sys/vfs/hammer/hammer.h

   1 /*
   2  * Copyright (c) 2007-2008 The DragonFly Project.  All rights reserved.
   3  *
   4  * This code is derived from software contributed to The DragonFly Project
   5  * by Matthew Dillon <dillon@backplane.com>
   6  *
   7  * Redistribution and use in source and binary forms, with or without
   8  * modification, are permitted provided that the following conditions
   9  * are met:
  10  *
  11  * 1. Redistributions of source code must retain the above copyright
  12  *    notice, this list of conditions and the following disclaimer.
  13  * 2. Redistributions in binary form must reproduce the above copyright
  14  *    notice, this list of conditions and the following disclaimer in
  15  *    the documentation and/or other materials provided with the
  16  *    distribution.
  17  * 3. Neither the name of The DragonFly Project nor the names of its
  18  *    contributors may be used to endorse or promote products derived
  19  *    from this software without specific, prior written permission.
  20  *
  21  * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
  22  * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
  23  * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS
  24  * FOR A PARTICULAR PURPOSE ARE DISCLAIMED.  IN NO EVENT SHALL THE
  25  * COPYRIGHT HOLDERS OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT,
  26  * INCIDENTAL, SPECIAL, EXEMPLARY OR CONSEQUENTIAL DAMAGES (INCLUDING,
  27  * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
  28  * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED
  29  * AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
  30  * OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT
  31  * OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
  32  * SUCH DAMAGE.
  33  *
  34  * $DragonFly: src/sys/vfs/hammer/hammer.h,v 1.130 2008/11/13 02:18:43 dillon Exp $
  35  */
  36 /*
  37  * This header file contains structures used internally by the HAMMERFS
  38  * implementation.  See hammer_disk.h for on-disk structures.
  39  */
  40
  41 #include <sys/param.h>
  42 #include <sys/types.h>
  43 #include <sys/kernel.h>
  44 #include <sys/conf.h>
  45 #include <sys/systm.h>
  46 #include <sys/tree.h>
  47 #include <sys/malloc.h>
  48 #include <sys/mount.h>
  49 #include <sys/mountctl.h>
  50 #include <sys/vnode.h>
  51 #include <sys/proc.h>
  52 #include <sys/priv.h>
  53 #include <sys/stat.h>
  54 #include <sys/globaldata.h>
  55 #include <sys/lockf.h>
  56 #include <sys/buf.h>
  57 #include <sys/queue.h>
  58 #include <sys/ktr.h>
  59 #include <sys/globaldata.h>
  60
  61 #include <sys/buf2.h>
  62 #include <sys/signal2.h>
  63 #include "hammer_disk.h"
  64 #include "hammer_mount.h"
  65 #include "hammer_ioctl.h"
  66
  67 #if defined(_KERNEL) || defined(_KERNEL_STRUCTURES)
  68
  69 MALLOC_DECLARE(M_HAMMER);
  70
  71 /*
  72  * Kernel trace
  73  */
  74 #if !defined(KTR_HAMMER)
  75 #define KTR_HAMMER      KTR_ALL
  76 #endif
  77 KTR_INFO_MASTER_EXTERN(hammer);
  78
  79 /*
  80  * Misc structures
  81  */
  82 struct hammer_mount;
  83
  84 /*
  85  * Key structure used for custom RB tree inode lookups.  This prototypes
  86  * the function hammer_ino_rb_tree_RB_LOOKUP_INFO(root, info).
  87  */
  88 typedef struct hammer_inode_info {
  89         int64_t         obj_id;         /* (key) object identifier */
  90         hammer_tid_t    obj_asof;       /* (key) snapshot transid or 0 */
  91         u_int32_t       obj_localization; /* (key) pseudo-fs */
  92         union {
  93                 struct hammer_btree_leaf_elm *leaf;
  94         } u;
  95 } *hammer_inode_info_t;
  96
  97 typedef enum hammer_transaction_type {
  98         HAMMER_TRANS_RO,
  99         HAMMER_TRANS_STD,
 100         HAMMER_TRANS_FLS
 101 } hammer_transaction_type_t;
 102
 103 /*
 104  * HAMMER Transaction tracking
 105  */
 106 struct hammer_transaction {
 107         hammer_transaction_type_t type;
 108         struct hammer_mount *hmp;
 109         hammer_tid_t    tid;
 110         u_int64_t       time;
 111         u_int32_t       time32;
 112         int             sync_lock_refs;
 113         int             flags;
 114         struct hammer_volume *rootvol;
 115 };
 116
 117 typedef struct hammer_transaction *hammer_transaction_t;
 118
 119 #define HAMMER_TRANSF_NEWINODE  0x0001
 120 #define HAMMER_TRANSF_DIDIO     0x0002
 121
 122 /*
 123  * HAMMER locks
 124  */
 125 struct hammer_lock {
 126         int     refs;           /* active references delay writes */
 127         int     lockcount;      /* lock count for exclusive/shared access */
 128         int     wanted;
 129         int     exwanted;       /* number of threads waiting for ex lock */
 130         struct thread *locktd;
 131 };
 132
 133 static __inline int
 134 hammer_islocked(struct hammer_lock *lock)
 135 {
 136         return(lock->lockcount != 0);
 137 }
 138
 139 static __inline int
 140 hammer_isactive(struct hammer_lock *lock)
 141 {
 142         return(lock->refs != 0);
 143 }
 144
 145 static __inline int
 146 hammer_islastref(struct hammer_lock *lock)
 147 {
 148         return(lock->refs == 1);
 149 }
 150
 151 /*
 152  * Return if we specifically own the lock exclusively.
 153  */
 154 static __inline int
 155 hammer_lock_excl_owned(struct hammer_lock *lock, thread_t td)
 156 {
 157         if (lock->lockcount > 0 && lock->locktd == td)
 158                 return(1);
 159         return(0);
 160 }
 161
 162 /*
 163  * Flush state, used by various structures
 164  */
 165 typedef enum hammer_inode_state {
 166         HAMMER_FST_IDLE,
 167         HAMMER_FST_SETUP,
 168         HAMMER_FST_FLUSH
 169 } hammer_inode_state_t;
 170
 171 TAILQ_HEAD(hammer_record_list, hammer_record);
 172
 173 /*
 174  * Pseudo-filesystem extended data tracking
 175  */
 176 struct hammer_pfs_rb_tree;
 177 struct hammer_pseudofs_inmem;
 178 RB_HEAD(hammer_pfs_rb_tree, hammer_pseudofs_inmem);
 179 RB_PROTOTYPE2(hammer_pfs_rb_tree, hammer_pseudofs_inmem, rb_node,
 180               hammer_pfs_rb_compare, u_int32_t);
 181
 182 struct hammer_pseudofs_inmem {
 183         RB_ENTRY(hammer_pseudofs_inmem) rb_node;
 184         struct hammer_lock      lock;
 185         u_int32_t               localization;
 186         hammer_tid_t            create_tid;
 187         int                     flags;
 188         udev_t                  fsid_udev;
 189         struct hammer_pseudofs_data pfsd;
 190 };
 191
 192 typedef struct hammer_pseudofs_inmem *hammer_pseudofs_inmem_t;
 193
 194 #define HAMMER_PFSM_DELETED     0x0001
 195
 196 /*
 197  * Cache object ids.  A fixed number of objid cache structures are
 198  * created to reserve object id's for newly created files in multiples
 199  * of 100,000, localized to a particular directory, and recycled as
 200  * needed.  This allows parallel create operations in different
 201  * directories to retain fairly localized object ids which in turn
 202  * improves reblocking performance and layout.
 203  */
 204 #define OBJID_CACHE_SIZE        1024
 205 #define OBJID_CACHE_BULK        100000
 206
 207 typedef struct hammer_objid_cache {
 208         TAILQ_ENTRY(hammer_objid_cache) entry;
 209         struct hammer_inode             *dip;
 210         hammer_tid_t                    next_tid;
 211         int                             count;
 212 } *hammer_objid_cache_t;
 213
 214 /*
 215  * Associate an inode with a B-Tree node to cache search start positions
 216  */
 217 typedef struct hammer_node_cache {
 218         TAILQ_ENTRY(hammer_node_cache)  entry;
 219         struct hammer_node              *node;
 220         struct hammer_inode             *ip;
 221 } *hammer_node_cache_t;
 222
 223 TAILQ_HEAD(hammer_node_cache_list, hammer_node_cache);
 224
 225 /*
 226  * Structure used to organize flush groups.  Flush groups must be
 227  * organized into chunks in order to avoid blowing out the UNDO FIFO.
