transaction_t trans);
 799 void    hammer_sync_lock_sh(hammer_transaction_t trans);
 800 int     hammer_sync_lock_sh_try(hammer_transaction_t trans);
 801 void    hammer_sync_unlock(hammer_transaction_t trans);
 802
 803 u_int32_t hammer_to_unix_xid(uuid_t *uuid);
 804 void hammer_guid_to_uuid(uuid_t *uuid, u_int32_t guid);
 805 void    hammer_time_to_timespec(u_int64_t xtime, struct timespec *ts);
 806 u_int64_t hammer_timespec_to_time(struct timespec *ts);
 807 hammer_tid_t hammer_str_to_tid(const char *str);
 808 hammer_tid_t hammer_alloc_objid(hammer_mount_t hmp, hammer_inode_t dip);
 809 void hammer_clear_objid(hammer_inode_t dip);
 810 void hammer_destroy_objid_cache(hammer_mount_t hmp);
 811
 812 int hammer_enter_undo_history(hammer_mount_t hmp, hammer_off_t offset,
 813                               int bytes);
 814 void hammer_clear_undo_history(hammer_mount_t hmp);
 815 enum vtype hammer_get_vnode_type(u_int8_t obj_type);
 816 int hammer_get_dtype(u_int8_t obj_type);
 817 u_int8_t hammer_get_obj_type(enum vtype vtype);
 818 int64_t hammer_directory_namekey(const void *name, int len);
 819 int     hammer_nohistory(hammer_inode_t ip);
 820
 821 int     hammer_init_cursor(hammer_transaction_t trans, hammer_cursor_t cursor,
 822                            hammer_node_cache_t cache, hammer_inode_t ip);
 823 int     hammer_reinit_cursor(hammer_cursor_t cursor);
 824 void    hammer_normalize_cursor(hammer_cursor_t cursor);
 825 void    hammer_done_cursor(hammer_cursor_t cursor);
 826 void    hammer_mem_done(hammer_cursor_t cursor);
 827
 828 int     hammer_btree_lookup(hammer_cursor_t cursor);
 829 int     hammer_btree_first(hammer_cursor_t cursor);
 830 int     hammer_btree_last(hammer_cursor_t cursor);
 831 int     hammer_btree_extract(hammer_cursor_t cursor, int flags);
 832 int     hammer_btree_iterate(hammer_cursor_t cursor);
 833 int     hammer_btree_iterate_reverse(hammer_cursor_t cursor);
 834 int     hammer_btree_insert(hammer_cursor_t cursor,
 835                             hammer_btree_leaf_elm_t elm);
 836 int     hammer_btree_delete(hammer_cursor_t cursor);
 837 int     hammer_btree_mirror_propagate(hammer_transaction_t trans,
 838                             hammer_node_t node, int index,
 839                             hammer_tid_t mirror_tid);
 840
 841 int     hammer_btree_cmp(hammer_base_elm_t key1, hammer_base_elm_t key2);
 842 int     hammer_btree_chkts(hammer_tid_t ts, hammer_base_elm_t key);
 843 int     hammer_btree_correct_rhb(hammer_cursor_t cursor, hammer_tid_t tid);
 844 int     hammer_btree_correct_lhb(hammer_cursor_t cursor, hammer_tid_t tid);
 845
 846 int     btree_set_parent(hammer_transaction_t trans, hammer_node_t node,
 847                         hammer_btree_elm_t elm);
 848 int     hammer_btree_lock_children(hammer_cursor_t cursor,
 849                         struct hammer_node_locklist **locklistp);
 850 void    hammer_btree_unlock_children(struct hammer_node_locklist **locklistp);
 851 int     hammer_btree_search_node(hammer_base_elm_t elm, hammer_node_ondisk_t node);
 852 hammer_node_t hammer_btree_get_parent(hammer_node_t node, int *parent_indexp,
 853                         int *errorp, int try_exclusive);
 854
 855 void    hammer_print_btree_node(hammer_node_ondisk_t ondisk);
 856 void    hammer_print_btree_elm(hammer_btree_elm_t elm, u_int8_t type, int i);
 857
 858 void    *hammer_bread(struct hammer_mount *hmp, hammer_off_t off,
 859                         int *errorp, struct hammer_buffer **bufferp);
 860 void    *hammer_bnew(struct hammer_mount *hmp, hammer_off_t off,
 861                         int *errorp, struct hammer_buffer **bufferp);
 862 void    *hammer_bread_ext(struct hammer_mount *hmp, hammer_off_t off, int bytes,
 863                         int *errorp, struct hammer_buffer **bufferp);
 864 void    *hammer_bnew_ext(struct hammer_mount *hmp, hammer_off_t off, int bytes,
 865                         int *errorp, struct hammer_buffer **bufferp);
 866
 867 hammer_volume_t hammer_get_root_volume(hammer_mount_t hmp, int *errorp);
 868
 869 hammer_volume_t hammer_get_volume(hammer_mount_t hmp,
 870                         int32_t vol_no, int *errorp);
