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 1.124 2008/07/31 04:42:04 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 MALLOC_DECLARE(M_HAMMER_INO);
  70
  71 /*
  72  * Kernel trace
  73  */
  74 #if !defined(KTR_HAMMER)
  75 #define KTR_HAMMER      KTR_ALL
  76 #endif
  77 KTR_INFO_MASTER_EXTERN(hammer);
  78
  79 /*
  80  * Misc structures
  81  */
  82 struct hammer_mount;
  83
  84 /*
  85  * Key structure used for custom RB tree inode lookups.  This prototypes
  86  * the function hammer_ino_rb_tree_RB_LOOKUP_INFO(root, info).
  87  */
  88 typedef struct hammer_inode_info {
  89         int64_t         obj_id;         /* (key) object identifier */
  90         hammer_tid_t    obj_asof;       /* (key) snapshot transid or 0 */
  91         u_int32_t       obj_localization; /* (key) pseudo-fs */
  92         union {
  93                 struct hammer_btree_leaf_elm *leaf;
  94         } u;
  95 } *hammer_inode_info_t;
  96
  97 typedef enum hammer_transaction_type {
  98         HAMMER_TRANS_RO,
  99         HAMMER_TRANS_STD,
 100         HAMMER_TRANS_FLS
 101 } hammer_transaction_type_t;
 102
 103 /*
 104  * HAMMER Transaction tracking
 105  */
 106 struct hammer_transaction {
 107         hammer_transaction_type_t type;
 108         struct hammer_mount *hmp;
 109         hammer_tid_t    tid;
 110         u_int64_t       time;
 111         u_int32_t       time32;
 112         int             sync_lock_refs;
 113         struct hammer_volume *rootvol;
 114 };
 115
 116 typedef struct hammer_transaction *hammer_transaction_t;
 117
 118 /*
 119  * HAMMER locks
 120  */
 121 struct hammer_lock {
 122         int     refs;           /* active references delay writes */
 123         int     lockcount;      /* lock count for exclusive/shared access */
 124         int     wanted;
 125         int     exwanted;       /* number of threads waiting for ex lock */
 126         struct thread *locktd;
 127 };
 128
 129 static __inline int
 130 hammer_islocked(struct hammer_lock *lock)
 131 {
 132         return(lock->lockcount != 0);
 133 }
 134
 135 static __inline int
 136 hammer_isactive(struct hammer_lock *lock)
 137 {
 138         return(lock->refs != 0);
 139 }
 140
 141 static __inline int
 142 hammer_islastref(struct hammer_lock *lock)
 143 {
 144         return(lock->refs == 1);
 145 }
 146
 147 /*
 148  * Return if we specifically own the lock exclusively.
 149  */
 150 static __inline int
 151 hammer_lock_excl_owned(struct hammer_lock *lock, thread_t td)
 152 {
 153         if (lock->lockcount > 0 && lock->locktd == td)
 154                 return(1);
 155         return(0);
 156 }
 157
 158 /*
 159  * Flush state, used by various structures
 160  */
 161 typedef enum hammer_inode_state {
 162         HAMMER_FST_IDLE,
 163         HAMMER_FST_SETUP,
 164         HAMMER_FST_FLUSH
 165 } hammer_inode_state_t;
 166
 167 TAILQ_HEAD(hammer_record_list, hammer_record);
 168
 169 /*
 170  * Pseudo-filesystem extended data tracking
 171  */
 172 struct hammer_pfs_rb_tree;
 173 struct hammer_pseudofs_inmem;
 174 RB_HEAD(hammer_pfs_rb_tree, hammer_pseudofs_inmem);
 175 RB_PROTOTYPE2(hammer_pfs_rb_tree, hammer_pseudofs_inmem, rb_node,
 176               hammer_pfs_rb_compare, u_int32_t);
 177
 178 struct hammer_pseudofs_inmem {
 179         RB_ENTRY(hammer_pseudofs_inmem) rb_node;
 180         struct hammer_lock      lock;
 181         u_int32_t               localization;
 182         hammer_tid_t            create_tid;
 183         int                     flags;
 184         udev_t                  fsid_udev;
 185         struct hammer_pseudofs_data pfsd;
 186 };
 187
 188 typedef struct hammer_pseudofs_inmem *hammer_pseudofs_inmem_t;
 189
 190 #define HAMMER_PFSM_DELETED     0x0001
 191
 192 /*
 193  * Cache object ids.  A fixed number of objid cache structures are
 194  * created to reserve object id's for newly created files in multiples
 195  * of 100,000, localized to a particular directory, and recycled as
 196  * needed.  This allows parallel create operations in different
 197  * directories to retain fairly localized object ids which in turn
 198  * improves reblocking performance and layout.
 199  */
 200 #define OBJID_CACHE_SIZE        1024
 201 #define OBJID_CACHE_BULK        100000
 202
 203 typedef struct hammer_objid_cache {
 204         TAILQ_ENTRY(hammer_objid_cache) entry;
 205         struct hammer_inode             *dip;
 206         hammer_tid_t                    next_tid;
 207         int                             count;
 208 } *hammer_objid_cache_t;
 209
 210 /*
 211  * Associate an inode with a B-Tree node to cache search start positions
 212  */
 213 typedef struct hammer_node_cache {
 214         TAILQ_ENTRY(hammer_node_cache)  entry;
 215         struct hammer_node              *node;
 216         struct hammer_inode             *ip;
 217 } *hammer_node_cache_t;
 218
 219 TAILQ_HEAD(hammer_node_cache_list, hammer_node_cache);
 220
 221 /*
 222  * Structure used to organize flush groups.  Flush groups must be
 223  * organized into chunks in order to avoid blowing out the UNDO FIFO.
 224  * Without this a 'sync' could end up flushing 50,000 inodes in a single
 225  * transaction.
 226  */
 227 struct hammer_flush_group {
 228         TAILQ_ENTRY(hammer_flush_group) flush_entry;
 229         TAILQ_HEAD(, hammer_inode)      flush_list;
 230         int                             inode_count;    /* inode load */
 231         int                             total_count;    /* record load */
 232         int                             running;        /* group is running */
 233         int                             closed;
 234         int                             refs;
 235 };
 236
 237 typedef struct hammer_flush_group *hammer_flush_group_t;
 238
 239 TAILQ_HEAD(hammer_flush_group_list, hammer_flush_group);
 240
 241 /*
 242  * Structure used to represent an inode in-memory.
 243  *
 244  * The record and data associated with an inode may be out of sync with
 245  * the disk (xDIRTY flags), or not even on the disk at all (ONDISK flag
 246  * clear).
 247  *
 248  * An inode may also hold a cache of unsynchronized records, used for
 249  * database and directories only.  Unsynchronized regular file data is
 250  * stored in the buffer cache.
 251  *
 252  * NOTE: A file which is created and destroyed within the initial
 253  * synchronization period can wind up not doing any disk I/O at all.
 254  *
 255  * Finally, an inode may cache numerous disk-referencing B-Tree cursors.
