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 1.107 2008/07/08 04:34:41 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  * Pseudo-filesystem extended data tracking
 170  */
 171 struct hammer_pfs_rb_tree;
 172 struct hammer_pseudofs_inmem;
 173 RB_HEAD(hammer_pfs_rb_tree, hammer_pseudofs_inmem);
 174 RB_PROTOTYPE2(hammer_pfs_rb_tree, hammer_pseudofs_inmem, rb_node,
 175               hammer_pfs_rb_compare, u_int32_t);
 176
 177 struct hammer_pseudofs_inmem {
 178         RB_ENTRY(hammer_pseudofs_inmem) rb_node;
 179         struct hammer_lock      lock;
 180         u_int32_t               localization;
 181         udev_t                  fsid_udev;
 182         struct hammer_pseudofs_data pfsd;
 183 };
 184
 185 typedef struct hammer_pseudofs_inmem *hammer_pseudofs_inmem_t;
 186
 187 /*
 188  * Cache object ids.  A fixed number of objid cache structures are
 189  * created to reserve object id's for newly created files in multiples
 190  * of 100,000, localized to a particular directory, and recycled as
 191  * needed.  This allows parallel create operations in different
 192  * directories to retain fairly localized object ids which in turn
 193  * improves reblocking performance and layout.
 194  */
 195 #define OBJID_CACHE_SIZE        1024
 196 #define OBJID_CACHE_BULK        100000
 197
 198 typedef struct hammer_objid_cache {
 199         TAILQ_ENTRY(hammer_objid_cache) entry;
 200         struct hammer_inode             *dip;
 201         hammer_tid_t                    next_tid;
 202         int                             count;
 203 } *hammer_objid_cache_t;
 204
 205 /*
 206  * Associate an inode with a B-Tree node to cache search start positions
 207  */
 208 typedef struct hammer_node_cache {
 209         TAILQ_ENTRY(hammer_node_cache) entry;
 210         struct hammer_node              *node;
 211         struct hammer_inode             *ip;
 212 } *hammer_node_cache_t;
 213
 214 TAILQ_HEAD(hammer_node_cache_list, hammer_node_cache);
 215
 216 /*
 217  * Structure used to represent an inode in-memory.
 218  *
 219  * The record and data associated with an inode may be out of sync with
 220  * the disk (xDIRTY flags), or not even on the disk at all (ONDISK flag
 221  * clear).
 222  *
 223  * An inode may also hold a cache of unsynchronized records, used for
 224  * database and directories only.  Unsynchronized regular file data is
 225  * stored in the buffer cache.
 226  *
 227  * NOTE: A file which is created and destroyed within the initial
 228  * synchronization period can wind up not doing any disk I/O at all.
 229  *
 230  * Finally, an inode may cache numerous disk-referencing B-Tree cursors.
 231  */
 232 struct hammer_ino_rb_tree;
 233 struct hammer_inode;
 234 RB_HEAD(hammer_ino_rb_tree, hammer_inode);
 235 RB_PROTOTYPEX(hammer_ino_rb_tree, INFO, hammer_inode, rb_node,
 236               hammer_ino_rb_compare, hammer_inode_info_t);
 237
 238 struct hammer_rec_rb_tree;
 239 struct hammer_record;
 240 RB_HEAD(hammer_rec_rb_tree, hammer_record);
 241 RB_PROTOTYPEX(hammer_rec_rb_tree, INFO, hammer_record, rb_node,
 242               hammer_rec_rb_compare, hammer_btree_leaf_elm_t);
 243
 244 TAILQ_HEAD(hammer_node_list, hammer_node);
 245
 246 struct hammer_inode {
 247         RB_ENTRY(hammer_inode)  rb_node;
 248         hammer_inode_state_t    flush_state;
 249         int                     flush_group;
 250         TAILQ_ENTRY(hammer_inode) flush_entry;
 251         struct hammer_record_list target_list;  /* target of dependant recs */
 252         int64_t                 obj_id;         /* (key) object identifier */
 253         hammer_tid_t            obj_asof;       /* (key) snapshot or 0 */
 254         u_int32_t               obj_localization; /* (key) pseudo-fs */
 255         struct hammer_mount     *hmp;
 256         hammer_objid_cache_t    objid_cache;
 257         int                     flags;
 258         int                     error;          /* flush error */
 259         int                     cursor_ip_refs; /* sanity */
 260         int                     rsv_recs;
 261         struct vnode            *vp;
 262         hammer_pseudofs_inmem_t pfsm;
 263         struct lockf            advlock;
 264         struct hammer_lock      lock;           /* sync copy interlock */
 265         off_t                   trunc_off;
 266         struct hammer_btree_leaf_elm ino_leaf;  /* in-memory cache */
 267         struct hammer_inode_data ino_data;      /* in-memory cache */
 268         struct hammer_rec_rb_tree rec_tree;     /* in-memory cache */
 269         struct hammer_node_cache cache[2];      /* search initiate cache */
 270
 271         /*
 272          * When a demark is created to synchronize an inode to
 273          * disk, certain fields are copied so the front-end VOPs
 274          * can continue to run in parallel with the synchronization
 275          * occuring in the background.