 871 hammer_buffer_t hammer_get_buffer(hammer_mount_t hmp, hammer_off_t buf_offset,
 872                         int bytes, int isnew, int *errorp);
 873 void            hammer_del_buffers(hammer_mount_t hmp, hammer_off_t base_offset,
 874                         hammer_off_t zone2_offset, int bytes);
 875
 876 int             hammer_ref_volume(hammer_volume_t volume);
 877 int             hammer_ref_buffer(hammer_buffer_t buffer);
 878 void            hammer_flush_buffer_nodes(hammer_buffer_t buffer);
 879
 880 void            hammer_rel_volume(hammer_volume_t volume, int flush);
 881 void            hammer_rel_buffer(hammer_buffer_t buffer, int flush);
 882
 883 int             hammer_vfs_export(struct mount *mp, int op,
 884                         const struct export_args *export);
 885 hammer_node_t   hammer_get_node(hammer_mount_t hmp, hammer_off_t node_offset,
 886                         int isnew, int *errorp);
 887 void            hammer_ref_node(hammer_node_t node);
 888 hammer_node_t   hammer_ref_node_safe(struct hammer_mount *hmp,
 889                         hammer_node_cache_t cache, int *errorp);
 890 void            hammer_rel_node(hammer_node_t node);
 891 void            hammer_delete_node(hammer_transaction_t trans,
 892                         hammer_node_t node);
 893 void            hammer_cache_node(hammer_node_cache_t cache,
 894                         hammer_node_t node);
 895 void            hammer_uncache_node(hammer_node_cache_t cache);
 896 void            hammer_flush_node(hammer_node_t node);
 897
 898 void hammer_dup_buffer(struct hammer_buffer **bufferp,
 899                         struct hammer_buffer *buffer);
 900 hammer_node_t hammer_alloc_btree(hammer_transaction_t trans, int *errorp);
 901 void *hammer_alloc_data(hammer_transaction_t trans, int32_t data_len,
 902                         u_int16_t rec_type, hammer_off_t *data_offsetp,
 903                         struct hammer_buffer **data_bufferp, int *errorp);
 904
 905 int hammer_generate_undo(hammer_transaction_t trans, hammer_io_t io,
 906                         hammer_off_t zone1_offset, void *base, int len);
 907
 908 void hammer_put_volume(struct hammer_volume *volume, int flush);
 909 void hammer_put_buffer(struct hammer_buffer *buffer, int flush);
 910
 911 hammer_off_t hammer_freemap_alloc(hammer_transaction_t trans,
 912                         hammer_off_t owner, int *errorp);
 913 void hammer_freemap_free(hammer_transaction_t trans, hammer_off_t phys_offset,
 914                         hammer_off_t owner, int *errorp);
 915 int hammer_checkspace(hammer_mount_t hmp);
 916 hammer_off_t hammer_blockmap_alloc(hammer_transaction_t trans, int zone,
 917                         int bytes, int *errorp);
 918 hammer_reserve_t hammer_blockmap_reserve(hammer_mount_t hmp, int zone,
 919                         int bytes, hammer_off_t *zone_offp, int *errorp);
 920 void hammer_blockmap_reserve_complete(hammer_mount_t hmp,
 921                         hammer_reserve_t resv);
 922 void hammer_reserve_setdelay(hammer_mount_t hmp, hammer_reserve_t resv,
 923                         hammer_off_t zone2_offset);
 924 void hammer_reserve_clrdelay(hammer_mount_t hmp, hammer_reserve_t resv);
 925 void hammer_blockmap_free(hammer_transaction_t trans,
 926                         hammer_off_t bmap_off, int bytes);
 927 void hammer_blockmap_finalize(hammer_transaction_t trans,
 928                         hammer_off_t bmap_off, int bytes);
 929 int hammer_blockmap_getfree(hammer_mount_t hmp, hammer_off_t bmap_off,
 930                         int *curp, int *errorp);
 931 hammer_off_t hammer_blockmap_lookup(hammer_mount_t hmp, hammer_off_t bmap_off,
 932                         int *errorp);
 933 hammer_off_t hammer_undo_lookup(hammer_mount_t hmp, hammer_off_t bmap_off,
 934                         int *errorp);
 935 int64_t hammer_undo_used(hammer_mount_t hmp);
 936 int64_t hammer_undo_space(hammer_mount_t hmp);
 937 int64_t hammer_undo_max(hammer_mount_t hmp);
 938
 939 void hammer_start_transaction(struct hammer_transaction *trans,
 940                               struct hammer_mount *hmp);
 941 void hammer_simple_transaction(struct hammer_transaction *trans,