 256  */
 257 struct hammer_ino_rb_tree;
 258 struct hammer_inode;
 259 RB_HEAD(hammer_ino_rb_tree, hammer_inode);
 260 RB_PROTOTYPEX(hammer_ino_rb_tree, INFO, hammer_inode, rb_node,
 261               hammer_ino_rb_compare, hammer_inode_info_t);
 262
 263 struct hammer_rec_rb_tree;
 264 struct hammer_record;
 265 RB_HEAD(hammer_rec_rb_tree, hammer_record);
 266 RB_PROTOTYPEX(hammer_rec_rb_tree, INFO, hammer_record, rb_node,
 267               hammer_rec_rb_compare, hammer_btree_leaf_elm_t);
 268
 269 TAILQ_HEAD(hammer_node_list, hammer_node);
 270
 271 struct hammer_inode {
 272         RB_ENTRY(hammer_inode)  rb_node;
 273         hammer_inode_state_t    flush_state;
 274         hammer_flush_group_t    flush_group;
 275         TAILQ_ENTRY(hammer_inode) flush_entry;
 276         struct hammer_record_list target_list;  /* target of dependant recs */
 277         int64_t                 obj_id;         /* (key) object identifier */
 278         hammer_tid_t            obj_asof;       /* (key) snapshot or 0 */
 279         u_int32_t               obj_localization; /* (key) pseudo-fs */
 280         struct hammer_mount     *hmp;
 281         hammer_objid_cache_t    objid_cache;
 282         int                     flags;
 283         int                     error;          /* flush error */
 284         int                     cursor_ip_refs; /* sanity */
 285         int                     rsv_recs;
 286         struct vnode            *vp;
 287         hammer_pseudofs_inmem_t pfsm;
 288         struct lockf            advlock;
 289         struct hammer_lock      lock;           /* sync copy interlock */
 290         off_t                   trunc_off;
 291         struct hammer_btree_leaf_elm ino_leaf;  /* in-memory cache */
 292         struct hammer_inode_data ino_data;      /* in-memory cache */
 293         struct hammer_rec_rb_tree rec_tree;     /* in-memory cache */
 294         struct hammer_node_cache cache[2];      /* search initiate cache */
 295
 296         /*
 297          * When a demark is created to synchronize an inode to
 298          * disk, certain fields are copied so the front-end VOPs
 299          * can continue to run in parallel with the synchronization
 300          * occuring in the background.
 301          */
 302         int             sync_flags;             /* to-sync flags cache */
 303         off_t           sync_trunc_off;         /* to-sync truncation */
 304         off_t           save_trunc_off;         /* write optimization */
 305         struct hammer_btree_leaf_elm sync_ino_leaf; /* to-sync cache */
 306         struct hammer_inode_data sync_ino_data; /* to-sync cache */
 307 };
 308
 309 typedef struct hammer_inode *hammer_inode_t;
 310
 311 #define VTOI(vp)        ((struct hammer_inode *)(vp)->v_data)
 312
 313 #define HAMMER_INODE_DDIRTY     0x0001  /* in-memory ino_data is dirty */
 314                                         /* (not including atime/mtime) */
 315 #define HAMMER_INODE_RSV_INODES 0x0002  /* hmp->rsv_inodes bumped */
 316 #define HAMMER_INODE_CONN_DOWN  0x0004  /* include in downward recursion */
 317 #define HAMMER_INODE_XDIRTY     0x0008  /* in-memory records */
 318 #define HAMMER_INODE_ONDISK     0x0010  /* inode is on-disk (else not yet) */
 319 #define HAMMER_INODE_FLUSH      0x0020  /* flush on last ref */
 320 #define HAMMER_INODE_DELETED    0x0080  /* inode delete (backend) */
 321 #define HAMMER_INODE_DELONDISK  0x0100  /* delete synchronized to disk */
 322 #define HAMMER_INODE_RO         0x0200  /* read-only (because of as-of) */
 323 #define HAMMER_INODE_VHELD      0x0400  /* vnode held on sync */
 324 #define HAMMER_INODE_DONDISK    0x0800  /* data records may be on disk */
 325 #define HAMMER_INODE_BUFS       0x1000  /* dirty high level bps present */
 326 #define HAMMER_INODE_REFLUSH    0x2000  /* pipelined flush during flush */
 327 #define HAMMER_INODE_RECLAIM    0x4000  /* trying to reclaim */
 328 #define HAMMER_INODE_FLUSHW     0x8000  /* Someone waiting for flush */
 329
 330 #define HAMMER_INODE_TRUNCATED  0x00010000
 331 #define HAMMER_INODE_DELETING   0x00020000 /* inode delete request (frontend)*/
 332 #define HAMMER_INODE_RESIGNAL   0x00040000 /* re-signal on re-flush */
 333 #define HAMMER_INODE_ATIME      0x00100000 /* in-memory atime modified */
 334 #define HAMMER_INODE_MTIME      0x00200000 /* in-memory mtime modified */
 335 #define HAMMER_INODE_WOULDBLOCK 0x00400000 /* re-issue to new flush group */
 336
 337 #define HAMMER_INODE_MODMASK    (HAMMER_INODE_DDIRTY|                       \
 338                                  HAMMER_INODE_XDIRTY|HAMMER_INODE_BUFS|     \
 339                                  HAMMER_INODE_ATIME|HAMMER_INODE_MTIME|     \
 340                                  HAMMER_INODE_TRUNCATED|HAMMER_INODE_DELETING)
 341
 342 #define HAMMER_INODE_MODMASK_NOXDIRTY \
 343                                 (HAMMER_INODE_MODMASK & ~HAMMER_INODE_XDIRTY)
 344
 345 #define HAMMER_FLUSH_GROUP_SIZE 64
 346
 347 #define HAMMER_FLUSH_SIGNAL     0x0001
 348 #define HAMMER_FLUSH_RECURSION  0x0002
 349
 350 /*
 351  * Used by the inode reclaim code to pipeline reclaims and avoid
 352  * blowing out kernel memory or letting the flusher get too far
 353  * behind.
 354  */
 355 struct hammer_reclaim {
 356         TAILQ_ENTRY(hammer_reclaim) entry;
 357         int     okydoky;
 358 };
 359
 360 #define HAMMER_RECLAIM_FLUSH    2000
 361 #define HAMMER_RECLAIM_WAIT     4000
 362
 363 /*
 364  * Structure used to represent an unsynchronized record in-memory.  These
 365  * records typically represent directory entries.  Only non-historical
 366  * records are kept in-memory.
 367  *
 368  * Records are organized as a per-inode RB-Tree.  If the inode is not
 369  * on disk then neither are any records and the in-memory record tree
 370  * represents the entire contents of the inode.  If the inode is on disk
 371  * then the on-disk B-Tree is scanned in parallel with the in-memory
 372  * RB-Tree to synthesize the current state of the file.
 373  *
 374  * Records are also used to enforce the ordering of directory create/delete
 375  * operations.  A new inode will not be flushed to disk unless its related
 376  * directory entry is also being flushed at the same time.  A directory entry
 377  * will not be removed unless its related inode is also being removed at the
 378  * same time.
 379  */
 380 typedef enum hammer_record_type {
 381         HAMMER_MEM_RECORD_GENERAL,      /* misc record */
 382         HAMMER_MEM_RECORD_INODE,        /* inode record */
 383         HAMMER_MEM_RECORD_ADD,          /* positive memory cache record */
 384         HAMMER_MEM_RECORD_DEL,          /* negative delete-on-disk record */
 385         HAMMER_MEM_RECORD_DATA          /* bulk-data record w/on-disk ref */
 386 } hammer_record_type_t;
 387
 388 struct hammer_record {
 389         RB_ENTRY(hammer_record)         rb_node;
 390         TAILQ_ENTRY(hammer_record)      target_entry;
 391         hammer_inode_state_t            flush_state;
 392         hammer_flush_group_t            flush_group;
 393         hammer_record_type_t            type;
 394         struct hammer_lock              lock;
 395         struct hammer_reserve           *resv;
 396         struct hammer_inode             *ip;
 397         struct hammer_inode             *target_ip;
 398         struct hammer_btree_leaf_elm    leaf;
 399         union hammer_data_ondisk        *data;
 400         int                             flags;
 401 };
 402
 403 typedef struct hammer_record *hammer_record_t;
 404
 405 /*
 406  * Record flags.  Note that FE can only be set by the frontend if the
 407  * record has not been interlocked by the backend w/ BE.