 228  * Without this a 'sync' could end up flushing 50,000 inodes in a single
 229  * transaction.
 230  */
 231 struct hammer_flush_group {
 232         TAILQ_ENTRY(hammer_flush_group) flush_entry;
 233         TAILQ_HEAD(, hammer_inode)      flush_list;
 234         int                             unused01;       /* inode load */
 235         int                             total_count;    /* record load */
 236         int                             running;        /* group is running */
 237         int                             closed;
 238         int                             refs;
 239 };
 240
 241 typedef struct hammer_flush_group *hammer_flush_group_t;
 242
 243 TAILQ_HEAD(hammer_flush_group_list, hammer_flush_group);
 244
 245 /*
 246  * Structure used to represent an inode in-memory.
 247  *
 248  * The record and data associated with an inode may be out of sync with
 249  * the disk (xDIRTY flags), or not even on the disk at all (ONDISK flag
 250  * clear).
 251  *
 252  * An inode may also hold a cache of unsynchronized records, used for
 253  * database and directories only.  Unsynchronized regular file data is
 254  * stored in the buffer cache.
 255  *
 256  * NOTE: A file which is created and destroyed within the initial
 257  * synchronization period can wind up not doing any disk I/O at all.
 258  *
 259  * Finally, an inode may cache numerous disk-referencing B-Tree cursors.
 260  */
 261 struct hammer_ino_rb_tree;
 262 struct hammer_inode;
 263 RB_HEAD(hammer_ino_rb_tree, hammer_inode);
 264 RB_PROTOTYPEX(hammer_ino_rb_tree, INFO, hammer_inode, rb_node,
 265               hammer_ino_rb_compare, hammer_inode_info_t);
 266
 267 struct hammer_rec_rb_tree;
 268 struct hammer_record;
 269 RB_HEAD(hammer_rec_rb_tree, hammer_record);
 270 RB_PROTOTYPEX(hammer_rec_rb_tree, INFO, hammer_record, rb_node,
 271               hammer_rec_rb_compare, hammer_btree_leaf_elm_t);
 272
 273 TAILQ_HEAD(hammer_node_list, hammer_node);
 274
 275 struct hammer_inode {
 276         RB_ENTRY(hammer_inode)  rb_node;
 277         hammer_inode_state_t    flush_state;
 278         hammer_flush_group_t    flush_group;
 279         TAILQ_ENTRY(hammer_inode) flush_entry;
 280         struct hammer_record_list target_list;  /* target of dependant recs */
 281         int64_t                 obj_id;         /* (key) object identifier */
 282         hammer_tid_t            obj_asof;       /* (key) snapshot or 0 */
 283         u_int32_t               obj_localization; /* (key) pseudo-fs */
 284         struct hammer_mount     *hmp;
 285         hammer_objid_cache_t    objid_cache;
 286         int                     flags;
 287         int                     error;          /* flush error */
 288         int                     cursor_ip_refs; /* sanity */
 289         int                     rsv_recs;
 290         struct vnode            *vp;
 291         hammer_pseudofs_inmem_t pfsm;
 292         struct lockf            advlock;
 293         struct hammer_lock      lock;           /* sync copy interlock */
 294         off_t                   trunc_off;
 295         struct hammer_btree_leaf_elm ino_leaf;  /* in-memory cache */
 296         struct hammer_inode_data ino_data;      /* in-memory cache */
 297         struct hammer_rec_rb_tree rec_tree;     /* in-memory cache */
 298         struct hammer_node_cache cache[2];      /* search initiate cache */
 299
 300         /*
 301          * When a demark is created to synchronize an inode to
 302          * disk, certain fields are copied so the front-end VOPs
 303          * can continue to run in parallel with the synchronization
 304          * occuring in the background.
 305          */
 306         int             sync_flags;             /* to-sync flags cache */
 307         off_t           sync_trunc_off;         /* to-sync truncation */
 308         off_t           save_trunc_off;         /* write optimization */
 309         struct hammer_btree_leaf_elm sync_ino_leaf; /* to-sync cache */
 310         struct hammer_inode_data sync_ino_data; /* to-sync cache */
 311 };
 312
 313 typedef struct hammer_inode *hammer_inode_t;
 314
 315 #define VTOI(vp)        ((struct hammer_inode *)(vp)->v_data)
 316
 317 #define HAMMER_INODE_DDIRTY     0x0001  /* in-memory ino_data is dirty */
 318                                         /* (not including atime/mtime) */
 319 #define HAMMER_INODE_RSV_INODES 0x0002  /* hmp->rsv_inodes bumped */
 320 #define HAMMER_INODE_CONN_DOWN  0x0004  /* include in downward recursion */
 321 #define HAMMER_INODE_XDIRTY     0x0008  /* in-memory records */
 322 #define HAMMER_INODE_ONDISK     0x0010  /* inode is on-disk (else not yet) */
 323 #define HAMMER_INODE_FLUSH      0x0020  /* flush on last ref */
 324 #define HAMMER_INODE_DELETED    0x0080  /* inode delete (backend) */
 325 #define HAMMER_INODE_DELONDISK  0x0100  /* delete synchronized to disk */
 326 #define HAMMER_INODE_RO         0x0200  /* read-only (because of as-of) */
 327 #define HAMMER_INODE_VHELD      0x0400  /* vnode held on sync */
 328 #define HAMMER_INODE_DONDISK    0x0800  /* data records may be on disk */
 329 #define HAMMER_INODE_BUFS       0x1000  /* dirty high level bps present */
 330 #define HAMMER_INODE_REFLUSH    0x2000  /* flush on dependancy / reflush */
 331 #define HAMMER_INODE_RECLAIM    0x4000  /* trying to reclaim */
 332 #define HAMMER_INODE_FLUSHW     0x8000  /* Someone waiting for flush */
 333
 334 #define HAMMER_INODE_TRUNCATED  0x00010000
 335 #define HAMMER_INODE_DELETING   0x00020000 /* inode delete request (frontend)*/
 336 #define HAMMER_INODE_RESIGNAL   0x00040000 /* re-signal on re-flush */
 337 #define HAMMER_INODE_ATIME      0x00100000 /* in-memory atime modified */
 338 #define HAMMER_INODE_MTIME      0x00200000 /* in-memory mtime modified */
 339 #define HAMMER_INODE_WOULDBLOCK 0x00400000 /* re-issue to new flush group */
 340
 341 #define HAMMER_INODE_MODMASK    (HAMMER_INODE_DDIRTY|                       \
 342                                  HAMMER_INODE_XDIRTY|HAMMER_INODE_BUFS|     \
 343                                  HAMMER_INODE_ATIME|HAMMER_INODE_MTIME|     \
 344                                  HAMMER_INODE_TRUNCATED|HAMMER_INODE_DELETING)
 345
 346 #define HAMMER_INODE_MODMASK_NOXDIRTY \
 347                                 (HAMMER_INODE_MODMASK & ~HAMMER_INODE_XDIRTY)
 348
 349 #define HAMMER_FLUSH_GROUP_SIZE 64
 350
 351 #define HAMMER_FLUSH_SIGNAL     0x0001
 352 #define HAMMER_FLUSH_RECURSION  0x0002
 353
 354 /*
 355  * Used by the inode reclaim code to pipeline reclaims and avoid
 356  * blowing out kernel memory or letting the flusher get too far
 357  * behind.  The reclaim wakes up when count reaches 0 or the
 358  * timer expires.
 359  */
 360 struct hammer_reclaim {
 361         TAILQ_ENTRY(hammer_reclaim) entry;
 362         int     count;
 363 };
 364
 365 #define HAMMER_RECLAIM_FLUSH    2000
 366 #define HAMMER_RECLAIM_WAIT     4000
 367
 368 /*
 369  * Structure used to represent an unsynchronized record in-memory.  These
 370  * records typically represent directory entries.  Only non-historical
 371  * records are kept in-memory.
 372  *
 373  * Records are organized as a per-inode RB-Tree.  If the inode is not
 374  * on disk then neither are any records and the in-memory record tree
 375  * represents the entire contents of the inode.  If the inode is on disk
 376  * then the on-disk B-Tree is scanned in parallel with the in-memory
 377  * RB-Tree to synthesize the current state of the file.