 276          */
 277         int             sync_flags;             /* to-sync flags cache */
 278         off_t           sync_trunc_off;         /* to-sync truncation */
 279         off_t           save_trunc_off;         /* write optimization */
 280         struct hammer_btree_leaf_elm sync_ino_leaf; /* to-sync cache */
 281         struct hammer_inode_data sync_ino_data; /* to-sync cache */
 282 };
 283
 284 typedef struct hammer_inode *hammer_inode_t;
 285
 286 #define VTOI(vp)        ((struct hammer_inode *)(vp)->v_data)
 287
 288 #define HAMMER_INODE_DDIRTY     0x0001  /* in-memory ino_data is dirty */
 289                                         /* (not including atime/mtime) */
 290 #define HAMMER_INODE_RSV_INODES 0x0002  /* hmp->rsv_inodes bumped */
 291 #define HAMMER_INODE_PFSD       0x0004  /* obj_asof set based on pfsd */
 292 #define HAMMER_INODE_XDIRTY     0x0008  /* in-memory records */
 293 #define HAMMER_INODE_ONDISK     0x0010  /* inode is on-disk (else not yet) */
 294 #define HAMMER_INODE_FLUSH      0x0020  /* flush on last ref */
 295 #define HAMMER_INODE_DELETED    0x0080  /* inode delete (backend) */
 296 #define HAMMER_INODE_DELONDISK  0x0100  /* delete synchronized to disk */
 297 #define HAMMER_INODE_RO         0x0200  /* read-only (because of as-of) */
 298 #define HAMMER_INODE_VHELD      0x0400  /* vnode held on sync */
 299 #define HAMMER_INODE_DONDISK    0x0800  /* data records may be on disk */
 300 #define HAMMER_INODE_BUFS       0x1000  /* dirty high level bps present */
 301 #define HAMMER_INODE_REFLUSH    0x2000  /* pipelined flush during flush */
 302 #define HAMMER_INODE_RECLAIM    0x4000  /* trying to reclaim */
 303 #define HAMMER_INODE_FLUSHW     0x8000  /* Someone waiting for flush */
 304
 305 #define HAMMER_INODE_TRUNCATED  0x00010000
 306 #define HAMMER_INODE_DELETING   0x00020000 /* inode delete request (frontend)*/
 307 #define HAMMER_INODE_RESIGNAL   0x00040000 /* re-signal on re-flush */
 308 #define HAMMER_INODE_ATIME      0x00100000 /* in-memory atime modified */
 309 #define HAMMER_INODE_MTIME      0x00200000 /* in-memory mtime modified */
 310 #define HAMMER_INODE_WOULDBLOCK 0x00400000 /* re-issue to new flush group */
 311
 312 #define HAMMER_INODE_MODMASK    (HAMMER_INODE_DDIRTY|                       \
 313                                  HAMMER_INODE_XDIRTY|HAMMER_INODE_BUFS|     \
 314                                  HAMMER_INODE_ATIME|HAMMER_INODE_MTIME|     \
 315                                  HAMMER_INODE_TRUNCATED|HAMMER_INODE_DELETING)
 316
 317 #define HAMMER_INODE_MODMASK_NOXDIRTY \
 318                                 (HAMMER_INODE_MODMASK & ~HAMMER_INODE_XDIRTY)
 319
 320 #define HAMMER_FLUSH_GROUP_SIZE 64
 321
 322 #define HAMMER_FLUSH_SIGNAL     0x0001
 323 #define HAMMER_FLUSH_RECURSION  0x0002
 324
 325 /*
 326  * Used by the inode reclaim code to pipeline reclaims and avoid
 327  * blowing out kernel memory or letting the flusher get too far
 328  * behind.
 329  */
 330 struct hammer_reclaim {
 331         TAILQ_ENTRY(hammer_reclaim) entry;
 332         int     okydoky;
 333 };
 334
 335 #define HAMMER_RECLAIM_FLUSH    2000
 336 #define HAMMER_RECLAIM_WAIT     4000
 337
 338 /*
 339  * Structure used to represent an unsynchronized record in-memory.  These
 340  * records typically represent directory entries.  Only non-historical
 341  * records are kept in-memory.
 342  *
 343  * Records are organized as a per-inode RB-Tree.  If the inode is not
 344  * on disk then neither are any records and the in-memory record tree
 345  * represents the entire contents of the inode.  If the inode is on disk
 346  * then the on-disk B-Tree is scanned in parallel with the in-memory
 347  * RB-Tree to synthesize the current state of the file.
 348  *
 349  * Records are also used to enforce the ordering of directory create/delete
 350  * operations.  A new inode will not be flushed to disk unless its related
 351  * directory entry is also being flushed at the same time.  A directory entry
 352  * will not be removed unless its related inode is also being removed at the
 353  * same time.
 354  */
 355 typedef enum hammer_record_type {
 356         HAMMER_MEM_RECORD_GENERAL,      /* misc record */
 357         HAMMER_MEM_RECORD_INODE,        /* inode record */
 358         HAMMER_MEM_RECORD_ADD,          /* positive memory cache record */
 359         HAMMER_MEM_RECORD_DEL,          /* negative delete-on-disk record */
 360         HAMMER_MEM_RECORD_DATA          /* bulk-data record w/on-disk ref */
 361 } hammer_record_type_t;
 362
 363 struct hammer_record {
 364         RB_ENTRY(hammer_record)         rb_node;
 365         TAILQ_ENTRY(hammer_record)      target_entry;
 366         hammer_inode_state_t            flush_state;
 367         int                             flush_group;
 368         hammer_record_type_t            type;
 369         struct hammer_lock              lock;
 370         struct hammer_reserve           *resv;
 371         struct hammer_inode             *ip;
 372         struct hammer_inode             *target_ip;
 373         struct hammer_btree_leaf_elm    leaf;
 374         union hammer_data_ondisk        *data;
 375         int                             flags;
 376 };
 377
 378 typedef struct hammer_record *hammer_record_t;
 379
 380 /*
 381  * Record flags.  Note that FE can only be set by the frontend if the
 382  * record has not been interlocked by the backend w/ BE.
 383  */
 384 #define HAMMER_RECF_ALLOCDATA           0x0001
 385 #define HAMMER_RECF_ONRBTREE            0x0002
 386 #define HAMMER_RECF_DELETED_FE          0x0004  /* deleted (frontend) */
 387 #define HAMMER_RECF_DELETED_BE          0x0008  /* deleted (backend) */
 388 #define HAMMER_RECF_UNUSED0010          0x0010
 389 #define HAMMER_RECF_INTERLOCK_BE        0x0020  /* backend interlock */
 390 #define HAMMER_RECF_WANTED              0x0040  /* wanted by the frontend */
 391 #define HAMMER_RECF_CONVERT_DELETE      0x0100 /* special case */
 392
 393 /*
 394  * hammer_delete_at_cursor() flags
 395  */
 396 #define HAMMER_DELETE_ADJUST            0x0001
 397 #define HAMMER_DELETE_DESTROY           0x0002
 398
 399 /*
 400  * In-memory structures representing on-disk structures.
 401  */
 402 struct hammer_volume;
 403 struct hammer_buffer;
 404 struct hammer_node;
 405 struct hammer_undo;
 406 struct hammer_reserve;
 407
 408 RB_HEAD(hammer_vol_rb_tree, hammer_volume);
 409 RB_HEAD(hammer_buf_rb_tree, hammer_buffer);
 410 RB_HEAD(hammer_nod_rb_tree, hammer_node);
 411 RB_HEAD(hammer_und_rb_tree, hammer_undo);
 412 RB_HEAD(hammer_res_rb_tree, hammer_reserve);
 413
 414 RB_PROTOTYPE2(hammer_vol_rb_tree, hammer_volume, rb_node,
 415               hammer_vol_rb_compare, int32_t);
 416 RB_PROTOTYPE2(hammer_buf_rb_tree, hammer_buffer, rb_node,
 417               hammer_buf_rb_compare, hammer_off_t);
 418 RB_PROTOTYPE2(hammer_nod_rb_tree, hammer_node, rb_node,
 419               hammer_nod_rb_compare, hammer_off_t);
 420 RB_PROTOTYPE2(hammer_und_rb_tree, hammer_undo, rb_node,
 421               hammer_und_rb_compare, hammer_off_t);
 422 RB_PROTOTYPE2(hammer_res_rb_tree, hammer_reserve, rb_node,
 423               hammer_res_rb_compare, hammer_off_t);
 424
 425 /*
 426  * IO management - embedded at the head of various in-memory structures
 427  *
 428  * VOLUME       - hammer_volume containing meta-data
 429  * META_BUFFER  - hammer_buffer containing meta-data
 430  * DATA_BUFFER  - hammer_buffer containing pure-data
 431  *
 432  * Dirty volume headers and dirty meta-data buffers are locked until the
 433  * flusher can sequence them out.  Dirty pure-data buffers can be written.