 408  */
 409 #define HAMMER_RECF_ALLOCDATA           0x0001
 410 #define HAMMER_RECF_ONRBTREE            0x0002
 411 #define HAMMER_RECF_DELETED_FE          0x0004  /* deleted (frontend) */
 412 #define HAMMER_RECF_DELETED_BE          0x0008  /* deleted (backend) */
 413 #define HAMMER_RECF_COMMITTED           0x0010  /* committed to the B-Tree */
 414 #define HAMMER_RECF_INTERLOCK_BE        0x0020  /* backend interlock */
 415 #define HAMMER_RECF_WANTED              0x0040  /* wanted by the frontend */
 416 #define HAMMER_RECF_CONVERT_DELETE      0x0100  /* special case */
 417 #define HAMMER_RECF_DIRECT_IO           0x0200  /* related direct I/O running*/
 418 #define HAMMER_RECF_DIRECT_WAIT         0x0400  /* related direct I/O running*/
 419
 420 /*
 421  * hammer_delete_at_cursor() flags
 422  */
 423 #define HAMMER_DELETE_ADJUST            0x0001
 424 #define HAMMER_DELETE_DESTROY           0x0002
 425
 426 /*
 427  * In-memory structures representing on-disk structures.
 428  */
 429 struct hammer_volume;
 430 struct hammer_buffer;
 431 struct hammer_node;
 432 struct hammer_undo;
 433 struct hammer_reserve;
 434
 435 RB_HEAD(hammer_vol_rb_tree, hammer_volume);
 436 RB_HEAD(hammer_buf_rb_tree, hammer_buffer);
 437 RB_HEAD(hammer_nod_rb_tree, hammer_node);
 438 RB_HEAD(hammer_und_rb_tree, hammer_undo);
 439 RB_HEAD(hammer_res_rb_tree, hammer_reserve);
 440
 441 RB_PROTOTYPE2(hammer_vol_rb_tree, hammer_volume, rb_node,
 442               hammer_vol_rb_compare, int32_t);
 443 RB_PROTOTYPE2(hammer_buf_rb_tree, hammer_buffer, rb_node,
 444               hammer_buf_rb_compare, hammer_off_t);
 445 RB_PROTOTYPE2(hammer_nod_rb_tree, hammer_node, rb_node,
 446               hammer_nod_rb_compare, hammer_off_t);
 447 RB_PROTOTYPE2(hammer_und_rb_tree, hammer_undo, rb_node,
 448               hammer_und_rb_compare, hammer_off_t);
 449 RB_PROTOTYPE2(hammer_res_rb_tree, hammer_reserve, rb_node,
 450               hammer_res_rb_compare, hammer_off_t);
 451
 452 /*
 453  * IO management - embedded at the head of various in-memory structures
 454  *
 455  * VOLUME       - hammer_volume containing meta-data
 456  * META_BUFFER  - hammer_buffer containing meta-data
 457  * DATA_BUFFER  - hammer_buffer containing pure-data
 458  *
 459  * Dirty volume headers and dirty meta-data buffers are locked until the
 460  * flusher can sequence them out.  Dirty pure-data buffers can be written.
 461  * Clean buffers can be passively released.
 462  */
 463 typedef enum hammer_io_type {
 464         HAMMER_STRUCTURE_VOLUME,
 465         HAMMER_STRUCTURE_META_BUFFER,
 466         HAMMER_STRUCTURE_UNDO_BUFFER,
 467         HAMMER_STRUCTURE_DATA_BUFFER
 468 } hammer_io_type_t;
 469
 470 union hammer_io_structure;
 471 struct hammer_io;
 472
 473 struct worklist {
 474         LIST_ENTRY(worklist) node;
 475 };
 476
 477 TAILQ_HEAD(hammer_io_list, hammer_io);
 478 typedef struct hammer_io_list *hammer_io_list_t;
 479
 480 struct hammer_io {
 481         struct worklist         worklist;
 482         struct hammer_lock      lock;
 483         enum hammer_io_type     type;
 484         struct hammer_mount     *hmp;
 485         TAILQ_ENTRY(hammer_io)  mod_entry; /* list entry if modified */
 486         hammer_io_list_t        mod_list;
 487         struct buf              *bp;
 488         int64_t                 offset;    /* zone-2 offset */
 489         int                     bytes;     /* buffer cache buffer size */
 490         int                     loading;   /* loading/unloading interlock */
 491         int                     modify_refs;
 492
 493         u_int           modified : 1;   /* bp's data was modified */
 494         u_int           released : 1;   /* bp released (w/ B_LOCKED set) */
 495         u_int           running : 1;    /* bp write IO in progress */
 496         u_int           waiting : 1;    /* someone is waiting on us */
 497         u_int           validated : 1;  /* ondisk has been validated */
 498         u_int           waitdep : 1;    /* flush waits for dependancies */
 499         u_int           recovered : 1;  /* has recovery ref */
 500         u_int           waitmod : 1;    /* waiting for modify_refs */
 501         u_int           reclaim : 1;    /* reclaim requested */
 502         u_int           gencrc : 1;     /* crc needs to be generated */
 503         u_int           ioerror : 1;    /* abort on io-error */
 504 };
 505
 506 typedef struct hammer_io *hammer_io_t;
 507
 508 #define HAMMER_CLUSTER_SIZE     (64 * 1024)
 509 #if HAMMER_CLUSTER_SIZE > MAXBSIZE
 510 #undef  HAMMER_CLUSTER_SIZE
 511 #define HAMMER_CLUSTER_SIZE     MAXBSIZE
 512 #endif
 513 #define HAMMER_CLUSTER_BUFS     (HAMMER_CLUSTER_SIZE / HAMMER_BUFSIZE)
 514
 515 /*
 516  * In-memory volume representing on-disk buffer
 517  */
 518 struct hammer_volume {
 519         struct hammer_io io;
 520         RB_ENTRY(hammer_volume) rb_node;
 521         struct hammer_volume_ondisk *ondisk;
 522         int32_t vol_no;
 523         int64_t nblocks;        /* note: special calculation for statfs */
 524         int64_t buffer_base;    /* base offset of buffer 0 */
 525         hammer_off_t maxbuf_off; /* Maximum buffer offset (zone-2) */
 526         hammer_off_t maxraw_off; /* Maximum raw offset for device */
 527         char    *vol_name;
 528         struct vnode *devvp;
 529         int     vol_flags;
 530 };
 531
 532 typedef struct hammer_volume *hammer_volume_t;
 533
 534 /*
 535  * In-memory buffer (other then volume, super-cluster, or cluster),
 536  * representing an on-disk buffer.
 537  */
 538 struct hammer_buffer {
 539         struct hammer_io io;
 540         RB_ENTRY(hammer_buffer) rb_node;
 541         void *ondisk;
 542         struct hammer_volume *volume;
 543         hammer_off_t zoneX_offset;
 544         hammer_off_t zone2_offset;
 545         struct hammer_reserve *resv;
 546         struct hammer_node_list clist;
 547 };
 548
 549 typedef struct hammer_buffer *hammer_buffer_t;
 550
 551 /*
 552  * In-memory B-Tree node, representing an on-disk B-Tree node.
 553  *
 554  * This is a hang-on structure which is backed by a hammer_buffer,
 555  * indexed by a hammer_cluster, and used for fine-grained locking of
 556  * B-Tree nodes in order to properly control lock ordering.  A hammer_buffer
 557  * can contain multiple nodes representing wildly disassociated portions
 558  * of the B-Tree so locking cannot be done on a buffer-by-buffer basis.
 559  *
 560  * This structure uses a cluster-relative index to reduce the number
 561  * of layers required to access it, and also because all on-disk B-Tree
 562  * references are cluster-relative offsets.