 378  *
 379  * Records are also used to enforce the ordering of directory create/delete
 380  * operations.  A new inode will not be flushed to disk unless its related
 381  * directory entry is also being flushed at the same time.  A directory entry
 382  * will not be removed unless its related inode is also being removed at the
 383  * same time.
 384  */
 385 typedef enum hammer_record_type {
 386         HAMMER_MEM_RECORD_GENERAL,      /* misc record */
 387         HAMMER_MEM_RECORD_INODE,        /* inode record */
 388         HAMMER_MEM_RECORD_ADD,          /* positive memory cache record */
 389         HAMMER_MEM_RECORD_DEL,          /* negative delete-on-disk record */
 390         HAMMER_MEM_RECORD_DATA          /* bulk-data record w/on-disk ref */
 391 } hammer_record_type_t;
 392
 393 struct hammer_record {
 394         RB_ENTRY(hammer_record)         rb_node;
 395         TAILQ_ENTRY(hammer_record)      target_entry;
 396         hammer_inode_state_t            flush_state;
 397         hammer_flush_group_t            flush_group;
 398         hammer_record_type_t            type;
 399         struct hammer_lock              lock;
 400         struct hammer_reserve           *resv;
 401         struct hammer_inode             *ip;
 402         struct hammer_inode             *target_ip;
 403         struct hammer_btree_leaf_elm    leaf;
 404         union hammer_data_ondisk        *data;
 405         int                             flags;
 406         hammer_off_t                    zone2_offset;   /* direct-write only */
 407 };
 408
 409 typedef struct hammer_record *hammer_record_t;
 410
 411 /*
 412  * Record flags.  Note that FE can only be set by the frontend if the
 413  * record has not been interlocked by the backend w/ BE.
 414  */
 415 #define HAMMER_RECF_ALLOCDATA           0x0001
 416 #define HAMMER_RECF_ONRBTREE            0x0002
 417 #define HAMMER_RECF_DELETED_FE          0x0004  /* deleted (frontend) */
 418 #define HAMMER_RECF_DELETED_BE          0x0008  /* deleted (backend) */
 419 #define HAMMER_RECF_COMMITTED           0x0010  /* committed to the B-Tree */
 420 #define HAMMER_RECF_INTERLOCK_BE        0x0020  /* backend interlock */
 421 #define HAMMER_RECF_WANTED              0x0040  /* wanted by the frontend */
 422 #define HAMMER_RECF_CONVERT_DELETE      0x0100  /* special case */
 423 #define HAMMER_RECF_DIRECT_IO           0x0200  /* related direct I/O running*/
 424 #define HAMMER_RECF_DIRECT_WAIT         0x0400  /* related direct I/O running*/
 425 #define HAMMER_RECF_DIRECT_INVAL        0x0800  /* buffer alias invalidation */
 426
 427 /*
 428  * hammer_delete_at_cursor() flags
 429  */
 430 #define HAMMER_DELETE_ADJUST            0x0001
 431 #define HAMMER_DELETE_DESTROY           0x0002
 432
 433 /*
 434  * In-memory structures representing on-disk structures.
 435  */
 436 struct hammer_volume;
 437 struct hammer_buffer;
 438 struct hammer_node;
 439 struct hammer_undo;
 440 struct hammer_reserve;
 441
 442 RB_HEAD(hammer_vol_rb_tree, hammer_volume);
 443 RB_HEAD(hammer_buf_rb_tree, hammer_buffer);
 444 RB_HEAD(hammer_nod_rb_tree, hammer_node);
 445 RB_HEAD(hammer_und_rb_tree, hammer_undo);
 446 RB_HEAD(hammer_res_rb_tree, hammer_reserve);
 447
 448 RB_PROTOTYPE2(hammer_vol_rb_tree, hammer_volume, rb_node,
 449               hammer_vol_rb_compare, int32_t);
 450 RB_PROTOTYPE2(hammer_buf_rb_tree, hammer_buffer, rb_node,
 451               hammer_buf_rb_compare, hammer_off_t);
 452 RB_PROTOTYPE2(hammer_nod_rb_tree, hammer_node, rb_node,
 453               hammer_nod_rb_compare, hammer_off_t);
 454 RB_PROTOTYPE2(hammer_und_rb_tree, hammer_undo, rb_node,
 455               hammer_und_rb_compare, hammer_off_t);
 456 RB_PROTOTYPE2(hammer_res_rb_tree, hammer_reserve, rb_node,
 457               hammer_res_rb_compare, hammer_off_t);
 458
 459 /*
 460  * IO management - embedded at the head of various in-memory structures
 461  *
 462  * VOLUME       - hammer_volume containing meta-data
 463  * META_BUFFER  - hammer_buffer containing meta-data
 464  * DATA_BUFFER  - hammer_buffer containing pure-data
 465  *
 466  * Dirty volume headers and dirty meta-data buffers are locked until the
 467  * flusher can sequence them out.  Dirty pure-data buffers can be written.
 468  * Clean buffers can be passively released.
 469  */
 470 typedef enum hammer_io_type {
 471         HAMMER_STRUCTURE_VOLUME,
 472         HAMMER_STRUCTURE_META_BUFFER,
 473         HAMMER_STRUCTURE_UNDO_BUFFER,
 474         HAMMER_STRUCTURE_DATA_BUFFER
 475 } hammer_io_type_t;
 476
 477 union hammer_io_structure;
 478 struct hammer_io;
 479
 480 struct worklist {
 481         LIST_ENTRY(worklist) node;
 482 };
 483
 484 TAILQ_HEAD(hammer_io_list, hammer_io);
 485 typedef struct hammer_io_list *hammer_io_list_t;
 486
 487 struct hammer_io {
 488         struct worklist         worklist;
 489         struct hammer_lock      lock;
 490         enum hammer_io_type     type;
 491         struct hammer_mount     *hmp;
 492         struct hammer_volume    *volume;
 493         TAILQ_ENTRY(hammer_io)  mod_entry; /* list entry if modified */
 494         hammer_io_list_t        mod_list;
 495         struct buf              *bp;
 496         int64_t                 offset;    /* zone-2 offset */
 497         int                     bytes;     /* buffer cache buffer size */
 498         int                     loading;   /* loading/unloading interlock */
 499         int                     modify_refs;
 500
 501         u_int           modified : 1;   /* bp's data was modified */
 502         u_int           released : 1;   /* bp released (w/ B_LOCKED set) */
 503         u_int           running : 1;    /* bp write IO in progress */
 504         u_int           waiting : 1;    /* someone is waiting on us */
 505         u_int           validated : 1;  /* ondisk has been validated */
 506         u_int           waitdep : 1;    /* flush waits for dependancies */
 507         u_int           recovered : 1;  /* has recovery ref */
 508         u_int           waitmod : 1;    /* waiting for modify_refs */
 509         u_int           reclaim : 1;    /* reclaim requested */
 510         u_int           gencrc : 1;     /* crc needs to be generated */
 511         u_int           ioerror : 1;    /* abort on io-error */
 512 };
 513
 514 typedef struct hammer_io *hammer_io_t;
 515
 516 #define HAMMER_CLUSTER_SIZE     (64 * 1024)
 517 #if HAMMER_CLUSTER_SIZE > MAXBSIZE
 518 #undef  HAMMER_CLUSTER_SIZE
 519 #define HAMMER_CLUSTER_SIZE     MAXBSIZE
 520 #endif
 521 #define HAMMER_CLUSTER_BUFS     (HAMMER_CLUSTER_SIZE / HAMMER_BUFSIZE)
 522
 523 /*
 524  * In-memory volume representing on-disk buffer
 525  */
 526 struct hammer_volume {
 527         struct hammer_io io;
 528         RB_ENTRY(hammer_volume) rb_node;
 529         struct hammer_volume_ondisk *ondisk;
 530         int32_t vol_no;
 531         int64_t nblocks;        /* note: special calculation for statfs */
 532         int64_t buffer_base;    /* base offset of buffer 0 */
 533         hammer_off_t maxbuf_off; /* Maximum buffer offset (zone-2) */
 534         hammer_off_t maxraw_off; /* Maximum raw offset for device */
 535         char    *vol_name;
 536         struct vnode *devvp;
 537         int     vol_flags;
 538 };
 539
 540 typedef struct hammer_volume *hammer_volume_t;
 541
 542 /*
 543  * In-memory buffer (other then volume, super-cluster, or cluster),
 544  * representing an on-disk buffer.