 942                               struct hammer_mount *hmp);
 943 void hammer_start_transaction_fls(struct hammer_transaction *trans,
 944                                   struct hammer_mount *hmp);
 945 void hammer_done_transaction(struct hammer_transaction *trans);
 946
 947 void hammer_modify_inode(hammer_inode_t ip, int flags);
 948 void hammer_flush_inode(hammer_inode_t ip, int flags);
 949 void hammer_flush_inode_done(hammer_inode_t ip);
 950 void hammer_wait_inode(hammer_inode_t ip);
 951
 952 int  hammer_create_inode(struct hammer_transaction *trans, struct vattr *vap,
 953                         struct ucred *cred, struct hammer_inode *dip,
 954                         int pseudofs, struct hammer_inode **ipp);
 955 void hammer_rel_inode(hammer_inode_t ip, int flush);
 956 int hammer_reload_inode(hammer_inode_t ip, void *arg __unused);
 957 int hammer_ino_rb_compare(hammer_inode_t ip1, hammer_inode_t ip2);
 958
 959 int hammer_sync_inode(hammer_inode_t ip);
 960 void hammer_test_inode(hammer_inode_t ip);
 961 void hammer_inode_unloadable_check(hammer_inode_t ip, int getvp);
 962
 963 int  hammer_ip_add_directory(struct hammer_transaction *trans,
 964                         hammer_inode_t dip, const char *name, int bytes,
 965                         hammer_inode_t nip);
 966 int  hammer_ip_del_directory(struct hammer_transaction *trans,
 967                         hammer_cursor_t cursor, hammer_inode_t dip,
 968                         hammer_inode_t ip);
 969 hammer_record_t hammer_ip_add_bulk(hammer_inode_t ip, off_t file_offset,
 970                         void *data, int bytes, int *errorp);
 971 int  hammer_ip_frontend_trunc(struct hammer_inode *ip, off_t file_size);
 972 int  hammer_ip_add_record(struct hammer_transaction *trans,
 973                         hammer_record_t record);
 974 int  hammer_ip_delete_range(hammer_cursor_t cursor, hammer_inode_t ip,
 975                         int64_t ran_beg, int64_t ran_end, int truncating);
 976 int  hammer_ip_delete_range_all(hammer_cursor_t cursor, hammer_inode_t ip,
 977                         int *countp);
 978 int  hammer_ip_sync_data(hammer_cursor_t cursor, hammer_inode_t ip,
 979                         int64_t offset, void *data, int bytes);
 980 int  hammer_ip_sync_record(hammer_transaction_t trans, hammer_record_t rec);
 981 int  hammer_ip_sync_record_cursor(hammer_cursor_t cursor, hammer_record_t rec);
 982
 983 int hammer_ioctl(hammer_inode_t ip, u_long com, caddr_t data, int fflag,
 984                         struct ucred *cred);
 985
 986 void hammer_io_init(hammer_io_t io, hammer_mount_t hmp,
 987                         enum hammer_io_type type);
 988 int hammer_io_read(struct vnode *devvp, struct hammer_io *io,
 989                         hammer_off_t limit);
 990 int hammer_io_new(struct vnode *devvp, struct hammer_io *io);
 991 void hammer_io_inval(hammer_volume_t volume, hammer_off_t zone2_offset);
 992 void hammer_io_release(struct hammer_io *io, int flush);
 993 void hammer_io_flush(struct hammer_io *io);
 994 void hammer_io_waitdep(struct hammer_io *io);
 995 void hammer_io_wait_all(hammer_mount_t hmp, const char *ident);
 996 int hammer_io_direct_read(hammer_mount_t hmp, struct bio *bio);
 997 int hammer_io_direct_write(hammer_mount_t hmp, hammer_btree_leaf_elm_t leaf,
 998                           struct bio *bio);
 999 void hammer_io_direct_uncache(hammer_mount_t hmp, hammer_btree_leaf_elm_t leaf);
1000 void hammer_io_write_interlock(hammer_io_t io);
1001 void hammer_io_done_interlock(hammer_io_t io);
1002 void hammer_io_clear_modify(struct hammer_io *io, int inval);
1003 void hammer_io_clear_modlist(struct hammer_io *io);
1004 void hammer_modify_volume(hammer_transaction_t trans, hammer_volume_t volume,
1005                         void *base, int len);
1006 void hammer_modify_buffer(hammer_transaction_t trans, hammer_buffer_t buffer,
1007                         void *base, int len);
1008 void hammer_modify_volume_done(hammer_volume_t volume);
1009 void hammer_modify_buffer_done(hammer_buffer_t buffer);
1010
1011 int hammer_ioc_reblock(hammer_transaction_t trans, hammer_inode_t ip,