 563  */
 564 struct hammer_node {
 565         struct hammer_lock      lock;           /* node-by-node lock */
 566         TAILQ_ENTRY(hammer_node) entry;         /* per-buffer linkage */
 567         RB_ENTRY(hammer_node)   rb_node;        /* per-cluster linkage */
 568         hammer_off_t            node_offset;    /* full offset spec */
 569         struct hammer_mount     *hmp;
 570         struct hammer_buffer    *buffer;        /* backing buffer */
 571         hammer_node_ondisk_t    ondisk;         /* ptr to on-disk structure */
 572         TAILQ_HEAD(, hammer_cursor) cursor_list;  /* deadlock recovery */
 573         struct hammer_node_cache_list cache_list; /* passive caches */
 574         int                     flags;
 575         int                     loading;        /* load interlock */
 576 };
 577
 578 #define HAMMER_NODE_DELETED     0x0001
 579 #define HAMMER_NODE_FLUSH       0x0002
 580 #define HAMMER_NODE_CRCGOOD     0x0004
 581 #define HAMMER_NODE_NEEDSCRC    0x0008
 582 #define HAMMER_NODE_NEEDSMIRROR 0x0010
 583
 584 typedef struct hammer_node      *hammer_node_t;
 585
 586 /*
 587  * List of locked nodes.
 588  */
 589 struct hammer_node_locklist {
 590         struct hammer_node_locklist *next;
 591         hammer_node_t   node;
 592 };
 593
 594 typedef struct hammer_node_locklist *hammer_node_locklist_t;
 595
 596
 597 /*
 598  * Common I/O management structure - embedded in in-memory structures
 599  * which are backed by filesystem buffers.
 600  */
 601 union hammer_io_structure {
 602         struct hammer_io        io;
 603         struct hammer_volume    volume;
 604         struct hammer_buffer    buffer;
 605 };
 606
 607 typedef union hammer_io_structure *hammer_io_structure_t;
 608
 609 /*
 610  * The reserve structure prevents the blockmap from allocating
 611  * out of a reserved bigblock.  Such reservations are used by
 612  * the direct-write mechanism.
 613  *
 614  * The structure is also used to hold off on reallocations of
 615  * big blocks from the freemap until flush dependancies have
 616  * been dealt with.
 617  */
 618 struct hammer_reserve {
 619         RB_ENTRY(hammer_reserve) rb_node;
 620         TAILQ_ENTRY(hammer_reserve) delay_entry;
 621         int             flush_group;
 622         int             flags;
 623         int             refs;
 624         int             zone;
 625         int             bytes_freed;
 626         int             append_off;
 627         hammer_off_t    zone_offset;
 628 };
 629
 630 typedef struct hammer_reserve *hammer_reserve_t;
 631
 632 #define HAMMER_RESF_ONDELAY     0x0001
 633
 634 #include "hammer_cursor.h"
 635
 636 /*
 637  * The undo structure tracks recent undos to avoid laying down duplicate
 638  * undos within a flush group, saving us a significant amount of overhead.
 639  *
 640  * This is strictly a heuristic.
 641  */
 642 #define HAMMER_MAX_UNDOS                1024
 643 #define HAMMER_MAX_FLUSHERS             4
 644
 645 struct hammer_undo {
 646         RB_ENTRY(hammer_undo)   rb_node;
 647         TAILQ_ENTRY(hammer_undo) lru_entry;
 648         hammer_off_t            offset;
 649         int                     bytes;
 650 };
 651
 652 typedef struct hammer_undo *hammer_undo_t;
 653
 654 struct hammer_flusher_info;
 655 TAILQ_HEAD(hammer_flusher_info_list, hammer_flusher_info);
 656
 657 struct hammer_flusher {
 658         int             signal;         /* flusher thread sequencer */
 659         int             act;            /* currently active flush group */
 660         int             done;           /* set to act when complete */
 661         int             next;           /* next flush group */
 662         int             group_lock;     /* lock sequencing of the next flush */
 663         int             exiting;        /* request master exit */
 664         thread_t        td;             /* master flusher thread */
 665         hammer_tid_t    tid;            /* last flushed transaction id */
 666         int             finalize_want;          /* serialize finalization */
 667         struct hammer_lock finalize_lock;       /* serialize finalization */
 668         struct hammer_transaction trans;        /* shared transaction */
 669         struct hammer_flusher_info_list run_list;
 670         struct hammer_flusher_info_list ready_list;
 671 };
 672
 673 /*
 674  * Internal hammer mount data structure
 675  */
 676 struct hammer_mount {
 677         struct mount *mp;
 678         /*struct vnode *rootvp;*/
 679         struct hammer_ino_rb_tree rb_inos_root;
 680         struct hammer_vol_rb_tree rb_vols_root;
 681         struct hammer_nod_rb_tree rb_nods_root;
 682         struct hammer_und_rb_tree rb_undo_root;
 683         struct hammer_res_rb_tree rb_resv_root;
 684         struct hammer_buf_rb_tree rb_bufs_root;
 685         struct hammer_pfs_rb_tree rb_pfsm_root;
 686         struct hammer_volume *rootvol;
 687         struct hammer_base_elm root_btree_beg;
 688         struct hammer_base_elm root_btree_end;
 689         int     flags;          /* HAMMER_MOUNT_xxx flags */
 690         int     hflags;
 691         int     ronly;
 692         int     nvolumes;
 693         int     volume_iterator;
 694         int     master_id;      /* -1 or 0-15 - clustering and mirroring */
 695         int     rsv_inodes;     /* reserved space due to dirty inodes */
 696         int64_t rsv_databytes;  /* reserved space due to record data */
 697         int     rsv_recs;       /* reserved space due to dirty records */
 698         int     rsv_fromdelay;  /* bigblocks reserved due to flush delay */
 699         int     undo_rec_limit; /* based on size of undo area */
 700         int     last_newrecords;
 701         int     count_newrecords;
 702
 703         int     inode_reclaims; /* inodes pending reclaim by flusher */
 704         int     count_inodes;   /* total number of inodes */
 705         int     count_iqueued;  /* inodes queued to flusher */
 706
 707         struct hammer_flusher flusher;
 708
 709         u_int   check_interrupt;
 710         uuid_t  fsid;
 711         struct hammer_io_list volu_list;        /* dirty undo buffers */
 712         struct hammer_io_list undo_list;        /* dirty undo buffers */
 713         struct hammer_io_list data_list;        /* dirty data buffers */
 714         struct hammer_io_list alt_data_list;    /* dirty data buffers */
 715         struct hammer_io_list meta_list;        /* dirty meta bufs    */
 716         struct hammer_io_list lose_list;        /* loose buffers      */
 717         int     locked_dirty_space;             /* meta/volu count    */
 718         int     io_running_space;
 719         int     objid_cache_count;
 720         int     error;                          /* critical I/O error */
 721         struct krate    krate;                  /* rate limited kprintf */
 722         hammer_tid_t    asof;                   /* snapshot mount */
 723         hammer_tid_t    next_tid;
 724         hammer_tid_t    flush_tid1;             /* flusher tid sequencing */
 725         hammer_tid_t    flush_tid2;             /* flusher tid sequencing */
 726         int64_t copy_stat_freebigblocks;        /* number of free bigblocks */
 727
 728         u_int32_t namekey_iterator;
 729         struct netexport export;
 730         struct hammer_lock sync_lock;
 731         struct hammer_lock free_lock;
 732         struct hammer_lock undo_lock;
 733         struct hammer_lock blkmap_lock;
 734         struct hammer_blockmap  blockmap[HAMMER_MAX_ZONES];
 735         struct hammer_undo      undos[HAMMER_MAX_UNDOS];
 736         int                     undo_alloc;
 737         TAILQ_HEAD(, hammer_undo)  undo_lru_list;
 738         TAILQ_HEAD(, hammer_reserve) delay_list;
 739         struct hammer_flush_group_list  flush_group_list;
 740         hammer_flush_group_t    next_flush_group;
 741         TAILQ_HEAD(, hammer_objid_cache) objid_cache_list;
 742         TAILQ_HEAD(, hammer_reclaim) reclaim_list;
 743 };
 744
 745 typedef struct hammer_mount     *hammer_mount_t;
 746
 747 #define HAMMER_MOUNT_CRITICAL_ERROR     0x0001
 748
 749 struct hammer_sync_info {
 750         int error;
 751         int waitfor;
 752 };
 753
 754 #endif
 755
 756 /*
 757  * checkspace slop (8MB chunks), higher numbers are more conservative.