 545  */
 546 struct hammer_buffer {
 547         struct hammer_io io;
 548         RB_ENTRY(hammer_buffer) rb_node;
 549         void *ondisk;
 550         hammer_off_t zoneX_offset;
 551         hammer_off_t zone2_offset;
 552         struct hammer_reserve *resv;
 553         struct hammer_node_list clist;
 554 };
 555
 556 typedef struct hammer_buffer *hammer_buffer_t;
 557
 558 /*
 559  * In-memory B-Tree node, representing an on-disk B-Tree node.
 560  *
 561  * This is a hang-on structure which is backed by a hammer_buffer,
 562  * indexed by a hammer_cluster, and used for fine-grained locking of
 563  * B-Tree nodes in order to properly control lock ordering.  A hammer_buffer
 564  * can contain multiple nodes representing wildly disassociated portions
 565  * of the B-Tree so locking cannot be done on a buffer-by-buffer basis.
 566  *
 567  * This structure uses a cluster-relative index to reduce the number
 568  * of layers required to access it, and also because all on-disk B-Tree
 569  * references are cluster-relative offsets.
 570  */
 571 struct hammer_node {
 572         struct hammer_lock      lock;           /* node-by-node lock */
 573         TAILQ_ENTRY(hammer_node) entry;         /* per-buffer linkage */
 574         RB_ENTRY(hammer_node)   rb_node;        /* per-cluster linkage */
 575         hammer_off_t            node_offset;    /* full offset spec */
 576         struct hammer_mount     *hmp;
 577         struct hammer_buffer    *buffer;        /* backing buffer */
 578         hammer_node_ondisk_t    ondisk;         /* ptr to on-disk structure */
 579         TAILQ_HEAD(, hammer_cursor) cursor_list;  /* deadlock recovery */
 580         struct hammer_node_cache_list cache_list; /* passive caches */
 581         int                     flags;
 582         int                     loading;        /* load interlock */
 583 };
 584
 585 #define HAMMER_NODE_DELETED     0x0001
 586 #define HAMMER_NODE_FLUSH       0x0002
 587 #define HAMMER_NODE_CRCGOOD     0x0004
 588 #define HAMMER_NODE_NEEDSCRC    0x0008
 589 #define HAMMER_NODE_NEEDSMIRROR 0x0010
 590
 591 typedef struct hammer_node      *hammer_node_t;
 592
 593 /*
 594  * List of locked nodes.  This structure is used to lock potentially large
 595  * numbers of nodes as an aid for complex B-Tree operations.
 596  */
 597 struct hammer_node_lock;
 598 TAILQ_HEAD(hammer_node_lock_list, hammer_node_lock);
 599
 600 struct hammer_node_lock {
 601         TAILQ_ENTRY(hammer_node_lock) entry;
 602         struct hammer_node_lock_list  list;
 603         struct hammer_node_lock       *parent;
 604         hammer_node_t   node;
 605         hammer_node_ondisk_t copy;      /* copy of on-disk data */
 606         int             index;          /* index of this node in parent */
 607         int             count;          /* count children */
 608         int             flags;
 609 };
 610
 611 typedef struct hammer_node_lock *hammer_node_lock_t;
 612
 613 #define HAMMER_NODE_LOCK_UPDATED        0x0001
 614
 615 /*
 616  * Common I/O management structure - embedded in in-memory structures
 617  * which are backed by filesystem buffers.
 618  */
 619 union hammer_io_structure {
 620         struct hammer_io        io;
 621         struct hammer_volume    volume;
 622         struct hammer_buffer    buffer;
 623 };
 624
 625 typedef union hammer_io_structure *hammer_io_structure_t;
 626
 627 /*
 628  * The reserve structure prevents the blockmap from allocating
 629  * out of a reserved bigblock.  Such reservations are used by
 630  * the direct-write mechanism.
 631  *
 632  * The structure is also used to hold off on reallocations of
 633  * big blocks from the freemap until flush dependancies have
 634  * been dealt with.
 635  */
 636 struct hammer_reserve {
 637         RB_ENTRY(hammer_reserve) rb_node;
 638         TAILQ_ENTRY(hammer_reserve) delay_entry;
 639         int             flush_group;
 640         int             flags;
 641         int             refs;
 642         int             zone;
 643         int             append_off;
 644         hammer_off_t    zone_offset;
 645 };
 646
 647 typedef struct hammer_reserve *hammer_reserve_t;
 648
 649 #define HAMMER_RESF_ONDELAY     0x0001
 650 #define HAMMER_RESF_LAYER2FREE  0x0002
 651
 652 #include "hammer_cursor.h"
 653
 654 /*
 655  * The undo structure tracks recent undos to avoid laying down duplicate
 656  * undos within a flush group, saving us a significant amount of overhead.
 657  *
 658  * This is strictly a heuristic.
 659  */
 660 #define HAMMER_MAX_UNDOS                1024
 661 #define HAMMER_MAX_FLUSHERS             4
 662
 663 struct hammer_undo {
 664         RB_ENTRY(hammer_undo)   rb_node;
 665         TAILQ_ENTRY(hammer_undo) lru_entry;
 666         hammer_off_t            offset;
 667         int                     bytes;
 668 };
 669
 670 typedef struct hammer_undo *hammer_undo_t;
 671
 672 struct hammer_flusher_info;
 673 TAILQ_HEAD(hammer_flusher_info_list, hammer_flusher_info);
 674
 675 struct hammer_flusher {
 676         int             signal;         /* flusher thread sequencer */
 677         int             act;            /* currently active flush group */
 678         int             done;           /* set to act when complete */
 679         int             next;           /* next flush group */
 680         int             group_lock;     /* lock sequencing of the next flush */
 681         int             exiting;        /* request master exit */
 682         thread_t        td;             /* master flusher thread */
 683         hammer_tid_t    tid;            /* last flushed transaction id */
 684         int             finalize_want;          /* serialize finalization */
 685         struct hammer_lock finalize_lock;       /* serialize finalization */
 686         struct hammer_transaction trans;        /* shared transaction */
 687         struct hammer_flusher_info_list run_list;
 688         struct hammer_flusher_info_list ready_list;
 689 };
 690
 691 /*
 692  * Internal hammer mount data structure
 693  */
 694 struct hammer_mount {
 695         struct mount *mp;
 696         /*struct vnode *rootvp;*/
 697         struct hammer_ino_rb_tree rb_inos_root;
 698         struct hammer_vol_rb_tree rb_vols_root;
 699         struct hammer_nod_rb_tree rb_nods_root;
 700         struct hammer_und_rb_tree rb_undo_root;
 701         struct hammer_res_rb_tree rb_resv_root;
 702         struct hammer_buf_rb_tree rb_bufs_root;
 703         struct hammer_pfs_rb_tree rb_pfsm_root;
 704         struct hammer_volume *rootvol;
 705         struct hammer_base_elm root_btree_beg;
 706         struct hammer_base_elm root_btree_end;
 707
 708         struct malloc_type      *m_misc;
 709         struct malloc_type      *m_inodes;
 710
 711         int     flags;          /* HAMMER_MOUNT_xxx flags */
 712         int     hflags;
 713         int     ronly;
 714         int     nvolumes;
 715         int     volume_iterator;
 716         int     master_id;      /* -1 or 0-15 - clustering and mirroring */
 717         int     version;        /* hammer filesystem version to use */
 718         int     rsv_inodes;     /* reserved space due to dirty inodes */
 719         int64_t rsv_databytes;  /* reserved space due to record data */
 720         int     rsv_recs;       /* reserved space due to dirty records */
 721         int     rsv_fromdelay;  /* bigblocks reserved due to flush delay */
 722         int     undo_rec_limit; /* based on size of undo area */
 723         int     last_newrecords;
 724         int     count_newrecords;
 725
 726         int     inode_reclaims; /* inodes pending reclaim by flusher */
 727         int     count_inodes;   /* total number of inodes */
 728         int     count_iqueued;  /* inodes queued to flusher */
 729
 730         struct hammer_flusher flusher;
 731
 732         u_int   check_interrupt;
 733         uuid_t  fsid;
 734         struct hammer_io_list volu_list;        /* dirty undo buffers */
 735         struct hammer_io_list undo_list;        /* dirty undo buffers */
 736         struct hammer_io_list data_list;        /* dirty data buffers */
 737         struct hammer_io_list alt_data_list;    /* dirty data buffers */
 738         struct hammer_io_list meta_list;        /* dirty meta bufs    */
 739         struct hammer_io_list lose_list;        /* loose buffers      */
 740         int     locked_dirty_space;             /* meta/volu count    */
 741         int     io_running_space;
 742         int     objid_cache_count;
 743         int     error;                          /* critical I/O error */
 744         struct krate    krate;                  /* rate limited kprintf */
 745         hammer_tid_t    asof;                   /* snapshot mount */
 746         hammer_tid_t    next_tid;
 747         hammer_tid_t    flush_tid1;             /* flusher tid sequencing */
 748         hammer_tid_t    flush_tid2;             /* flusher tid sequencing */
 749         int64_t copy_stat_freebigblocks;        /* number of free bigblocks */
 750
 751         struct netexport export;
 752         struct hammer_lock sync_lock;
 753         struct hammer_lock free_lock;
 754         struct hammer_lock undo_lock;
 755         struct hammer_lock blkmap_lock;
 756         struct hammer_blockmap  blockmap[HAMMER_MAX_ZONES];
 757         struct hammer_undo      undos[HAMMER_MAX_UNDOS];
 758         int                     undo_alloc;
 759         TAILQ_HEAD(, hammer_undo)  undo_lru_list;
 760         TAILQ_HEAD(, hammer_reserve) delay_list;
 761         struct hammer_flush_group_list  flush_group_list;
 762         hammer_flush_group_t    next_flush_group;
 763         TAILQ_HEAD(, hammer_objid_cache) objid_cache_list;
 764         TAILQ_HEAD(, hammer_reclaim) reclaim_list;
 765 };
 766
 767 typedef struct hammer_mount     *hammer_mount_t;
 768
 769 #define HAMMER_MOUNT_CRITICAL_ERROR     0x0001
 770 #define HAMMER_MOUNT_FLUSH_RECOVERY     0x0002
 771
 772 struct hammer_sync_info {
 773         int error;
 774         int waitfor;
 775 };
 776
 777 #endif
 778
 779 /*
 780  * checkspace slop (8MB chunks), higher numbers are more conservative.