 434  * Clean buffers can be passively released.
 435  */
 436 typedef enum hammer_io_type {
 437         HAMMER_STRUCTURE_VOLUME,
 438         HAMMER_STRUCTURE_META_BUFFER,
 439         HAMMER_STRUCTURE_UNDO_BUFFER,
 440         HAMMER_STRUCTURE_DATA_BUFFER
 441 } hammer_io_type_t;
 442
 443 union hammer_io_structure;
 444 struct hammer_io;
 445
 446 struct worklist {
 447         LIST_ENTRY(worklist) node;
 448 };
 449
 450 TAILQ_HEAD(hammer_io_list, hammer_io);
 451 typedef struct hammer_io_list *hammer_io_list_t;
 452
 453 struct hammer_io {
 454         struct worklist         worklist;
 455         struct hammer_lock      lock;
 456         enum hammer_io_type     type;
 457         struct hammer_mount     *hmp;
 458         TAILQ_ENTRY(hammer_io)  mod_entry; /* list entry if modified */
 459         hammer_io_list_t        mod_list;
 460         struct buf              *bp;
 461         int64_t                 offset;    /* zone-2 offset */
 462         int                     bytes;     /* buffer cache buffer size */
 463         int                     loading;   /* loading/unloading interlock */
 464         int                     modify_refs;
 465
 466         u_int           modified : 1;   /* bp's data was modified */
 467         u_int           released : 1;   /* bp released (w/ B_LOCKED set) */
 468         u_int           running : 1;    /* bp write IO in progress */
 469         u_int           waiting : 1;    /* someone is waiting on us */
 470         u_int           validated : 1;  /* ondisk has been validated */
 471         u_int           waitdep : 1;    /* flush waits for dependancies */
 472         u_int           recovered : 1;  /* has recovery ref */
 473         u_int           waitmod : 1;    /* waiting for modify_refs */
 474         u_int           reclaim : 1;    /* reclaim requested */
 475         u_int           gencrc : 1;     /* crc needs to be generated */
 476 };
 477
 478 typedef struct hammer_io *hammer_io_t;
 479
 480 #define HAMMER_CLUSTER_SIZE     (64 * 1024)
 481 #if HAMMER_CLUSTER_SIZE > MAXBSIZE
 482 #undef  HAMMER_CLUSTER_SIZE
 483 #define HAMMER_CLUSTER_SIZE     MAXBSIZE
 484 #endif
 485 #define HAMMER_CLUSTER_BUFS     (HAMMER_CLUSTER_SIZE / HAMMER_BUFSIZE)
 486
 487 /*
 488  * In-memory volume representing on-disk buffer
 489  */
 490 struct hammer_volume {
 491         struct hammer_io io;
 492         RB_ENTRY(hammer_volume) rb_node;
 493         struct hammer_volume_ondisk *ondisk;
 494         int32_t vol_no;
 495         int64_t nblocks;        /* note: special calculation for statfs */
 496         int64_t buffer_base;    /* base offset of buffer 0 */
 497         hammer_off_t maxbuf_off; /* Maximum buffer offset (zone-2) */
 498         hammer_off_t maxraw_off; /* Maximum raw offset for device */
 499         char    *vol_name;
 500         struct vnode *devvp;
 501         int     vol_flags;
 502 };
 503
 504 typedef struct hammer_volume *hammer_volume_t;
 505
 506 /*
 507  * In-memory buffer (other then volume, super-cluster, or cluster),
 508  * representing an on-disk buffer.
 509  */
 510 struct hammer_buffer {
 511         struct hammer_io io;
 512         RB_ENTRY(hammer_buffer) rb_node;
 513         void *ondisk;
 514         struct hammer_volume *volume;
 515         hammer_off_t zoneX_offset;
 516         hammer_off_t zone2_offset;
 517         struct hammer_reserve *resv;
 518         struct hammer_node_list clist;
 519 };
 520
 521 typedef struct hammer_buffer *hammer_buffer_t;
 522
 523 /*
 524  * In-memory B-Tree node, representing an on-disk B-Tree node.
 525  *
 526  * This is a hang-on structure which is backed by a hammer_buffer,
 527  * indexed by a hammer_cluster, and used for fine-grained locking of
 528  * B-Tree nodes in order to properly control lock ordering.  A hammer_buffer
 529  * can contain multiple nodes representing wildly disassociated portions
 530  * of the B-Tree so locking cannot be done on a buffer-by-buffer basis.
 531  *
 532  * This structure uses a cluster-relative index to reduce the number
 533  * of layers required to access it, and also because all on-disk B-Tree
 534  * references are cluster-relative offsets.
 535  */
 536 struct hammer_node {
 537         struct hammer_lock      lock;           /* node-by-node lock */
 538         TAILQ_ENTRY(hammer_node) entry;         /* per-buffer linkage */
 539         RB_ENTRY(hammer_node)   rb_node;        /* per-cluster linkage */
 540         hammer_off_t            node_offset;    /* full offset spec */
 541         struct hammer_mount     *hmp;
 542         struct hammer_buffer    *buffer;        /* backing buffer */
 543         hammer_node_ondisk_t    ondisk;         /* ptr to on-disk structure */
 544         TAILQ_HEAD(, hammer_cursor) cursor_list;  /* deadlock recovery */
 545         struct hammer_node_cache_list cache_list; /* passive caches */
 546         int                     flags;
 547         int                     loading;        /* load interlock */
 548 };
 549
 550 #define HAMMER_NODE_DELETED     0x0001
 551 #define HAMMER_NODE_FLUSH       0x0002
 552 #define HAMMER_NODE_CRCGOOD     0x0004
 553 #define HAMMER_NODE_NEEDSCRC    0x0008
 554 #define HAMMER_NODE_NEEDSMIRROR 0x0010
 555
 556 typedef struct hammer_node      *hammer_node_t;
 557
 558 /*
 559  * List of locked nodes.
 560  */
 561 struct hammer_node_locklist {
 562         struct hammer_node_locklist *next;
 563         hammer_node_t   node;
 564 };
 565
 566 typedef struct hammer_node_locklist *hammer_node_locklist_t;
 567
 568
 569 /*
 570  * Common I/O management structure - embedded in in-memory structures
 571  * which are backed by filesystem buffers.