 758  */
 759 #define HAMMER_CHKSPC_REBLOCK   25
 760 #define HAMMER_CHKSPC_MIRROR    20
 761 #define HAMMER_CHKSPC_WRITE     20
 762 #define HAMMER_CHKSPC_CREATE    20
 763 #define HAMMER_CHKSPC_REMOVE    10
 764 #define HAMMER_CHKSPC_EMERGENCY 0
 765
 766 #if defined(_KERNEL)
 767
 768 extern struct vop_ops hammer_vnode_vops;
 769 extern struct vop_ops hammer_spec_vops;
 770 extern struct vop_ops hammer_fifo_vops;
 771 extern struct bio_ops hammer_bioops;
 772
 773 extern int hammer_debug_io;
 774 extern int hammer_debug_general;
 775 extern int hammer_debug_debug;
 776 extern int hammer_debug_inode;
 777 extern int hammer_debug_locks;
 778 extern int hammer_debug_btree;
 779 extern int hammer_debug_tid;
 780 extern int hammer_debug_recover;
 781 extern int hammer_debug_recover_faults;
 782 extern int hammer_cluster_enable;
 783 extern int hammer_count_fsyncs;
 784 extern int hammer_count_inodes;
 785 extern int hammer_count_iqueued;
 786 extern int hammer_count_reclaiming;
 787 extern int hammer_count_records;
 788 extern int hammer_count_record_datas;
 789 extern int hammer_count_volumes;
 790 extern int hammer_count_buffers;
 791 extern int hammer_count_nodes;
 792 extern int64_t hammer_count_extra_space_used;
 793 extern int64_t hammer_stats_btree_lookups;
 794 extern int64_t hammer_stats_btree_searches;
 795 extern int64_t hammer_stats_btree_inserts;
 796 extern int64_t hammer_stats_btree_deletes;
 797 extern int64_t hammer_stats_btree_elements;
 798 extern int64_t hammer_stats_btree_splits;
 799 extern int64_t hammer_stats_btree_iterations;
 800 extern int64_t hammer_stats_record_iterations;
 801 extern int64_t hammer_stats_file_read;
 802 extern int64_t hammer_stats_file_write;
 803 extern int64_t hammer_stats_file_iopsr;
 804 extern int64_t hammer_stats_file_iopsw;
 805 extern int64_t hammer_stats_disk_read;
 806 extern int64_t hammer_stats_disk_write;
 807 extern int64_t hammer_stats_inode_flushes;
 808 extern int64_t hammer_stats_commits;
 809 extern int hammer_count_dirtybufspace;
 810 extern int hammer_count_refedbufs;
 811 extern int hammer_count_reservations;
 812 extern int hammer_count_io_running_read;
 813 extern int hammer_count_io_running_write;
 814 extern int hammer_count_io_locked;
 815 extern int hammer_limit_dirtybufspace;
 816 extern int hammer_limit_iqueued;
 817 extern int hammer_limit_recs;
 818 extern int hammer_bio_count;
 819 extern int hammer_verify_zone;
 820 extern int hammer_verify_data;
 821 extern int hammer_write_mode;
 822 extern int64_t hammer_contention_count;
 823
 824 void    hammer_critical_error(hammer_mount_t hmp, hammer_inode_t ip,
 825                         int error, const char *msg);
 826 int     hammer_vop_inactive(struct vop_inactive_args *);
 827 int     hammer_vop_reclaim(struct vop_reclaim_args *);
 828 int     hammer_get_vnode(struct hammer_inode *ip, struct vnode **vpp);
 829 struct hammer_inode *hammer_get_inode(hammer_transaction_t trans,
 830                         hammer_inode_t dip, int64_t obj_id,
 831                         hammer_tid_t asof, u_int32_t localization,
 832                         int flags, int *errorp);
 833 void    hammer_scan_inode_snapshots(hammer_mount_t hmp,
 834                         hammer_inode_info_t iinfo,
 835                         int (*callback)(hammer_inode_t ip, void *data),
 836                         void *data);
 837 void    hammer_put_inode(struct hammer_inode *ip);
 838 void    hammer_put_inode_ref(struct hammer_inode *ip);
 839 void    hammer_inode_waitreclaims(hammer_mount_t hmp);
 840
 841 int     hammer_unload_volume(hammer_volume_t volume, void *data __unused);
 842 int     hammer_adjust_volume_mode(hammer_volume_t volume, void *data __unused);
 843
 844 int     hammer_unload_buffer(hammer_buffer_t buffer, void *data __unused);
 845 int     hammer_install_volume(hammer_mount_t hmp, const char *volname,
 846                         struct vnode *devvp);
 847 int     hammer_mountcheck_volumes(hammer_mount_t hmp);
 848
 849 int     hammer_ip_lookup(hammer_cursor_t cursor);
 850 int     hammer_ip_first(hammer_cursor_t cursor);
 851 int     hammer_ip_next(hammer_cursor_t cursor);
 852 int     hammer_ip_resolve_data(hammer_cursor_t cursor);
 853 int     hammer_ip_delete_record(hammer_cursor_t cursor, hammer_inode_t ip,
 854                         hammer_tid_t tid);
 855 int     hammer_delete_at_cursor(hammer_cursor_t cursor, int delete_flags,
 856                         hammer_tid_t delete_tid, u_int32_t delete_ts,
 857                         int track, int64_t *stat_bytes);
 858 int     hammer_ip_check_directory_empty(hammer_transaction_t trans,
 859                         hammer_inode_t ip);
 860 int     hammer_sync_hmp(hammer_mount_t hmp, int waitfor);
 861 int     hammer_queue_inodes_flusher(hammer_mount_t hmp, int waitfor);
 862
 863 hammer_record_t
 864         hammer_alloc_mem_record(hammer_inode_t ip, int data_len);
 865 void    hammer_flush_record_done(hammer_record_t record, int error);
 866 void    hammer_wait_mem_record_ident(hammer_record_t record, const char *ident);
 867 void    hammer_rel_mem_record(hammer_record_t record);
 868
 869 int     hammer_cursor_up(hammer_cursor_t cursor);
 870 int     hammer_cursor_up_locked(hammer_cursor_t cursor);
 871 int     hammer_cursor_down(hammer_cursor_t cursor);
 872 int     hammer_cursor_upgrade(hammer_cursor_t cursor);
 873 int     hammer_cursor_upgrade_node(hammer_cursor_t cursor);
 874 void    hammer_cursor_downgrade(hammer_cursor_t cursor);
 875 int     hammer_cursor_seek(hammer_cursor_t cursor, hammer_node_t node,
 876                         int index);
 877 void    hammer_lock_ex_ident(struct hammer_lock *lock, const char *ident);
 878 int     hammer_lock_ex_try(struct hammer_lock *lock);
 879 void    hammer_lock_sh(struct hammer_lock *lock);
 880 int     hammer_lock_sh_try(struct hammer_lock *lock);
 881 int     hammer_lock_upgrade(struct hammer_lock *lock);
 882 void    hammer_lock_downgrade(struct hammer_lock *lock);
 883 int     hammer_lock_status(struct hammer_lock *lock);
 884 void    hammer_unlock(struct hammer_lock *lock);
 885 void    hammer_ref(struct hammer_lock *lock);
 886 void    hammer_unref(struct hammer_lock *lock);
 887
 888 void    hammer_sync_lock_ex(hammer_transaction_t trans);
 889 void    hammer_sync_lock_sh(hammer_transaction_t trans);
 890 int     hammer_sync_lock_sh_try(hammer_transaction_t trans);
 891 void    hammer_sync_unlock(hammer_transaction_t trans);
 892
 893 u_int32_t hammer_to_unix_xid(uuid_t *uuid);
 894 void hammer_guid_to_uuid(uuid_t *uuid, u_int32_t guid);