 781  */
 782 #define HAMMER_CHKSPC_REBLOCK   25
 783 #define HAMMER_CHKSPC_MIRROR    20
 784 #define HAMMER_CHKSPC_WRITE     20
 785 #define HAMMER_CHKSPC_CREATE    20
 786 #define HAMMER_CHKSPC_REMOVE    10
 787 #define HAMMER_CHKSPC_EMERGENCY 0
 788
 789 #if defined(_KERNEL)
 790
 791 extern struct vop_ops hammer_vnode_vops;
 792 extern struct vop_ops hammer_spec_vops;
 793 extern struct vop_ops hammer_fifo_vops;
 794 extern struct bio_ops hammer_bioops;
 795
 796 extern int hammer_debug_io;
 797 extern int hammer_debug_general;
 798 extern int hammer_debug_debug;
 799 extern int hammer_debug_inode;
 800 extern int hammer_debug_locks;
 801 extern int hammer_debug_btree;
 802 extern int hammer_debug_tid;
 803 extern int hammer_debug_recover;
 804 extern int hammer_debug_recover_faults;
 805 extern int hammer_cluster_enable;
 806 extern int hammer_count_fsyncs;
 807 extern int hammer_count_inodes;
 808 extern int hammer_count_iqueued;
 809 extern int hammer_count_reclaiming;
 810 extern int hammer_count_records;
 811 extern int hammer_count_record_datas;
 812 extern int hammer_count_volumes;
 813 extern int hammer_count_buffers;
 814 extern int hammer_count_nodes;
 815 extern int64_t hammer_count_extra_space_used;
 816 extern int64_t hammer_stats_btree_lookups;
 817 extern int64_t hammer_stats_btree_searches;
 818 extern int64_t hammer_stats_btree_inserts;
 819 extern int64_t hammer_stats_btree_deletes;
 820 extern int64_t hammer_stats_btree_elements;
 821 extern int64_t hammer_stats_btree_splits;
 822 extern int64_t hammer_stats_btree_iterations;
 823 extern int64_t hammer_stats_record_iterations;
 824 extern int64_t hammer_stats_file_read;
 825 extern int64_t hammer_stats_file_write;
 826 extern int64_t hammer_stats_file_iopsr;
 827 extern int64_t hammer_stats_file_iopsw;
 828 extern int64_t hammer_stats_disk_read;
 829 extern int64_t hammer_stats_disk_write;
 830 extern int64_t hammer_stats_inode_flushes;
 831 extern int64_t hammer_stats_commits;
 832 extern int hammer_count_dirtybufspace;
 833 extern int hammer_count_refedbufs;
 834 extern int hammer_count_reservations;
 835 extern int hammer_count_io_running_read;
 836 extern int hammer_count_io_running_write;
 837 extern int hammer_count_io_locked;
 838 extern int hammer_limit_dirtybufspace;
 839 extern int hammer_limit_recs;
 840 extern int hammer_bio_count;
 841 extern int hammer_verify_zone;
 842 extern int hammer_verify_data;
 843 extern int hammer_write_mode;
 844 extern int hammer_autoflush;
 845 extern int64_t hammer_contention_count;
 846
 847 void    hammer_critical_error(hammer_mount_t hmp, hammer_inode_t ip,
 848                         int error, const char *msg);
 849 int     hammer_vop_inactive(struct vop_inactive_args *);
 850 int     hammer_vop_reclaim(struct vop_reclaim_args *);
 851 int     hammer_get_vnode(struct hammer_inode *ip, struct vnode **vpp);
 852 struct hammer_inode *hammer_get_inode(hammer_transaction_t trans,
 853                         hammer_inode_t dip, int64_t obj_id,
 854                         hammer_tid_t asof, u_int32_t localization,
 855                         int flags, int *errorp);
 856 void    hammer_scan_inode_snapshots(hammer_mount_t hmp,
 857                         hammer_inode_info_t iinfo,
 858                         int (*callback)(hammer_inode_t ip, void *data),
 859                         void *data);
 860 void    hammer_put_inode(struct hammer_inode *ip);
 861 void    hammer_put_inode_ref(struct hammer_inode *ip);
 862 void    hammer_inode_waitreclaims(hammer_mount_t hmp);
 863 void    hammer_inode_waithard(hammer_mount_t hmp);
 864
 865 int     hammer_unload_volume(hammer_volume_t volume, void *data __unused);
 866 int     hammer_adjust_volume_mode(hammer_volume_t volume, void *data __unused);
 867
 868 int     hammer_unload_buffer(hammer_buffer_t buffer, void *data __unused);
 869 int     hammer_install_volume(hammer_mount_t hmp, const char *volname,
 870                         struct vnode *devvp);
 871 int     hammer_mountcheck_volumes(hammer_mount_t hmp);
 872
 873 int     hammer_mem_add(hammer_record_t record);
 874 int     hammer_ip_lookup(hammer_cursor_t cursor);
 875 int     hammer_ip_first(hammer_cursor_t cursor);
 876 int     hammer_ip_next(hammer_cursor_t cursor);
 877 int     hammer_ip_resolve_data(hammer_cursor_t cursor);
 878 int     hammer_ip_delete_record(hammer_cursor_t cursor, hammer_inode_t ip,
 879                         hammer_tid_t tid);
 880 int     hammer_delete_at_cursor(hammer_cursor_t cursor, int delete_flags,
 881                         hammer_tid_t delete_tid, u_int32_t delete_ts,
 882                         int track, int64_t *stat_bytes);
 883 int     hammer_ip_check_directory_empty(hammer_transaction_t trans,
 884                         hammer_inode_t ip);
 885 int     hammer_sync_hmp(hammer_mount_t hmp, int waitfor);
 886 int     hammer_queue_inodes_flusher(hammer_mount_t hmp, int waitfor);
 887
 888 hammer_record_t
 889         hammer_alloc_mem_record(hammer_inode_t ip, int data_len);
 890 void    hammer_flush_record_done(hammer_record_t record, int error);
 891 void    hammer_wait_mem_record_ident(hammer_record_t record, const char *ident);
 892 void    hammer_rel_mem_record(hammer_record_t record);
 893
 894 int     hammer_cursor_up(hammer_cursor_t cursor);
 895 int     hammer_cursor_up_locked(hammer_cursor_t cursor);
 896 int     hammer_cursor_down(hammer_cursor_t cursor);
 897 int     hammer_cursor_upgrade(hammer_cursor_t cursor);
 898 int     hammer_cursor_upgrade_node(hammer_cursor_t cursor);
 899 void    hammer_cursor_downgrade(hammer_cursor_t cursor);
 900 int     hammer_cursor_seek(hammer_cursor_t cursor, hammer_node_t node,
 901                         int index);
 902 void    hammer_lock_ex_ident(struct hammer_lock *lock, const char *ident);
 903 int     hammer_lock_ex_try(struct hammer_lock *lock);
 904 void    hammer_lock_sh(struct hammer_lock *lock);
 905 int     hammer_lock_sh_try(struct hammer_lock *lock);
 906 int     hammer_lock_upgrade(struct hammer_lock *lock);
 907 void    hammer_lock_downgrade(struct hammer_lock *lock);
 908 int     hammer_lock_status(struct hammer_lock *lock);
 909 void    hammer_unlock(struct hammer_lock *lock);
 910 void    hammer_ref(struct hammer_lock *lock);
 911 void    hammer_unref(struct hammer_lock *lock);
 912
 913 void    hammer_sync_lock_ex(hammer_transaction_t trans);
 914 void    hammer_sync_lock_sh(hammer_transaction_t trans);
 915 int     hammer_sync_lock_sh_try(hammer_transaction_t trans);
 916 void    hammer_sync_unlock(hammer_transaction_t trans);
 917
 918 u_int32_t hammer_to_unix_xid(uuid_t *uuid);
 919 void hammer_guid_to_uuid(uuid_t *uuid, u_int32_t guid);
 920 void    hammer_time_to_timespec(u_int64_t xtime, struct timespec *ts);
 921 u_int64_t hammer_timespec_to_time(struct timespec *ts);
 922 int     hammer_str_to_tid(const char *str, int *ispfsp,