 572  */
 573 union hammer_io_structure {
 574         struct hammer_io        io;
 575         struct hammer_volume    volume;
 576         struct hammer_buffer    buffer;
 577 };
 578
 579 typedef union hammer_io_structure *hammer_io_structure_t;
 580
 581 /*
 582  * The reserve structure prevents the blockmap from allocating
 583  * out of a reserved bigblock.  Such reservations are used by
 584  * the direct-write mechanism.
 585  *
 586  * The structure is also used to hold off on reallocations of
 587  * big blocks from the freemap until flush dependancies have
 588  * been dealt with.
 589  */
 590 struct hammer_reserve {
 591         RB_ENTRY(hammer_reserve) rb_node;
 592         TAILQ_ENTRY(hammer_reserve) delay_entry;
 593         int             flush_group;
 594         int             flags;
 595         int             refs;
 596         int             zone;
 597         int             append_off;
 598         hammer_off_t    zone_offset;
 599 };
 600
 601 typedef struct hammer_reserve *hammer_reserve_t;
 602
 603 #define HAMMER_RESF_ONDELAY     0x0001
 604
 605 #include "hammer_cursor.h"
 606
 607 /*
 608  * The undo structure tracks recent undos to avoid laying down duplicate
 609  * undos within a flush group, saving us a significant amount of overhead.
 610  *
 611  * This is strictly a heuristic.
 612  */
 613 #define HAMMER_MAX_UNDOS        1024
 614 #define HAMMER_MAX_FLUSHERS     4
 615
 616 struct hammer_undo {
 617         RB_ENTRY(hammer_undo)   rb_node;
 618         TAILQ_ENTRY(hammer_undo) lru_entry;
 619         hammer_off_t            offset;
 620         int                     bytes;
 621 };
 622
 623 typedef struct hammer_undo *hammer_undo_t;
 624
 625 struct hammer_flusher_info;
 626
 627 struct hammer_flusher {
 628         int             signal;         /* flusher thread sequencer */
 629         int             act;            /* currently active flush group */
 630         int             done;           /* set to act when complete */
 631         int             next;           /* next flush group */
 632         int             group_lock;     /* lock sequencing of the next flush */
 633         int             exiting;        /* request master exit */
 634         int             count;          /* number of slave flushers */
 635         int             running;        /* number of slave flushers running */
 636         thread_t        td;             /* master flusher thread */
 637         hammer_tid_t    tid;            /* last flushed transaction id */
 638         int             finalize_want;          /* serialize finalization */
 639         struct hammer_lock finalize_lock;       /* serialize finalization */
 640         struct hammer_transaction trans;        /* shared transaction */
 641         struct hammer_flusher_info *info[HAMMER_MAX_FLUSHERS];
 642 };
 643
 644 /*
 645  * Internal hammer mount data structure
 646  */
 647 struct hammer_mount {
 648         struct mount *mp;
 649         /*struct vnode *rootvp;*/
 650         struct hammer_ino_rb_tree rb_inos_root;
 651         struct hammer_vol_rb_tree rb_vols_root;
 652         struct hammer_nod_rb_tree rb_nods_root;
 653         struct hammer_und_rb_tree rb_undo_root;
 654         struct hammer_res_rb_tree rb_resv_root;
 655         struct hammer_buf_rb_tree rb_bufs_root;
 656         struct hammer_pfs_rb_tree rb_pfsm_root;
 657         struct hammer_volume *rootvol;
 658         struct hammer_base_elm root_btree_beg;
 659         struct hammer_base_elm root_btree_end;
 660         int     flags;
 661         int     hflags;
 662         int     ronly;
 663         int     nvolumes;
 664         int     volume_iterator;
 665         int     masterid;       /* -1 or 0-15 - clustering and mirroring */
 666         int     rsv_inodes;     /* reserved space due to dirty inodes */
 667         int64_t rsv_databytes;  /* reserved space due to record data */
 668         int     rsv_recs;       /* reserved space due to dirty records */
 669         int     rsv_fromdelay;  /* bigblocks reserved due to flush delay */
 670         int     last_newrecords;
 671         int     count_newrecords;
 672
 673         int     inode_reclaims; /* inodes pending reclaim by flusher */
 674         int     count_inodes;   /* total number of inodes */
 675         int     count_iqueued;  /* inodes queued to flusher */
 676
 677         struct hammer_flusher flusher;
 678
 679         u_int   check_interrupt;
 680         uuid_t  fsid;
 681         struct hammer_io_list volu_list;        /* dirty undo buffers */
 682         struct hammer_io_list undo_list;        /* dirty undo buffers */
 683         struct hammer_io_list data_list;        /* dirty data buffers */
 684         struct hammer_io_list alt_data_list;    /* dirty data buffers */
 685         struct hammer_io_list meta_list;        /* dirty meta bufs    */
 686         struct hammer_io_list lose_list;        /* loose buffers      */
 687         int     locked_dirty_space;             /* meta/volu count    */
 688         int     io_running_space;
 689         int     objid_cache_count;
 690         hammer_tid_t    asof;                   /* snapshot mount */
 691         hammer_off_t    next_tid;
 692         int64_t copy_stat_freebigblocks;        /* number of free bigblocks */
 693
 694         u_int32_t namekey_iterator;
 695         struct netexport export;
 696         struct hammer_lock sync_lock;
 697         struct hammer_lock free_lock;
 698         struct hammer_lock undo_lock;
 699         struct hammer_lock blkmap_lock;
 700         struct hammer_blockmap  blockmap[HAMMER_MAX_ZONES];
 701         struct hammer_undo      undos[HAMMER_MAX_UNDOS];
 702         int                     undo_alloc;
 703         TAILQ_HEAD(, hammer_undo)  undo_lru_list;
 704         TAILQ_HEAD(, hammer_inode) flush_list;
 705         TAILQ_HEAD(, hammer_reserve) delay_list;
 706         TAILQ_HEAD(, hammer_objid_cache) objid_cache_list;
 707         TAILQ_HEAD(, hammer_reclaim) reclaim_list;
 708 };
 709
 710 typedef struct hammer_mount     *hammer_mount_t;
 711
 712 #define HAMMER_MOUNT_UNUSED0001 0x0001
 713
 714 struct hammer_sync_info {
 715         int error;
 716         int waitfor;
 717 };
 718
 719 #endif
 720
 721 /*
 722  * checkspace slop (8MB chunks), higher numbers are more conservative.