1012                         struct hammer_ioc_reblock *reblock);
1013 int hammer_ioc_prune(hammer_transaction_t trans, hammer_inode_t ip,
1014                         struct hammer_ioc_prune *prune);
1015 int hammer_ioc_mirror_read(hammer_transaction_t trans, hammer_inode_t ip,
1016                         struct hammer_ioc_mirror_rw *mirror);
1017 int hammer_ioc_mirror_write(hammer_transaction_t trans, hammer_inode_t ip,
1018                         struct hammer_ioc_mirror_rw *mirror);
1019
1020 int hammer_signal_check(hammer_mount_t hmp);
1021
1022 void hammer_flusher_create(hammer_mount_t hmp);
1023 void hammer_flusher_destroy(hammer_mount_t hmp);
1024 void hammer_flusher_sync(hammer_mount_t hmp);
1025 void hammer_flusher_async(hammer_mount_t hmp);
1026
1027 int hammer_recover(hammer_mount_t hmp, hammer_volume_t rootvol);
1028 void hammer_recover_flush_buffers(hammer_mount_t hmp,
1029                         hammer_volume_t root_volume);
1030
1031 void hammer_crc_set_blockmap(hammer_blockmap_t blockmap);
1032 void hammer_crc_set_volume(hammer_volume_ondisk_t ondisk);
1033 void hammer_crc_set_leaf(void *data, hammer_btree_leaf_elm_t leaf);
1034
1035 int hammer_crc_test_blockmap(hammer_blockmap_t blockmap);
1036 int hammer_crc_test_volume(hammer_volume_ondisk_t ondisk);
1037 int hammer_crc_test_btree(hammer_node_ondisk_t ondisk);
1038 int hammer_crc_test_leaf(void *data, hammer_btree_leaf_elm_t leaf);
1039 void hkprintf(const char *ctl, ...);
1040
1041 int hammer_blocksize(int64_t file_offset);
1042 int64_t hammer_blockdemarc(int64_t file_offset1, int64_t file_offset2);
1043
1044 #endif
1045
1046 static __inline void
1047 hammer_wait_mem_record(hammer_record_t record)
1048 {
1049         hammer_wait_mem_record_ident(record, "hmmwai");
1050 }
1051
1052 static __inline void
1053 hammer_lock_ex(struct hammer_lock *lock)
1054 {
1055         hammer_lock_ex_ident(lock, "hmrlck");
1056 }
1057
1058 /*
1059  * Indicate that a B-Tree node is being modified.
1060  */
1061 static __inline void
1062 hammer_modify_node_noundo(hammer_transaction_t trans, hammer_node_t node)
1063 {
1064         hammer_modify_buffer(trans, node->buffer, NULL, 0);
1065 }
1066
1067 static __inline void
1068 hammer_modify_node_all(hammer_transaction_t trans, struct hammer_node *node)
1069 {
1070         hammer_modify_buffer(trans, node->buffer,
1071                              node->ondisk, sizeof(*node->ondisk));
1072 }
1073
1074 static __inline void
1075 hammer_modify_node(hammer_transaction_t trans, hammer_node_t node,
1076                    void *base, int len)
1077 {
1078         hammer_crc_t *crcptr;
1079
1080         KKASSERT((char *)base >= (char *)node->ondisk &&
1081                  (char *)base + len <=
1082                     (char *)node->ondisk + sizeof(*node->ondisk));
1083         hammer_modify_buffer(trans, node->buffer, base, len);
1084         crcptr = &node->ondisk->crc;
1085         hammer_modify_buffer(trans, node->buffer, crcptr, sizeof(hammer_crc_t));
1086         --node->buffer->io.modify_refs; /* only want one ref */
1087 }
1088
1089 /*
1090  * Indicate that the specified modifications have been completed.
1091  *
1092  * Do not try to generate the crc here, it's very expensive to do and a
1093  * sequence of insertions or deletions can result in many calls to this
1094  * function on the same node.
1095  */
1096 static __inline void
1097 hammer_modify_node_done(hammer_node_t node)
1098 {
1099         node->flags |= HAMMER_NODE_CRCGOOD;
1100         if ((node->flags & HAMMER_NODE_NEEDSCRC) == 0) {
1101                 node->flags |= HAMMER_NODE_NEEDSCRC;
1102                 node->buffer->io.gencrc = 1;
1103                 hammer_ref_node(node);
1104         }
1105         hammer_modify_buffer_done(node->buffer);
1106 }
1107
1108 #define hammer_modify_volume_field(trans, vol, field)           \
1109         hammer_modify_volume(trans, vol, &(vol)->ondisk->field, \
1110                              sizeof((vol)->ondisk->field))
1111
1112 #define hammer_modify_node_field(trans, node, field)            \
1113         hammer_modify_node(trans, node, &(node)->ondisk->field, \
1114                              sizeof((node)->ondisk->field))
1115