 923                         hammer_tid_t *tidp, u_int32_t *localizationp);
 924 int     hammer_is_atatext(const char *name, int len);
 925 hammer_tid_t hammer_alloc_objid(hammer_mount_t hmp, hammer_inode_t dip);
 926 void hammer_clear_objid(hammer_inode_t dip);
 927 void hammer_destroy_objid_cache(hammer_mount_t hmp);
 928
 929 int hammer_enter_undo_history(hammer_mount_t hmp, hammer_off_t offset,
 930                         int bytes);
 931 void hammer_clear_undo_history(hammer_mount_t hmp);
 932 enum vtype hammer_get_vnode_type(u_int8_t obj_type);
 933 int hammer_get_dtype(u_int8_t obj_type);
 934 u_int8_t hammer_get_obj_type(enum vtype vtype);
 935 int64_t hammer_directory_namekey(hammer_inode_t dip, const void *name, int len,
 936                         u_int32_t *max_iterationsp);
 937 int     hammer_nohistory(hammer_inode_t ip);
 938
 939 int     hammer_init_cursor(hammer_transaction_t trans, hammer_cursor_t cursor,
 940                         hammer_node_cache_t cache, hammer_inode_t ip);
 941 void    hammer_normalize_cursor(hammer_cursor_t cursor);
 942 void    hammer_done_cursor(hammer_cursor_t cursor);
 943 int     hammer_recover_cursor(hammer_cursor_t cursor);
 944 void    hammer_unlock_cursor(hammer_cursor_t cursor);
 945 int     hammer_lock_cursor(hammer_cursor_t cursor);
 946 hammer_cursor_t hammer_push_cursor(hammer_cursor_t ocursor);
 947 void    hammer_pop_cursor(hammer_cursor_t ocursor, hammer_cursor_t ncursor);
 948
 949 void    hammer_cursor_replaced_node(hammer_node_t onode, hammer_node_t nnode);
 950 void    hammer_cursor_removed_node(hammer_node_t onode, hammer_node_t parent,
 951                         int index);
 952 void    hammer_cursor_split_node(hammer_node_t onode, hammer_node_t nnode,
 953                         int index);
 954 void    hammer_cursor_moved_element(hammer_node_t onode, hammer_node_t nnode,
 955                         int oindex, int nindex);
 956 void    hammer_cursor_parent_changed(hammer_node_t node, hammer_node_t oparent,
 957                         hammer_node_t nparent, int nindex);
 958 void    hammer_cursor_inserted_element(hammer_node_t node, int index);
 959 void    hammer_cursor_deleted_element(hammer_node_t node, int index);
 960
 961 int     hammer_btree_lookup(hammer_cursor_t cursor);
 962 int     hammer_btree_first(hammer_cursor_t cursor);
 963 int     hammer_btree_last(hammer_cursor_t cursor);
 964 int     hammer_btree_extract(hammer_cursor_t cursor, int flags);
 965 int     hammer_btree_iterate(hammer_cursor_t cursor);
 966 int     hammer_btree_iterate_reverse(hammer_cursor_t cursor);
 967 int     hammer_btree_insert(hammer_cursor_t cursor,
 968                             hammer_btree_leaf_elm_t elm, int *doprop);
 969 int     hammer_btree_delete(hammer_cursor_t cursor);
 970 void    hammer_btree_do_propagation(hammer_cursor_t cursor,
 971                             hammer_pseudofs_inmem_t pfsm,
 972                             hammer_btree_leaf_elm_t leaf);
 973 int     hammer_btree_cmp(hammer_base_elm_t key1, hammer_base_elm_t key2);
 974 int     hammer_btree_chkts(hammer_tid_t ts, hammer_base_elm_t key);
 975 int     hammer_btree_correct_rhb(hammer_cursor_t cursor, hammer_tid_t tid);
 976 int     hammer_btree_correct_lhb(hammer_cursor_t cursor, hammer_tid_t tid);
 977
 978 int     btree_set_parent(hammer_transaction_t trans, hammer_node_t node,
 979                         hammer_btree_elm_t elm);
 980 void    hammer_node_lock_init(hammer_node_lock_t parent, hammer_node_t node);
 981 int     hammer_btree_lock_children(hammer_cursor_t cursor, int depth,
 982                         hammer_node_lock_t parent);
 983 void    hammer_btree_lock_copy(hammer_cursor_t cursor,
 984                         hammer_node_lock_t parent);
 985 void    hammer_btree_sync_copy(hammer_cursor_t cursor,
 986                         hammer_node_lock_t parent);
 987 void    hammer_btree_unlock_children(hammer_cursor_t cursor,
 988                         hammer_node_lock_t parent);
 989 int     hammer_btree_search_node(hammer_base_elm_t elm, hammer_node_ondisk_t node);
 990 hammer_node_t hammer_btree_get_parent(hammer_transaction_t trans,
 991                         hammer_node_t node, int *parent_indexp,
 992                         int *errorp, int try_exclusive);
 993
 994 void    hammer_print_btree_node(hammer_node_ondisk_t ondisk);
 995 void    hammer_print_btree_elm(hammer_btree_elm_t elm, u_int8_t type, int i);
 996
 997 void    *hammer_bread(struct hammer_mount *hmp, hammer_off_t off,
 998                         int *errorp, struct hammer_buffer **bufferp);
 999 void    *hammer_bnew(struct hammer_mount *hmp, hammer_off_t off,
1000                         int *errorp, struct hammer_buffer **bufferp);
1001 void    *hammer_bread_ext(struct hammer_mount *hmp, hammer_off_t off, int bytes,
1002                         int *errorp, struct hammer_buffer **bufferp);
1003 void    *hammer_bnew_ext(struct hammer_mount *hmp, hammer_off_t off, int bytes,
1004                         int *errorp, struct hammer_buffer **bufferp);
1005
1006 hammer_volume_t hammer_get_root_volume(hammer_mount_t hmp, int *errorp);
1007
1008 hammer_volume_t hammer_get_volume(hammer_mount_t hmp,
1009                         int32_t vol_no, int *errorp);
1010 hammer_buffer_t hammer_get_buffer(hammer_mount_t hmp, hammer_off_t buf_offset,
1011                         int bytes, int isnew, int *errorp);
1012 void            hammer_sync_buffers(hammer_mount_t hmp,
1013                         hammer_off_t base_offset, int bytes);
1014 int             hammer_del_buffers(hammer_mount_t hmp,
1015                         hammer_off_t base_offset,
1016                         hammer_off_t zone2_offset, int bytes,
1017                         int report_conflicts);
1018
1019 int             hammer_ref_volume(hammer_volume_t volume);
1020 int             hammer_ref_buffer(hammer_buffer_t buffer);
1021 void            hammer_flush_buffer_nodes(hammer_buffer_t buffer);
1022
1023 void            hammer_rel_volume(hammer_volume_t volume, int flush);
1024 void            hammer_rel_buffer(hammer_buffer_t buffer, int flush);
1025
1026 int             hammer_vfs_export(struct mount *mp, int op,
1027                         const struct export_args *export);
1028 hammer_node_t   hammer_get_node(hammer_transaction_t trans,
1029                         hammer_off_t node_offset, int isnew, int *errorp);
1030 void            hammer_ref_node(hammer_node_t node);
1031 hammer_node_t   hammer_ref_node_safe(struct hammer_mount *hmp,
1032                         hammer_node_cache_t cache, int *errorp);
1033 void            hammer_rel_node(hammer_node_t node);
1034 void            hammer_delete_node(hammer_transaction_t trans,
1035                         hammer_node_t node);
1036 void            hammer_cache_node(hammer_node_cache_t cache,
1037                         hammer_node_t node);
1038 void            hammer_uncache_node(hammer_node_cache_t cache);