 723  */
 724 #define HAMMER_CHECKSPACE_SLOP_REBLOCK  25
 725 #define HAMMER_CHECKSPACE_SLOP_WRITE    20
 726 #define HAMMER_CHECKSPACE_SLOP_CREATE   20
 727 #define HAMMER_CHECKSPACE_SLOP_REMOVE   10
 728
 729 #if defined(_KERNEL)
 730
 731 extern struct vop_ops hammer_vnode_vops;
 732 extern struct vop_ops hammer_spec_vops;
 733 extern struct vop_ops hammer_fifo_vops;
 734 extern struct bio_ops hammer_bioops;
 735
 736 extern int hammer_debug_io;
 737 extern int hammer_debug_general;
 738 extern int hammer_debug_debug;
 739 extern int hammer_debug_inode;
 740 extern int hammer_debug_locks;
 741 extern int hammer_debug_btree;
 742 extern int hammer_debug_tid;
 743 extern int hammer_debug_recover;
 744 extern int hammer_debug_recover_faults;
 745 extern int hammer_debug_cluster_enable;
 746 extern int hammer_count_inodes;
 747 extern int hammer_count_iqueued;
 748 extern int hammer_count_reclaiming;
 749 extern int hammer_count_records;
 750 extern int hammer_count_record_datas;
 751 extern int hammer_count_volumes;
 752 extern int hammer_count_buffers;
 753 extern int hammer_count_nodes;
 754 extern int64_t hammer_count_extra_space_used;
 755 extern int64_t hammer_stats_btree_lookups;
 756 extern int64_t hammer_stats_btree_searches;
 757 extern int64_t hammer_stats_btree_inserts;
 758 extern int64_t hammer_stats_btree_deletes;
 759 extern int64_t hammer_stats_btree_elements;
 760 extern int64_t hammer_stats_btree_splits;
 761 extern int64_t hammer_stats_btree_iterations;
 762 extern int64_t hammer_stats_record_iterations;
 763 extern int hammer_count_dirtybufspace;
 764 extern int hammer_count_refedbufs;
 765 extern int hammer_count_reservations;
 766 extern int hammer_count_io_running_read;
 767 extern int hammer_count_io_running_write;
 768 extern int hammer_count_io_locked;
 769 extern int hammer_limit_dirtybufspace;
 770 extern int hammer_limit_iqueued;
 771 extern int hammer_limit_recs;
 772 extern int hammer_bio_count;
 773 extern int hammer_verify_zone;
 774 extern int hammer_write_mode;
 775 extern int64_t hammer_contention_count;
 776
 777 int     hammer_vop_inactive(struct vop_inactive_args *);
 778 int     hammer_vop_reclaim(struct vop_reclaim_args *);
 779 int     hammer_get_vnode(struct hammer_inode *ip, struct vnode **vpp);
 780 struct hammer_inode *hammer_get_inode(hammer_transaction_t trans,
 781                         hammer_inode_t dip, int64_t obj_id,
 782                         hammer_tid_t asof, u_int32_t localization,
 783                         int flags, int *errorp);
 784 void    hammer_scan_inode_snapshots(hammer_mount_t hmp,
 785                         hammer_inode_info_t iinfo,
 786                         int (*callback)(hammer_inode_t ip, void *data),
 787                         void *data);
 788 void    hammer_put_inode(struct hammer_inode *ip);
 789 void    hammer_put_inode_ref(struct hammer_inode *ip);
 790 void    hammer_inode_waitreclaims(hammer_mount_t hmp);
 791
 792 int     hammer_unload_volume(hammer_volume_t volume, void *data __unused);
 793 int     hammer_adjust_volume_mode(hammer_volume_t volume, void *data __unused);
 794
 795 int     hammer_unload_buffer(hammer_buffer_t buffer, void *data __unused);
 796 int     hammer_install_volume(hammer_mount_t hmp, const char *volname);
 797
 798 int     hammer_ip_lookup(hammer_cursor_t cursor);
 799 int     hammer_ip_first(hammer_cursor_t cursor);
 800 int     hammer_ip_next(hammer_cursor_t cursor);
 801 int     hammer_ip_resolve_data(hammer_cursor_t cursor);
 802 int     hammer_ip_delete_record(hammer_cursor_t cursor, hammer_inode_t ip,
 803                         hammer_tid_t tid);
 804 int     hammer_delete_at_cursor(hammer_cursor_t cursor, int delete_flags,
 805                         int64_t *stat_bytes);
 806 int     hammer_ip_check_directory_empty(hammer_transaction_t trans,
 807                         hammer_inode_t ip);
 808 int     hammer_sync_hmp(hammer_mount_t hmp, int waitfor);
 809 int     hammer_queue_inodes_flusher(hammer_mount_t hmp, int waitfor);
 810
 811 hammer_record_t
 812         hammer_alloc_mem_record(hammer_inode_t ip, int data_len);
 813 void    hammer_flush_record_done(hammer_record_t record, int error);
 814 void    hammer_wait_mem_record_ident(hammer_record_t record, const char *ident);
 815 void    hammer_rel_mem_record(hammer_record_t record);
 816
 817 int     hammer_cursor_up(hammer_cursor_t cursor);
 818 int     hammer_cursor_up_locked(hammer_cursor_t cursor);
 819 int     hammer_cursor_down(hammer_cursor_t cursor);
 820 int     hammer_cursor_upgrade(hammer_cursor_t cursor);
 821 int     hammer_cursor_upgrade_node(hammer_cursor_t cursor);
 822 void    hammer_cursor_downgrade(hammer_cursor_t cursor);
 823 int     hammer_cursor_seek(hammer_cursor_t cursor, hammer_node_t node,
 824                         int index);
 825 void    hammer_lock_ex_ident(struct hammer_lock *lock, const char *ident);
 826 int     hammer_lock_ex_try(struct hammer_lock *lock);
 827 void    hammer_lock_sh(struct hammer_lock *lock);
 828 void    hammer_lock_sh_lowpri(struct hammer_lock *lock);
 829 int     hammer_lock_sh_try(struct hammer_lock *lock);
 830 int     hammer_lock_upgrade(struct hammer_lock *lock);
 831 void    hammer_lock_downgrade(struct hammer_lock *lock);
 832 int     hammer_lock_status(struct hammer_lock *lock);
 833 void    hammer_unlock(struct hammer_lock *lock);
 834 void    hammer_ref(struct hammer_lock *lock);
 835 void    hammer_unref(struct hammer_lock *lock);
 836
 837 void    hammer_sync_lock_ex(hammer_transaction_t trans);
 838 void    hammer_sync_lock_sh(hammer_transaction_t trans);
 839 int     hammer_sync_lock_sh_try(hammer_transaction_t trans);
 840 void    hammer_sync_unlock(hammer_transaction_t trans);
 841
 842 u_int32_t hammer_to_unix_xid(uuid_t *uuid);
 843 void hammer_guid_to_uuid(uuid_t *uuid, u_int32_t guid);
 844 void    hammer_time_to_timespec(u_int64_t xtime, struct timespec *ts);