 895 void    hammer_time_to_timespec(u_int64_t xtime, struct timespec *ts);
 896 u_int64_t hammer_timespec_to_time(struct timespec *ts);
 897 hammer_tid_t hammer_str_to_tid(const char *str, int *ispfs,
 898                         u_int32_t *localizationp);
 899 hammer_tid_t hammer_alloc_objid(hammer_mount_t hmp, hammer_inode_t dip);
 900 void hammer_clear_objid(hammer_inode_t dip);
 901 void hammer_destroy_objid_cache(hammer_mount_t hmp);
 902
 903 int hammer_enter_undo_history(hammer_mount_t hmp, hammer_off_t offset,
 904                         int bytes);
 905 void hammer_clear_undo_history(hammer_mount_t hmp);
 906 enum vtype hammer_get_vnode_type(u_int8_t obj_type);
 907 int hammer_get_dtype(u_int8_t obj_type);
 908 u_int8_t hammer_get_obj_type(enum vtype vtype);
 909 int64_t hammer_directory_namekey(const void *name, int len);
 910 int     hammer_nohistory(hammer_inode_t ip);
 911
 912 int     hammer_init_cursor(hammer_transaction_t trans, hammer_cursor_t cursor,
 913                         hammer_node_cache_t cache, hammer_inode_t ip);
 914 void    hammer_normalize_cursor(hammer_cursor_t cursor);
 915 void    hammer_done_cursor(hammer_cursor_t cursor);
 916 int     hammer_recover_cursor(hammer_cursor_t cursor);
 917 void    hammer_unlock_cursor(hammer_cursor_t cursor, int also_ip);
 918 int     hammer_lock_cursor(hammer_cursor_t cursor, int also_ip);
 919 hammer_cursor_t hammer_push_cursor(hammer_cursor_t ocursor);
 920 void    hammer_pop_cursor(hammer_cursor_t ocursor, hammer_cursor_t ncursor);
 921
 922 void    hammer_cursor_replaced_node(hammer_node_t onode, hammer_node_t nnode);
 923 void    hammer_cursor_removed_node(hammer_node_t onode, hammer_node_t parent,
 924                         int index);
 925 void    hammer_cursor_split_node(hammer_node_t onode, hammer_node_t nnode,
 926                         int index);
 927 void    hammer_cursor_inserted_element(hammer_node_t node, int index);
 928 void    hammer_cursor_deleted_element(hammer_node_t node, int index);
 929
 930 int     hammer_btree_lookup(hammer_cursor_t cursor);
 931 int     hammer_btree_first(hammer_cursor_t cursor);
 932 int     hammer_btree_last(hammer_cursor_t cursor);
 933 int     hammer_btree_extract(hammer_cursor_t cursor, int flags);
 934 int     hammer_btree_iterate(hammer_cursor_t cursor);
 935 int     hammer_btree_iterate_reverse(hammer_cursor_t cursor);
 936 int     hammer_btree_insert(hammer_cursor_t cursor,
 937                             hammer_btree_leaf_elm_t elm, int *doprop);
 938 int     hammer_btree_delete(hammer_cursor_t cursor);
 939 void    hammer_btree_do_propagation(hammer_cursor_t cursor,
 940                             hammer_pseudofs_inmem_t pfsm,
 941                             hammer_btree_leaf_elm_t leaf);
 942 int     hammer_btree_cmp(hammer_base_elm_t key1, hammer_base_elm_t key2);
 943 int     hammer_btree_chkts(hammer_tid_t ts, hammer_base_elm_t key);
 944 int     hammer_btree_correct_rhb(hammer_cursor_t cursor, hammer_tid_t tid);
 945 int     hammer_btree_correct_lhb(hammer_cursor_t cursor, hammer_tid_t tid);
 946
 947 int     btree_set_parent(hammer_transaction_t trans, hammer_node_t node,
 948                         hammer_btree_elm_t elm);
 949 int     hammer_btree_lock_children(hammer_cursor_t cursor,
 950                         struct hammer_node_locklist **locklistp);
 951 void    hammer_btree_unlock_children(struct hammer_node_locklist **locklistp);
 952 int     hammer_btree_search_node(hammer_base_elm_t elm, hammer_node_ondisk_t node);
 953 hammer_node_t hammer_btree_get_parent(hammer_node_t node, int *parent_indexp,
 954                         int *errorp, int try_exclusive);
 955
 956 void    hammer_print_btree_node(hammer_node_ondisk_t ondisk);
 957 void    hammer_print_btree_elm(hammer_btree_elm_t elm, u_int8_t type, int i);
 958
 959 void    *hammer_bread(struct hammer_mount *hmp, hammer_off_t off,
 960                         int *errorp, struct hammer_buffer **bufferp);
 961 void    *hammer_bnew(struct hammer_mount *hmp, hammer_off_t off,
 962                         int *errorp, struct hammer_buffer **bufferp);
 963 void    *hammer_bread_ext(struct hammer_mount *hmp, hammer_off_t off, int bytes,
 964                         int *errorp, struct hammer_buffer **bufferp);
 965 void    *hammer_bnew_ext(struct hammer_mount *hmp, hammer_off_t off, int bytes,
 966                         int *errorp, struct hammer_buffer **bufferp);
 967
 968 hammer_volume_t hammer_get_root_volume(hammer_mount_t hmp, int *errorp);
 969
 970 hammer_volume_t hammer_get_volume(hammer_mount_t hmp,
 971                         int32_t vol_no, int *errorp);
 972 hammer_buffer_t hammer_get_buffer(hammer_mount_t hmp, hammer_off_t buf_offset,
 973                         int bytes, int isnew, int *errorp);
 974 void            hammer_sync_buffers(hammer_mount_t hmp,
 975                         hammer_off_t base_offset, int bytes);
 976 void            hammer_del_buffers(hammer_mount_t hmp,
 977                         hammer_off_t base_offset,
 978                         hammer_off_t zone2_offset, int bytes);
 979
 980 int             hammer_ref_volume(hammer_volume_t volume);
 981 int             hammer_ref_buffer(hammer_buffer_t buffer);
 982 void            hammer_flush_buffer_nodes(hammer_buffer_t buffer);
 983
 984 void            hammer_rel_volume(hammer_volume_t volume, int flush);
 985 void            hammer_rel_buffer(hammer_buffer_t buffer, int flush);
 986
 987 int             hammer_vfs_export(struct mount *mp, int op,
 988                         const struct export_args *export);
 989 hammer_node_t   hammer_get_node(hammer_mount_t hmp, hammer_off_t node_offset,
 990                         int isnew, int *errorp);
 991 void            hammer_ref_node(hammer_node_t node);
 992 hammer_node_t   hammer_ref_node_safe(struct hammer_mount *hmp,
 993                         hammer_node_cache_t cache, int *errorp);
 994 void            hammer_rel_node(hammer_node_t node);
 995 void            hammer_delete_node(hammer_transaction_t trans,
 996                         hammer_node_t node);
 997 void            hammer_cache_node(hammer_node_cache_t cache,
 998                         hammer_node_t node);
 999 void            hammer_uncache_node(hammer_node_cache_t cache);
1000 void            hammer_flush_node(hammer_node_t node);
1001
1002 void hammer_dup_buffer(struct hammer_buffer **bufferp,
1003                         struct hammer_buffer *buffer);
1004 hammer_node_t hammer_alloc_btree(hammer_transaction_t trans, int *errorp);
1005 void *hammer_alloc_data(hammer_transaction_t trans, int32_t data_len,