1039 void            hammer_flush_node(hammer_node_t node);
1040
1041 void hammer_dup_buffer(struct hammer_buffer **bufferp,
1042                         struct hammer_buffer *buffer);
1043 hammer_node_t hammer_alloc_btree(hammer_transaction_t trans, int *errorp);
1044 void *hammer_alloc_data(hammer_transaction_t trans, int32_t data_len,
1045                         u_int16_t rec_type, hammer_off_t *data_offsetp,
1046                         struct hammer_buffer **data_bufferp, int *errorp);
1047
1048 int hammer_generate_undo(hammer_transaction_t trans, hammer_io_t io,
1049                         hammer_off_t zone1_offset, void *base, int len);
1050
1051 void hammer_put_volume(struct hammer_volume *volume, int flush);
1052 void hammer_put_buffer(struct hammer_buffer *buffer, int flush);
1053
1054 hammer_off_t hammer_freemap_alloc(hammer_transaction_t trans,
1055                         hammer_off_t owner, int *errorp);
1056 void hammer_freemap_free(hammer_transaction_t trans, hammer_off_t phys_offset,
1057                         hammer_off_t owner, int *errorp);
1058 int hammer_checkspace(hammer_mount_t hmp, int slop);
1059 hammer_off_t hammer_blockmap_alloc(hammer_transaction_t trans, int zone,
1060                         int bytes, int *errorp);
1061 hammer_reserve_t hammer_blockmap_reserve(hammer_mount_t hmp, int zone,
1062                         int bytes, hammer_off_t *zone_offp, int *errorp);
1063 void hammer_blockmap_reserve_undo(hammer_mount_t hmp, hammer_reserve_t resv,
1064                         hammer_off_t zone_offset, int bytes);
1065 void hammer_blockmap_reserve_complete(hammer_mount_t hmp,
1066                         hammer_reserve_t resv);
1067 void hammer_reserve_clrdelay(hammer_mount_t hmp, hammer_reserve_t resv);
1068 void hammer_blockmap_free(hammer_transaction_t trans,
1069                         hammer_off_t bmap_off, int bytes);
1070 int hammer_blockmap_finalize(hammer_transaction_t trans,
1071                         hammer_reserve_t resv,
1072                         hammer_off_t bmap_off, int bytes);
1073 int hammer_blockmap_getfree(hammer_mount_t hmp, hammer_off_t bmap_off,
1074                         int *curp, int *errorp);
1075 hammer_off_t hammer_blockmap_lookup(hammer_mount_t hmp, hammer_off_t bmap_off,
1076                         int *errorp);
1077 hammer_off_t hammer_undo_lookup(hammer_mount_t hmp, hammer_off_t bmap_off,
1078                         int *errorp);
1079 int64_t hammer_undo_used(hammer_transaction_t trans);
1080 int64_t hammer_undo_space(hammer_transaction_t trans);
1081 int64_t hammer_undo_max(hammer_mount_t hmp);
1082
1083 void hammer_start_transaction(struct hammer_transaction *trans,
1084                               struct hammer_mount *hmp);
1085 void hammer_simple_transaction(struct hammer_transaction *trans,
1086                               struct hammer_mount *hmp);
1087 void hammer_start_transaction_fls(struct hammer_transaction *trans,
1088                                   struct hammer_mount *hmp);
1089 void hammer_done_transaction(struct hammer_transaction *trans);
1090
1091 void hammer_modify_inode(hammer_inode_t ip, int flags);
1092 void hammer_flush_inode(hammer_inode_t ip, int flags);
1093 void hammer_flush_inode_done(hammer_inode_t ip, int error);
1094 void hammer_wait_inode(hammer_inode_t ip);
1095
1096 int  hammer_create_inode(struct hammer_transaction *trans, struct vattr *vap,
1097                         struct ucred *cred, struct hammer_inode *dip,
1098                         hammer_pseudofs_inmem_t pfsm,
1099                         struct hammer_inode **ipp);
1100 void hammer_rel_inode(hammer_inode_t ip, int flush);
1101 int hammer_reload_inode(hammer_inode_t ip, void *arg __unused);
1102 int hammer_ino_rb_compare(hammer_inode_t ip1, hammer_inode_t ip2);
1103 int hammer_destroy_inode_callback(hammer_inode_t ip, void *data __unused);
1104
1105 int hammer_sync_inode(hammer_transaction_t trans, hammer_inode_t ip);
1106 void hammer_test_inode(hammer_inode_t dip);
1107 void hammer_inode_unloadable_check(hammer_inode_t ip, int getvp);
1108
1109 int  hammer_ip_add_directory(struct hammer_transaction *trans,
1110                         hammer_inode_t dip, const char *name, int bytes,
1111                         hammer_inode_t nip);
1112 int  hammer_ip_del_directory(struct hammer_transaction *trans,
1113                         hammer_cursor_t cursor, hammer_inode_t dip,
1114                         hammer_inode_t ip);
1115 hammer_record_t hammer_ip_add_bulk(hammer_inode_t ip, off_t file_offset,
1116                         void *data, int bytes, int *errorp);
1117 int  hammer_ip_frontend_trunc(struct hammer_inode *ip, off_t file_size);
1118 int  hammer_ip_add_record(struct hammer_transaction *trans,
1119                         hammer_record_t record);
1120 int  hammer_ip_delete_range(hammer_cursor_t cursor, hammer_inode_t ip,
1121                         int64_t ran_beg, int64_t ran_end, int truncating);
1122 int  hammer_ip_delete_clean(hammer_cursor_t cursor, hammer_inode_t ip,
1123                         int *countp);
1124 int  hammer_ip_sync_data(hammer_cursor_t cursor, hammer_inode_t ip,
1125                         int64_t offset, void *data, int bytes);
1126 int  hammer_ip_sync_record(hammer_transaction_t trans, hammer_record_t rec);
1127 int  hammer_ip_sync_record_cursor(hammer_cursor_t cursor, hammer_record_t rec);
1128 hammer_pseudofs_inmem_t  hammer_load_pseudofs(hammer_transaction_t trans,
1129                         u_int32_t localization, int *errorp);
1130 int  hammer_mkroot_pseudofs(hammer_transaction_t trans, struct ucred *cred,
1131                         hammer_pseudofs_inmem_t pfsm);
1132 int  hammer_save_pseudofs(hammer_transaction_t trans,
1133                         hammer_pseudofs_inmem_t pfsm);
1134 int  hammer_unload_pseudofs(hammer_transaction_t trans, u_int32_t localization);
1135 void hammer_rel_pseudofs(hammer_mount_t hmp, hammer_pseudofs_inmem_t pfsm);
1136 int hammer_ioctl(hammer_inode_t ip, u_long com, caddr_t data, int fflag,
1137                         struct ucred *cred);
1138
1139 void hammer_io_init(hammer_io_t io, hammer_volume_t volume,
1140                         enum hammer_io_type type);
1141 int hammer_io_read(struct vnode *devvp, struct hammer_io *io,
1142                         hammer_off_t limit);
1143 int hammer_io_new(struct vnode *devvp, struct hammer_io *io);
1144 int hammer_io_inval(hammer_volume_t volume, hammer_off_t zone2_offset);
1145 struct buf *hammer_io_release(struct hammer_io *io, int flush);
1146 void hammer_io_flush(struct hammer_io *io);
1147 void hammer_io_wait(struct hammer_io *io);
1148 void hammer_io_waitdep(struct hammer_io *io);
1149 void hammer_io_wait_all(hammer_mount_t hmp, const char *ident);
1150 int hammer_io_direct_read(hammer_mount_t hmp, struct bio *bio,
1151                         hammer_btree_leaf_elm_t leaf);