 845 u_int64_t hammer_timespec_to_time(struct timespec *ts);
 846 hammer_tid_t hammer_str_to_tid(const char *str);
 847 hammer_tid_t hammer_alloc_objid(hammer_mount_t hmp, hammer_inode_t dip);
 848 void hammer_clear_objid(hammer_inode_t dip);
 849 void hammer_destroy_objid_cache(hammer_mount_t hmp);
 850
 851 int hammer_enter_undo_history(hammer_mount_t hmp, hammer_off_t offset,
 852                         int bytes);
 853 void hammer_clear_undo_history(hammer_mount_t hmp);
 854 enum vtype hammer_get_vnode_type(u_int8_t obj_type);
 855 int hammer_get_dtype(u_int8_t obj_type);
 856 u_int8_t hammer_get_obj_type(enum vtype vtype);
 857 int64_t hammer_directory_namekey(const void *name, int len);
 858 int     hammer_nohistory(hammer_inode_t ip);
 859
 860 int     hammer_init_cursor(hammer_transaction_t trans, hammer_cursor_t cursor,
 861                         hammer_node_cache_t cache, hammer_inode_t ip);
 862 void    hammer_normalize_cursor(hammer_cursor_t cursor);
 863 void    hammer_done_cursor(hammer_cursor_t cursor);
 864 int     hammer_recover_cursor(hammer_cursor_t cursor);
 865 void    hammer_unlock_cursor(hammer_cursor_t cursor, int also_ip);
 866 int     hammer_lock_cursor(hammer_cursor_t cursor, int also_ip);
 867 hammer_cursor_t hammer_push_cursor(hammer_cursor_t ocursor);
 868 void    hammer_pop_cursor(hammer_cursor_t ocursor, hammer_cursor_t ncursor);
 869
 870 void    hammer_cursor_replaced_node(hammer_node_t onode, hammer_node_t nnode);
 871 void    hammer_cursor_removed_node(hammer_node_t onode, hammer_node_t parent,
 872                         int index);
 873 void    hammer_cursor_split_node(hammer_node_t onode, hammer_node_t nnode,
 874                         int index);
 875 void    hammer_cursor_inserted_element(hammer_node_t node, int index);
 876 void    hammer_cursor_deleted_element(hammer_node_t node, int index);
 877
 878 int     hammer_btree_lookup(hammer_cursor_t cursor);
 879 int     hammer_btree_first(hammer_cursor_t cursor);
 880 int     hammer_btree_last(hammer_cursor_t cursor);
 881 int     hammer_btree_extract(hammer_cursor_t cursor, int flags);
 882 int     hammer_btree_iterate(hammer_cursor_t cursor);
 883 int     hammer_btree_iterate_reverse(hammer_cursor_t cursor);
 884 int     hammer_btree_insert(hammer_cursor_t cursor,
 885                             hammer_btree_leaf_elm_t elm, int *doprop);
 886 int     hammer_btree_delete(hammer_cursor_t cursor);
 887 void    hammer_btree_do_propagation(hammer_cursor_t cursor, hammer_inode_t ip,
 888                             hammer_btree_leaf_elm_t leaf);
 889 int     hammer_btree_cmp(hammer_base_elm_t key1, hammer_base_elm_t key2);
 890 int     hammer_btree_chkts(hammer_tid_t ts, hammer_base_elm_t key);
 891 int     hammer_btree_correct_rhb(hammer_cursor_t cursor, hammer_tid_t tid);
 892 int     hammer_btree_correct_lhb(hammer_cursor_t cursor, hammer_tid_t tid);
 893
 894 int     btree_set_parent(hammer_transaction_t trans, hammer_node_t node,
 895                         hammer_btree_elm_t elm);
 896 int     hammer_btree_lock_children(hammer_cursor_t cursor,
 897                         struct hammer_node_locklist **locklistp);
 898 void    hammer_btree_unlock_children(struct hammer_node_locklist **locklistp);
 899 int     hammer_btree_search_node(hammer_base_elm_t elm, hammer_node_ondisk_t node);
 900 hammer_node_t hammer_btree_get_parent(hammer_node_t node, int *parent_indexp,
 901                         int *errorp, int try_exclusive);
 902
 903 void    hammer_print_btree_node(hammer_node_ondisk_t ondisk);
 904 void    hammer_print_btree_elm(hammer_btree_elm_t elm, u_int8_t type, int i);
 905
 906 void    *hammer_bread(struct hammer_mount *hmp, hammer_off_t off,
 907                         int *errorp, struct hammer_buffer **bufferp);
 908 void    *hammer_bnew(struct hammer_mount *hmp, hammer_off_t off,
 909                         int *errorp, struct hammer_buffer **bufferp);
 910 void    *hammer_bread_ext(struct hammer_mount *hmp, hammer_off_t off, int bytes,
 911                         int *errorp, struct hammer_buffer **bufferp);
 912 void    *hammer_bnew_ext(struct hammer_mount *hmp, hammer_off_t off, int bytes,
 913                         int *errorp, struct hammer_buffer **bufferp);
 914
 915 hammer_volume_t hammer_get_root_volume(hammer_mount_t hmp, int *errorp);
 916
 917 hammer_volume_t hammer_get_volume(hammer_mount_t hmp,
 918                         int32_t vol_no, int *errorp);
 919 hammer_buffer_t hammer_get_buffer(hammer_mount_t hmp, hammer_off_t buf_offset,
 920                         int bytes, int isnew, int *errorp);
 921 void            hammer_del_buffers(hammer_mount_t hmp, hammer_off_t base_offset,
 922                         hammer_off_t zone2_offset, int bytes);
 923
 924 int             hammer_ref_volume(hammer_volume_t volume);
 925 int             hammer_ref_buffer(hammer_buffer_t buffer);
 926 void            hammer_flush_buffer_nodes(hammer_buffer_t buffer);
 927
 928 void            hammer_rel_volume(hammer_volume_t volume, int flush);
 929 void            hammer_rel_buffer(hammer_buffer_t buffer, int flush);
 930
 931 int             hammer_vfs_export(struct mount *mp, int op,
 932                         const struct export_args *export);
 933 hammer_node_t   hammer_get_node(hammer_mount_t hmp, hammer_off_t node_offset,
 934                         int isnew, int *errorp);
 935 void            hammer_ref_node(hammer_node_t node);
 936 hammer_node_t   hammer_ref_node_safe(struct hammer_mount *hmp,
 937                         hammer_node_cache_t cache, int *errorp);
 938 void            hammer_rel_node(hammer_node_t node);
 939 void            hammer_delete_node(hammer_transaction_t trans,
 940                         hammer_node_t node);
 941 void            hammer_cache_node(hammer_node_cache_t cache,
 942                         hammer_node_t node);