1006                         u_int16_t rec_type, hammer_off_t *data_offsetp,
1007                         struct hammer_buffer **data_bufferp, int *errorp);
1008
1009 int hammer_generate_undo(hammer_transaction_t trans, hammer_io_t io,
1010                         hammer_off_t zone1_offset, void *base, int len);
1011
1012 void hammer_put_volume(struct hammer_volume *volume, int flush);
1013 void hammer_put_buffer(struct hammer_buffer *buffer, int flush);
1014
1015 hammer_off_t hammer_freemap_alloc(hammer_transaction_t trans,
1016                         hammer_off_t owner, int *errorp);
1017 void hammer_freemap_free(hammer_transaction_t trans, hammer_off_t phys_offset,
1018                         hammer_off_t owner, int *errorp);
1019 int hammer_checkspace(hammer_mount_t hmp, int slop);
1020 hammer_off_t hammer_blockmap_alloc(hammer_transaction_t trans, int zone,
1021                         int bytes, int *errorp);
1022 hammer_reserve_t hammer_blockmap_reserve(hammer_mount_t hmp, int zone,
1023                         int bytes, hammer_off_t *zone_offp, int *errorp);
1024 void hammer_blockmap_reserve_undo(hammer_reserve_t resv,
1025                         hammer_off_t zone_offset, int bytes);
1026 void hammer_blockmap_reserve_complete(hammer_mount_t hmp,
1027                         hammer_reserve_t resv);
1028 void hammer_reserve_clrdelay(hammer_mount_t hmp, hammer_reserve_t resv);
1029 void hammer_blockmap_free(hammer_transaction_t trans,
1030                         hammer_off_t bmap_off, int bytes);
1031 int hammer_blockmap_finalize(hammer_transaction_t trans,
1032                         hammer_off_t bmap_off, int bytes);
1033 int hammer_blockmap_getfree(hammer_mount_t hmp, hammer_off_t bmap_off,
1034                         int *curp, int *errorp);
1035 hammer_off_t hammer_blockmap_lookup(hammer_mount_t hmp, hammer_off_t bmap_off,
1036                         int *errorp);
1037 hammer_off_t hammer_undo_lookup(hammer_mount_t hmp, hammer_off_t bmap_off,
1038                         int *errorp);
1039 int64_t hammer_undo_used(hammer_transaction_t trans);
1040 int64_t hammer_undo_space(hammer_transaction_t trans);
1041 int64_t hammer_undo_max(hammer_mount_t hmp);
1042
1043 void hammer_start_transaction(struct hammer_transaction *trans,
1044                               struct hammer_mount *hmp);
1045 void hammer_simple_transaction(struct hammer_transaction *trans,
1046                               struct hammer_mount *hmp);
1047 void hammer_start_transaction_fls(struct hammer_transaction *trans,
1048                                   struct hammer_mount *hmp);
1049 void hammer_done_transaction(struct hammer_transaction *trans);
1050
1051 void hammer_modify_inode(hammer_inode_t ip, int flags);
1052 void hammer_flush_inode(hammer_inode_t ip, int flags);
1053 void hammer_flush_inode_done(hammer_inode_t ip, int error);
1054 void hammer_wait_inode(hammer_inode_t ip);
1055
1056 int  hammer_create_inode(struct hammer_transaction *trans, struct vattr *vap,
1057                         struct ucred *cred, struct hammer_inode *dip,
1058                         hammer_pseudofs_inmem_t pfsm,
1059                         struct hammer_inode **ipp);
1060 void hammer_rel_inode(hammer_inode_t ip, int flush);
1061 int hammer_reload_inode(hammer_inode_t ip, void *arg __unused);
1062 int hammer_ino_rb_compare(hammer_inode_t ip1, hammer_inode_t ip2);
1063 int hammer_destroy_inode_callback(hammer_inode_t ip, void *data __unused);
1064
1065 int hammer_sync_inode(hammer_transaction_t trans, hammer_inode_t ip);
1066 void hammer_test_inode(hammer_inode_t dip);
1067 void hammer_inode_unloadable_check(hammer_inode_t ip, int getvp);
1068
1069 int  hammer_ip_add_directory(struct hammer_transaction *trans,
1070                         hammer_inode_t dip, const char *name, int bytes,
1071                         hammer_inode_t nip);
1072 int  hammer_ip_del_directory(struct hammer_transaction *trans,
1073                         hammer_cursor_t cursor, hammer_inode_t dip,
1074                         hammer_inode_t ip);
1075 hammer_record_t hammer_ip_add_bulk(hammer_inode_t ip, off_t file_offset,
1076                         void *data, int bytes, int *errorp);
1077 int  hammer_ip_frontend_trunc(struct hammer_inode *ip, off_t file_size);
1078 int  hammer_ip_add_record(struct hammer_transaction *trans,
1079                         hammer_record_t record);
1080 int  hammer_ip_delete_range(hammer_cursor_t cursor, hammer_inode_t ip,
1081                         int64_t ran_beg, int64_t ran_end, int truncating);
1082 int  hammer_ip_delete_clean(hammer_cursor_t cursor, hammer_inode_t ip,
1083                         int *countp);
1084 int  hammer_ip_sync_data(hammer_cursor_t cursor, hammer_inode_t ip,
1085                         int64_t offset, void *data, int bytes);
1086 int  hammer_ip_sync_record(hammer_transaction_t trans, hammer_record_t rec);
1087 int  hammer_ip_sync_record_cursor(hammer_cursor_t cursor, hammer_record_t rec);
1088 hammer_pseudofs_inmem_t  hammer_load_pseudofs(hammer_transaction_t trans,
1089                         u_int32_t localization, int *errorp);
1090 int  hammer_mkroot_pseudofs(hammer_transaction_t trans, struct ucred *cred,
1091                         hammer_pseudofs_inmem_t pfsm);
1092 int  hammer_save_pseudofs(hammer_transaction_t trans,
1093                         hammer_pseudofs_inmem_t pfsm);
1094 int  hammer_unload_pseudofs(hammer_transaction_t trans, u_int32_t localization);
1095 void hammer_rel_pseudofs(hammer_mount_t hmp, hammer_pseudofs_inmem_t pfsm);
1096 int hammer_ioctl(hammer_inode_t ip, u_long com, caddr_t data, int fflag,
1097                         struct ucred *cred);
1098
1099 void hammer_io_init(hammer_io_t io, hammer_mount_t hmp,
1100                         enum hammer_io_type type);
1101 int hammer_io_read(struct vnode *devvp, struct hammer_io *io,
1102                         hammer_off_t limit);
1103 int hammer_io_new(struct vnode *devvp, struct hammer_io *io);
1104 void hammer_io_inval(hammer_volume_t volume, hammer_off_t zone2_offset);
1105 struct buf *hammer_io_release(struct hammer_io *io, int flush);
1106 void hammer_io_flush(struct hammer_io *io);
1107 void hammer_io_wait(struct hammer_io *io);
1108 void hammer_io_waitdep(struct hammer_io *io);
1109 void hammer_io_wait_all(hammer_mount_t hmp, const char *ident);
1110 int hammer_io_direct_read(hammer_mount_t hmp, struct bio *bio,
1111                         hammer_btree_leaf_elm_t leaf);
1112 int hammer_io_direct_write(hammer_mount_t hmp, hammer_record_t record,