1152 int hammer_io_direct_write(hammer_mount_t hmp, hammer_record_t record,
1153                         struct bio *bio);
1154 void hammer_io_direct_wait(hammer_record_t record);
1155 void hammer_io_direct_uncache(hammer_mount_t hmp, hammer_btree_leaf_elm_t leaf);
1156 void hammer_io_write_interlock(hammer_io_t io);
1157 void hammer_io_done_interlock(hammer_io_t io);
1158 void hammer_io_clear_modify(struct hammer_io *io, int inval);
1159 void hammer_io_clear_modlist(struct hammer_io *io);
1160 void hammer_io_flush_sync(hammer_mount_t hmp);
1161
1162 void hammer_modify_volume(hammer_transaction_t trans, hammer_volume_t volume,
1163                         void *base, int len);
1164 void hammer_modify_buffer(hammer_transaction_t trans, hammer_buffer_t buffer,
1165                         void *base, int len);
1166 void hammer_modify_volume_done(hammer_volume_t volume);
1167 void hammer_modify_buffer_done(hammer_buffer_t buffer);
1168
1169 int hammer_ioc_reblock(hammer_transaction_t trans, hammer_inode_t ip,
1170                         struct hammer_ioc_reblock *reblock);
1171 int hammer_ioc_rebalance(hammer_transaction_t trans, hammer_inode_t ip,
1172                         struct hammer_ioc_rebalance *rebal);
1173 int hammer_ioc_prune(hammer_transaction_t trans, hammer_inode_t ip,
1174                         struct hammer_ioc_prune *prune);
1175 int hammer_ioc_mirror_read(hammer_transaction_t trans, hammer_inode_t ip,
1176                         struct hammer_ioc_mirror_rw *mirror);
1177 int hammer_ioc_mirror_write(hammer_transaction_t trans, hammer_inode_t ip,
1178                         struct hammer_ioc_mirror_rw *mirror);
1179 int hammer_ioc_set_pseudofs(hammer_transaction_t trans, hammer_inode_t ip,
1180                         struct ucred *cred, struct hammer_ioc_pseudofs_rw *pfs);
1181 int hammer_ioc_get_pseudofs(hammer_transaction_t trans, hammer_inode_t ip,
1182                         struct hammer_ioc_pseudofs_rw *pfs);
1183 int hammer_ioc_destroy_pseudofs(hammer_transaction_t trans, hammer_inode_t ip,
1184                         struct hammer_ioc_pseudofs_rw *pfs);
1185 int hammer_ioc_downgrade_pseudofs(hammer_transaction_t trans, hammer_inode_t ip,
1186                         struct hammer_ioc_pseudofs_rw *pfs);
1187 int hammer_ioc_upgrade_pseudofs(hammer_transaction_t trans, hammer_inode_t ip,
1188                         struct hammer_ioc_pseudofs_rw *pfs);
1189 int hammer_ioc_wait_pseudofs(hammer_transaction_t trans, hammer_inode_t ip,
1190                         struct hammer_ioc_pseudofs_rw *pfs);
1191
1192 int hammer_signal_check(hammer_mount_t hmp);
1193
1194 void hammer_flusher_create(hammer_mount_t hmp);
1195 void hammer_flusher_destroy(hammer_mount_t hmp);
1196 void hammer_flusher_sync(hammer_mount_t hmp);
1197 int  hammer_flusher_async(hammer_mount_t hmp, hammer_flush_group_t flg);
1198 int  hammer_flusher_async_one(hammer_mount_t hmp);
1199 void hammer_flusher_wait(hammer_mount_t hmp, int seq);
1200 void hammer_flusher_wait_next(hammer_mount_t hmp);
1201 int  hammer_flusher_meta_limit(hammer_mount_t hmp);
1202 int  hammer_flusher_meta_halflimit(hammer_mount_t hmp);
1203 int  hammer_flusher_undo_exhausted(hammer_transaction_t trans, int quarter);
1204 void hammer_flusher_clean_loose_ios(hammer_mount_t hmp);
1205 void hammer_flusher_finalize(hammer_transaction_t trans, int final);
1206 int  hammer_flusher_haswork(hammer_mount_t hmp);
1207
1208
1209 int hammer_recover(hammer_mount_t hmp, hammer_volume_t rootvol);
1210 void hammer_recover_flush_buffers(hammer_mount_t hmp,
1211                         hammer_volume_t root_volume, int final);
1212
1213 void hammer_crc_set_blockmap(hammer_blockmap_t blockmap);
1214 void hammer_crc_set_volume(hammer_volume_ondisk_t ondisk);
1215 void hammer_crc_set_leaf(void *data, hammer_btree_leaf_elm_t leaf);
1216
1217 int hammer_crc_test_blockmap(hammer_blockmap_t blockmap);
1218 int hammer_crc_test_volume(hammer_volume_ondisk_t ondisk);
1219 int hammer_crc_test_btree(hammer_node_ondisk_t ondisk);
1220 int hammer_crc_test_leaf(void *data, hammer_btree_leaf_elm_t leaf);
1221 void hkprintf(const char *ctl, ...);
1222 udev_t hammer_fsid_to_udev(uuid_t *uuid);
1223
1224
1225 int hammer_blocksize(int64_t file_offset);
1226 int64_t hammer_blockdemarc(int64_t file_offset1, int64_t file_offset2);
1227
1228 #endif
1229
1230 static __inline void
1231 hammer_wait_mem_record(hammer_record_t record)
1232 {
1233         hammer_wait_mem_record_ident(record, "hmmwai");
1234 }
1235
1236 static __inline void
1237 hammer_lock_ex(struct hammer_lock *lock)
1238 {
1239         hammer_lock_ex_ident(lock, "hmrlck");
1240 }
1241
1242 /*
1243  * Indicate that a B-Tree node is being modified.
1244  */
1245 static __inline void
1246 hammer_modify_node_noundo(hammer_transaction_t trans, hammer_node_t node)
1247 {
1248         hammer_modify_buffer(trans, node->buffer, NULL, 0);
1249 }
1250
1251 static __inline void
1252 hammer_modify_node_all(hammer_transaction_t trans, struct hammer_node *node)
1253 {
1254         hammer_modify_buffer(trans, node->buffer,
1255                              node->ondisk, sizeof(*node->ondisk));
1256 }
1257
1258 static __inline void
1259 hammer_modify_node(hammer_transaction_t trans, hammer_node_t node,
1260                    void *base, int len)
1261 {
1262         hammer_crc_t *crcptr;
1263
1264         KKASSERT((char *)base >= (char *)node->ondisk &&
1265                  (char *)base + len <=
1266                     (char *)node->ondisk + sizeof(*node->ondisk));
1267         hammer_modify_buffer(trans, node->buffer, base, len);
1268         crcptr = &node->ondisk->crc;
1269         hammer_modify_buffer(trans, node->buffer, crcptr, sizeof(hammer_crc_t));
1270         --node->buffer->io.modify_refs; /* only want one ref */
1271 }
1272
1273 /*
1274  * Indicate that the specified modifications have been completed.
1275  *
1276  * Do not try to generate the crc here, it's very expensive to do and a
1277  * sequence of insertions or deletions can result in many calls to this
1278  * function on the same node.
1279  */
1280 static __inline void
1281 hammer_modify_node_done(hammer_node_t node)
1282 {
1283         node->flags |= HAMMER_NODE_CRCGOOD;
1284         if ((node->flags & HAMMER_NODE_NEEDSCRC) == 0) {
1285                 node->flags |= HAMMER_NODE_NEEDSCRC;
1286                 node->buffer->io.gencrc = 1;
1287                 hammer_ref_node(node);
1288         }
1289         hammer_modify_buffer_done(node->buffer);
1290 }
1291
1292 #define hammer_modify_volume_field(trans, vol, field)           \
1293         hammer_modify_volume(trans, vol, &(vol)->ondisk->field, \
1294                              sizeof((vol)->ondisk->field))
1295
1296 #define hammer_modify_node_field(trans, node, field)            \
1297         hammer_modify_node(trans, node, &(node)->ondisk->field, \
1298                              sizeof((node)->ondisk->field))
1299