 943 void            hammer_uncache_node(hammer_node_cache_t cache);
 944 void            hammer_flush_node(hammer_node_t node);
 945
 946 void hammer_dup_buffer(struct hammer_buffer **bufferp,
 947                         struct hammer_buffer *buffer);
 948 hammer_node_t hammer_alloc_btree(hammer_transaction_t trans, int *errorp);
 949 void *hammer_alloc_data(hammer_transaction_t trans, int32_t data_len,
 950                         u_int16_t rec_type, hammer_off_t *data_offsetp,
 951                         struct hammer_buffer **data_bufferp, int *errorp);
 952
 953 int hammer_generate_undo(hammer_transaction_t trans, hammer_io_t io,
 954                         hammer_off_t zone1_offset, void *base, int len);
 955
 956 void hammer_put_volume(struct hammer_volume *volume, int flush);
 957 void hammer_put_buffer(struct hammer_buffer *buffer, int flush);
 958
 959 hammer_off_t hammer_freemap_alloc(hammer_transaction_t trans,
 960                         hammer_off_t owner, int *errorp);
 961 void hammer_freemap_free(hammer_transaction_t trans, hammer_off_t phys_offset,
 962                         hammer_off_t owner, int *errorp);
 963 int hammer_checkspace(hammer_mount_t hmp, int slop);
 964 hammer_off_t hammer_blockmap_alloc(hammer_transaction_t trans, int zone,
 965                         int bytes, int *errorp);
 966 hammer_reserve_t hammer_blockmap_reserve(hammer_mount_t hmp, int zone,
 967                         int bytes, hammer_off_t *zone_offp, int *errorp);
 968 void hammer_blockmap_reserve_complete(hammer_mount_t hmp,
 969                         hammer_reserve_t resv);
 970 void hammer_reserve_clrdelay(hammer_mount_t hmp, hammer_reserve_t resv);
 971 void hammer_blockmap_free(hammer_transaction_t trans,
 972                         hammer_off_t bmap_off, int bytes);
 973 void hammer_blockmap_finalize(hammer_transaction_t trans,
 974                         hammer_off_t bmap_off, int bytes);
 975 int hammer_blockmap_getfree(hammer_mount_t hmp, hammer_off_t bmap_off,
 976                         int *curp, int *errorp);
 977 hammer_off_t hammer_blockmap_lookup(hammer_mount_t hmp, hammer_off_t bmap_off,
 978                         int *errorp);
 979 hammer_off_t hammer_undo_lookup(hammer_mount_t hmp, hammer_off_t bmap_off,
 980                         int *errorp);
 981 int64_t hammer_undo_used(hammer_transaction_t trans);
 982 int64_t hammer_undo_space(hammer_transaction_t trans);
 983 int64_t hammer_undo_max(hammer_mount_t hmp);
 984
 985 void hammer_start_transaction(struct hammer_transaction *trans,
 986                               struct hammer_mount *hmp);
 987 void hammer_simple_transaction(struct hammer_transaction *trans,
 988                               struct hammer_mount *hmp);
 989 void hammer_start_transaction_fls(struct hammer_transaction *trans,
 990                                   struct hammer_mount *hmp);
 991 void hammer_done_transaction(struct hammer_transaction *trans);
 992
 993 void hammer_modify_inode(hammer_inode_t ip, int flags);
 994 void hammer_flush_inode(hammer_inode_t ip, int flags);
 995 void hammer_flush_inode_done(hammer_inode_t ip);
 996 void hammer_wait_inode(hammer_inode_t ip);
 997
 998 int  hammer_create_inode(struct hammer_transaction *trans, struct vattr *vap,
 999                         struct ucred *cred, struct hammer_inode *dip,
1000                         int pseudofs, struct hammer_inode **ipp);
1001 void hammer_rel_inode(hammer_inode_t ip, int flush);
1002 int hammer_reload_inode(hammer_inode_t ip, void *arg __unused);
1003 int hammer_ino_rb_compare(hammer_inode_t ip1, hammer_inode_t ip2);
1004
1005 int hammer_sync_inode(hammer_inode_t ip);
1006 void hammer_test_inode(hammer_inode_t ip);
1007 void hammer_inode_unloadable_check(hammer_inode_t ip, int getvp);
1008
1009 int  hammer_ip_add_directory(struct hammer_transaction *trans,
1010                         hammer_inode_t dip, const char *name, int bytes,
1011                         hammer_inode_t nip);
1012 int  hammer_ip_del_directory(struct hammer_transaction *trans,
1013                         hammer_cursor_t cursor, hammer_inode_t dip,
1014                         hammer_inode_t ip);
1015 hammer_record_t hammer_ip_add_bulk(hammer_inode_t ip, off_t file_offset,
1016                         void *data, int bytes, int *errorp);
1017 int  hammer_ip_frontend_trunc(struct hammer_inode *ip, off_t file_size);
1018 int  hammer_ip_add_record(struct hammer_transaction *trans,
1019                         hammer_record_t record);
1020 int  hammer_ip_delete_range(hammer_cursor_t cursor, hammer_inode_t ip,
1021                         int64_t ran_beg, int64_t ran_end, int truncating);
1022 int  hammer_ip_delete_clean(hammer_cursor_t cursor, hammer_inode_t ip,
1023                         int *countp);
1024 int  hammer_ip_sync_data(hammer_cursor_t cursor, hammer_inode_t ip,
1025                         int64_t offset, void *data, int bytes);
1026 int  hammer_ip_sync_record(hammer_transaction_t trans, hammer_record_t rec);
1027 int  hammer_ip_sync_record_cursor(hammer_cursor_t cursor, hammer_record_t rec);
1028 int  hammer_load_pseudofs(hammer_transaction_t trans, hammer_inode_t ip);
1029 int  hammer_save_pseudofs(hammer_transaction_t trans, hammer_inode_t ip);
1030 void hammer_rel_pseudofs(hammer_mount_t hmp, hammer_pseudofs_inmem_t pfsm);
1031 int hammer_ioctl(hammer_inode_t ip, u_long com, caddr_t data, int fflag,
1032                         struct ucred *cred);
1033
1034 void hammer_io_init(hammer_io_t io, hammer_mount_t hmp,
1035                         enum hammer_io_type type);
1036 int hammer_io_read(struct vnode *devvp, struct hammer_io *io,
1037                         hammer_off_t limit);
1038 int hammer_io_new(struct vnode *devvp, struct hammer_io *io);
1039 void hammer_io_inval(hammer_volume_t volume, hammer_off_t zone2_offset);
1040 struct buf *hammer_io_release(struct hammer_io *io, int flush);