1113                         struct bio *bio);
1114 void hammer_io_direct_wait(hammer_record_t record);
1115 void hammer_io_direct_uncache(hammer_mount_t hmp, hammer_btree_leaf_elm_t leaf);
1116 void hammer_io_write_interlock(hammer_io_t io);
1117 void hammer_io_done_interlock(hammer_io_t io);
1118 void hammer_io_clear_modify(struct hammer_io *io, int inval);
1119 void hammer_io_clear_modlist(struct hammer_io *io);
1120 void hammer_modify_volume(hammer_transaction_t trans, hammer_volume_t volume,
1121                         void *base, int len);
1122 void hammer_modify_buffer(hammer_transaction_t trans, hammer_buffer_t buffer,
1123                         void *base, int len);
1124 void hammer_modify_volume_done(hammer_volume_t volume);
1125 void hammer_modify_buffer_done(hammer_buffer_t buffer);
1126
1127 int hammer_ioc_reblock(hammer_transaction_t trans, hammer_inode_t ip,
1128                         struct hammer_ioc_reblock *reblock);
1129 int hammer_ioc_prune(hammer_transaction_t trans, hammer_inode_t ip,
1130                         struct hammer_ioc_prune *prune);
1131 int hammer_ioc_mirror_read(hammer_transaction_t trans, hammer_inode_t ip,
1132                         struct hammer_ioc_mirror_rw *mirror);
1133 int hammer_ioc_mirror_write(hammer_transaction_t trans, hammer_inode_t ip,
1134                         struct hammer_ioc_mirror_rw *mirror);
1135 int hammer_ioc_set_pseudofs(hammer_transaction_t trans, hammer_inode_t ip,
1136                         struct ucred *cred, struct hammer_ioc_pseudofs_rw *pfs);
1137 int hammer_ioc_get_pseudofs(hammer_transaction_t trans, hammer_inode_t ip,
1138                         struct hammer_ioc_pseudofs_rw *pfs);
1139 int hammer_ioc_destroy_pseudofs(hammer_transaction_t trans, hammer_inode_t ip,
1140                         struct hammer_ioc_pseudofs_rw *pfs);
1141 int hammer_ioc_downgrade_pseudofs(hammer_transaction_t trans, hammer_inode_t ip,
1142                         struct hammer_ioc_pseudofs_rw *pfs);
1143 int hammer_ioc_upgrade_pseudofs(hammer_transaction_t trans, hammer_inode_t ip,
1144                         struct hammer_ioc_pseudofs_rw *pfs);
1145 int hammer_ioc_wait_pseudofs(hammer_transaction_t trans, hammer_inode_t ip,
1146                         struct hammer_ioc_pseudofs_rw *pfs);
1147
1148 int hammer_signal_check(hammer_mount_t hmp);
1149
1150 void hammer_flusher_create(hammer_mount_t hmp);
1151 void hammer_flusher_destroy(hammer_mount_t hmp);
1152 void hammer_flusher_sync(hammer_mount_t hmp);
1153 int  hammer_flusher_async(hammer_mount_t hmp, hammer_flush_group_t flg);
1154 int  hammer_flusher_async_one(hammer_mount_t hmp);
1155 void hammer_flusher_wait(hammer_mount_t hmp, int seq);
1156 int  hammer_flusher_meta_limit(hammer_mount_t hmp);
1157 int  hammer_flusher_meta_halflimit(hammer_mount_t hmp);
1158 int  hammer_flusher_undo_exhausted(hammer_transaction_t trans, int quarter);
1159 void hammer_flusher_clean_loose_ios(hammer_mount_t hmp);
1160 void hammer_flusher_finalize(hammer_transaction_t trans, int final);
1161 int  hammer_flusher_haswork(hammer_mount_t hmp);
1162
1163
1164 int hammer_recover(hammer_mount_t hmp, hammer_volume_t rootvol);
1165 void hammer_recover_flush_buffers(hammer_mount_t hmp,
1166                         hammer_volume_t root_volume, int final);
1167
1168 void hammer_crc_set_blockmap(hammer_blockmap_t blockmap);
1169 void hammer_crc_set_volume(hammer_volume_ondisk_t ondisk);
1170 void hammer_crc_set_leaf(void *data, hammer_btree_leaf_elm_t leaf);
1171
1172 int hammer_crc_test_blockmap(hammer_blockmap_t blockmap);
1173 int hammer_crc_test_volume(hammer_volume_ondisk_t ondisk);
1174 int hammer_crc_test_btree(hammer_node_ondisk_t ondisk);
1175 int hammer_crc_test_leaf(void *data, hammer_btree_leaf_elm_t leaf);
1176 void hkprintf(const char *ctl, ...);
1177 udev_t hammer_fsid_to_udev(uuid_t *uuid);
1178
1179
1180 int hammer_blocksize(int64_t file_offset);
1181 int64_t hammer_blockdemarc(int64_t file_offset1, int64_t file_offset2);
1182
1183 #endif
1184
1185 static __inline void
1186 hammer_wait_mem_record(hammer_record_t record)
1187 {
1188         hammer_wait_mem_record_ident(record, "hmmwai");
1189 }
1190
1191 static __inline void
1192 hammer_lock_ex(struct hammer_lock *lock)
1193 {
1194         hammer_lock_ex_ident(lock, "hmrlck");
1195 }
1196
1197 /*
1198  * Indicate that a B-Tree node is being modified.
1199  */
1200 static __inline void
1201 hammer_modify_node_noundo(hammer_transaction_t trans, hammer_node_t node)
1202 {
1203         hammer_modify_buffer(trans, node->buffer, NULL, 0);
1204 }
1205
1206 static __inline void
1207 hammer_modify_node_all(hammer_transaction_t trans, struct hammer_node *node)
1208 {
1209         hammer_modify_buffer(trans, node->buffer,
1210                              node->ondisk, sizeof(*node->ondisk));
1211 }
1212
1213 static __inline void
1214 hammer_modify_node(hammer_transaction_t trans, hammer_node_t node,
1215                    void *base, int len)
1216 {
1217         hammer_crc_t *crcptr;
1218
1219         KKASSERT((char *)base >= (char *)node->ondisk &&
1220                  (char *)base + len <=
1221                     (char *)node->ondisk + sizeof(*node->ondisk));
1222         hammer_modify_buffer(trans, node->buffer, base, len);
1223         crcptr = &node->ondisk->crc;
1224         hammer_modify_buffer(trans, node->buffer, crcptr, sizeof(hammer_crc_t));
1225         --node->buffer->io.modify_refs; /* only want one ref */
1226 }
1227
1228 /*
1229  * Indicate that the specified modifications have been completed.
1230  *
1231  * Do not try to generate the crc here, it's very expensive to do and a
1232  * sequence of insertions or deletions can result in many calls to this
1233  * function on the same node.
1234  */
1235 static __inline void
1236 hammer_modify_node_done(hammer_node_t node)
1237 {
1238         node->flags |= HAMMER_NODE_CRCGOOD;
1239         if ((node->flags & HAMMER_NODE_NEEDSCRC) == 0) {
1240                 node->flags |= HAMMER_NODE_NEEDSCRC;
1241                 node->buffer->io.gencrc = 1;
1242                 hammer_ref_node(node);
1243         }
1244         hammer_modify_buffer_done(node->buffer);
1245 }
1246
1247 #define hammer_modify_volume_field(trans, vol, field)           \
1248         hammer_modify_volume(trans, vol, &(vol)->ondisk->field, \
1249                              sizeof((vol)->ondisk->field))
1250
1251 #define hammer_modify_node_field(trans, node, field)            \
1252         hammer_modify_node(trans, node, &(node)->ondisk->field, \
1253                              sizeof((node)->ondisk->field))
1254