1041 void hammer_io_flush(struct hammer_io *io);
1042 void hammer_io_waitdep(struct hammer_io *io);
1043 void hammer_io_wait_all(hammer_mount_t hmp, const char *ident);
1044 int hammer_io_direct_read(hammer_mount_t hmp, struct bio *bio);
1045 int hammer_io_direct_write(hammer_mount_t hmp, hammer_btree_leaf_elm_t leaf,
1046                           struct bio *bio);
1047 void hammer_io_direct_uncache(hammer_mount_t hmp, hammer_btree_leaf_elm_t leaf);
1048 void hammer_io_write_interlock(hammer_io_t io);
1049 void hammer_io_done_interlock(hammer_io_t io);
1050 void hammer_io_clear_modify(struct hammer_io *io, int inval);
1051 void hammer_io_clear_modlist(struct hammer_io *io);
1052 void hammer_modify_volume(hammer_transaction_t trans, hammer_volume_t volume,
1053                         void *base, int len);
1054 void hammer_modify_buffer(hammer_transaction_t trans, hammer_buffer_t buffer,
1055                         void *base, int len);
1056 void hammer_modify_volume_done(hammer_volume_t volume);
1057 void hammer_modify_buffer_done(hammer_buffer_t buffer);
1058
1059 int hammer_ioc_reblock(hammer_transaction_t trans, hammer_inode_t ip,
1060                         struct hammer_ioc_reblock *reblock);
1061 int hammer_ioc_prune(hammer_transaction_t trans, hammer_inode_t ip,
1062                         struct hammer_ioc_prune *prune);
1063 int hammer_ioc_mirror_read(hammer_transaction_t trans, hammer_inode_t ip,
1064                         struct hammer_ioc_mirror_rw *mirror);
1065 int hammer_ioc_mirror_write(hammer_transaction_t trans, hammer_inode_t ip,
1066                         struct hammer_ioc_mirror_rw *mirror);
1067 int hammer_ioc_set_pseudofs(hammer_transaction_t trans, hammer_inode_t ip,
1068                         struct hammer_ioc_pseudofs_rw *pfs);
1069 int hammer_ioc_get_pseudofs(hammer_transaction_t trans, hammer_inode_t ip,
1070                         struct hammer_ioc_pseudofs_rw *pfs);
1071
1072 int hammer_signal_check(hammer_mount_t hmp);
1073
1074 void hammer_flusher_create(hammer_mount_t hmp);
1075 void hammer_flusher_destroy(hammer_mount_t hmp);
1076 void hammer_flusher_sync(hammer_mount_t hmp);
1077 void hammer_flusher_async(hammer_mount_t hmp);
1078 int  hammer_flusher_meta_limit(hammer_mount_t hmp);
1079 int  hammer_flusher_undo_exhausted(hammer_transaction_t trans, int quarter);
1080 void hammer_flusher_clean_loose_ios(hammer_mount_t hmp);
1081
1082 int hammer_recover(hammer_mount_t hmp, hammer_volume_t rootvol);
1083 void hammer_recover_flush_buffers(hammer_mount_t hmp,
1084                         hammer_volume_t root_volume, int final);
1085
1086 void hammer_crc_set_blockmap(hammer_blockmap_t blockmap);
1087 void hammer_crc_set_volume(hammer_volume_ondisk_t ondisk);
1088 void hammer_crc_set_leaf(void *data, hammer_btree_leaf_elm_t leaf);
1089
1090 int hammer_crc_test_blockmap(hammer_blockmap_t blockmap);
1091 int hammer_crc_test_volume(hammer_volume_ondisk_t ondisk);
1092 int hammer_crc_test_btree(hammer_node_ondisk_t ondisk);
1093 int hammer_crc_test_leaf(void *data, hammer_btree_leaf_elm_t leaf);
1094 void hkprintf(const char *ctl, ...);
1095 udev_t hammer_fsid_to_udev(uuid_t *uuid);
1096
1097
1098 int hammer_blocksize(int64_t file_offset);
1099 int64_t hammer_blockdemarc(int64_t file_offset1, int64_t file_offset2);
1100
1101 #endif
1102
1103 static __inline void
1104 hammer_wait_mem_record(hammer_record_t record)
1105 {
1106         hammer_wait_mem_record_ident(record, "hmmwai");
1107 }
1108
1109 static __inline void
1110 hammer_lock_ex(struct hammer_lock *lock)
1111 {
1112         hammer_lock_ex_ident(lock, "hmrlck");
1113 }
1114
1115 /*
1116  * Indicate that a B-Tree node is being modified.
1117  */
1118 static __inline void
1119 hammer_modify_node_noundo(hammer_transaction_t trans, hammer_node_t node)
1120 {
1121         hammer_modify_buffer(trans, node->buffer, NULL, 0);
1122 }
1123
1124 static __inline void
1125 hammer_modify_node_all(hammer_transaction_t trans, struct hammer_node *node)
1126 {
1127         hammer_modify_buffer(trans, node->buffer,
1128                              node->ondisk, sizeof(*node->ondisk));
1129 }
1130
1131 static __inline void
1132 hammer_modify_node(hammer_transaction_t trans, hammer_node_t node,
1133                    void *base, int len)
1134 {
1135         hammer_crc_t *crcptr;
1136
1137         KKASSERT((char *)base >= (char *)node->ondisk &&
1138                  (char *)base + len <=
1139                     (char *)node->ondisk + sizeof(*node->ondisk));
1140         hammer_modify_buffer(trans, node->buffer, base, len);
1141         crcptr = &node->ondisk->crc;
1142         hammer_modify_buffer(trans, node->buffer, crcptr, sizeof(hammer_crc_t));
1143         --node->buffer->io.modify_refs; /* only want one ref */
1144 }
1145
1146 /*
1147  * Indicate that the specified modifications have been completed.
1148  *
1149  * Do not try to generate the crc here, it's very expensive to do and a
1150  * sequence of insertions or deletions can result in many calls to this
1151  * function on the same node.
1152  */
1153 static __inline void
1154 hammer_modify_node_done(hammer_node_t node)
1155 {
1156         node->flags |= HAMMER_NODE_CRCGOOD;
1157         if ((node->flags & HAMMER_NODE_NEEDSCRC) == 0) {
1158                 node->flags |= HAMMER_NODE_NEEDSCRC;
1159                 node->buffer->io.gencrc = 1;
1160                 hammer_ref_node(node);
1161         }
1162         hammer_modify_buffer_done(node->buffer);
1163 }
1164
1165 #define hammer_modify_volume_field(trans, vol, field)           \
1166         hammer_modify_volume(trans, vol, &(vol)->ondisk->field, \
1167                              sizeof((vol)->ondisk->field))
1168
1169 #define hammer_modify_node_field(trans, node, field)            \
1170         hammer_modify_node(trans, node, &(node)->ondisk->field, \
1171                              sizeof((node)->ondisk->field))
1172