sys/vfs/hammer/hammer.h

   1 /*
   2  * Copyright (c) 2007-2008 The DragonFly Project.  All rights reserved.
   3  *
   4  * This code is derived from software contributed to The DragonFly Project
   5  * by Matthew Dillon <dillon@backplane.com>
   6  *
   7  * Redistribution and use in source and binary forms, with or without
   8  * modification, are permitted provided that the following conditions
   9  * are met:
  10  *
  11  * 1. Redistributions of source code must retain the above copyright
  12  *    notice, this list of conditions and the following disclaimer.
  13  * 2. Redistributions in binary form must reproduce the above copyright
  14  *    notice, this list of conditions and the following disclaimer in
  15  *    the documentation and/or other materials provided with the
  16  *    distribution.
  17  * 3. Neither the name of The DragonFly Project nor the names of its
  18  *    contributors may be used to endorse or promote products derived
  19  *    from this software without specific, prior written permission.
  20  *
  21  * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
  22  * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
  23  * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS
  24  * FOR A PARTICULAR PURPOSE ARE DISCLAIMED.  IN NO EVENT SHALL THE
  25  * COPYRIGHT HOLDERS OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT,
  26  * INCIDENTAL, SPECIAL, EXEMPLARY OR CONSEQUENTIAL DAMAGES (INCLUDING,
  27  * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
  28  * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED
  29  * AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
  30  * OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT
  31  * OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
  32  * SUCH DAMAGE.
  33  *
  34  * $DragonFly: src/sys/vfs/hammer/hammer.h,v 1.130 2008/11/13 02:18:43 dillon Exp $
  35  */
  36 /*
  37  * This header file contains structures used internally by the HAMMERFS
  38  * implementation.  See hammer_disk.h for on-disk structures.
  39  */
  40
  41 #include <sys/param.h>
  42 #include <sys/types.h>
  43 #include <sys/kernel.h>
  44 #include <sys/conf.h>
  45 #include <sys/systm.h>
  46 #include <sys/tree.h>
  47 #include <sys/malloc.h>
  48 #include <sys/mount.h>
  49 #include <sys/mountctl.h>
  50 #include <sys/vnode.h>
  51 #include <sys/proc.h>
  52 #include <sys/priv.h>
  53 #include <sys/stat.h>
  54 #include <sys/globaldata.h>
  55 #include <sys/lockf.h>
  56 #include <sys/buf.h>
  57 #include <sys/queue.h>
  58 #include <sys/ktr.h>
  59 #include <sys/globaldata.h>
  60
  61 #include <sys/buf2.h>
  62 #include <sys/signal2.h>
  63 #include "hammer_disk.h"
  64 #include "hammer_mount.h"
  65 #include "hammer_ioctl.h"
  66
  67 #if defined(_KERNEL) || defined(_KERNEL_STRUCTURES)
  68
  69 MALLOC_DECLARE(M_HAMMER);
  70
  71 /*
  72  * Kernel trace
  73  */
  74 #if !defined(KTR_HAMMER)
  75 #define KTR_HAMMER      KTR_ALL
  76 #endif
  77 KTR_INFO_MASTER_EXTERN(hammer);
  78
  79 /*
  80  * Misc structures
  81  */
  82 struct hammer_mount;
  83
  84 /*
  85  * Key structure used for custom RB tree inode lookups.  This prototypes
  86  * the function hammer_ino_rb_tree_RB_LOOKUP_INFO(root, info).
  87  */
  88 typedef struct hammer_inode_info {
  89         int64_t         obj_id;         /* (key) object identifier */
  90         hammer_tid_t    obj_asof;       /* (key) snapshot transid or 0 */
  91         u_int32_t       obj_localization; /* (key) pseudo-fs */
  92         union {
  93                 struct hammer_btree_leaf_elm *leaf;
  94         } u;
  95 } *hammer_inode_info_t;
  96
  97 typedef enum hammer_transaction_type {
  98         HAMMER_TRANS_RO,
  99         HAMMER_TRANS_STD,
 100         HAMMER_TRANS_FLS
 101 } hammer_transaction_type_t;
 102
 103 /*
 104  * HAMMER Transaction tracking
 105  */
 106 struct hammer_transaction {
 107         hammer_transaction_type_t type;
 108         struct hammer_mount *hmp;
 109         hammer_tid_t    tid;
 110         u_int64_t       time;
 111         u_int32_t       time32;
 112         int             sync_lock_refs;
 113         int             flags;
 114         struct hammer_volume *rootvol;
 115 };
 116
 117 typedef struct hammer_transaction *hammer_transaction_t;
 118
 119 #define HAMMER_TRANSF_NEWINODE  0x0001
 120 #define HAMMER_TRANSF_DIDIO     0x0002
 121 #define HAMMER_TRANSF_CRCDOM    0x0004  /* EDOM on CRC error, less critical */
 122
 123 /*
 124  * HAMMER locks
 125  */
 126 struct hammer_lock {
 127         int     refs;           /* active references delay writes */
 128         int     lockcount;      /* lock count for exclusive/shared access */
 129         int     wanted;
 130         int     exwanted;       /* number of threads waiting for ex lock */
 131         struct thread *locktd;
 132 };
 133
 134 static __inline int
 135 hammer_islocked(struct hammer_lock *lock)
 136 {
 137         return(lock->lockcount != 0);
 138 }
 139
 140 static __inline int
 141 hammer_isactive(struct hammer_lock *lock)
 142 {
 143         return(lock->refs != 0);
 144 }
 145
 146 static __inline int
 147 hammer_islastref(struct hammer_lock *lock)
 148 {
 149         return(lock->refs == 1);
 150 }
 151
 152 /*
 153  * Return if we specifically own the lock exclusively.
 154  */
 155 static __inline int
 156 hammer_lock_excl_owned(struct hammer_lock *lock, thread_t td)
 157 {
 158         if (lock->lockcount > 0 && lock->locktd == td)
 159                 return(1);
 160         return(0);
 161 }
 162
 163 /*
 164  * Flush state, used by various structures
 165  */
 166 typedef enum hammer_inode_state {
 167         HAMMER_FST_IDLE,
 168         HAMMER_FST_SETUP,
 169         HAMMER_FST_FLUSH
 170 } hammer_inode_state_t;
 171
 172 TAILQ_HEAD(hammer_record_list, hammer_record);
 173
 174 /*
 175  * Pseudo-filesystem extended data tracking
 176  */
 177 struct hammer_pfs_rb_tree;
 178 struct hammer_pseudofs_inmem;
 179 RB_HEAD(hammer_pfs_rb_tree, hammer_pseudofs_inmem);
 180 RB_PROTOTYPE2(hammer_pfs_rb_tree, hammer_pseudofs_inmem, rb_node,
 181               hammer_pfs_rb_compare, u_int32_t);
 182
 183 struct hammer_pseudofs_inmem {
 184         RB_ENTRY(hammer_pseudofs_inmem) rb_node;
 185         struct hammer_lock      lock;
 186         u_int32_t               localization;
 187         hammer_tid_t            create_tid;
 188         int                     flags;
 189         udev_t                  fsid_udev;
 190         struct hammer_pseudofs_data pfsd;
 191 };
 192
 193 typedef struct hammer_pseudofs_inmem *hammer_pseudofs_inmem_t;
 194
 195 #define HAMMER_PFSM_DELETED     0x0001
 196
 197 /*
 198  * Cache object ids.  A fixed number of objid cache structures are
 199  * created to reserve object id's for newly created files in multiples
 200  * of 100,000, localized to a particular directory, and recycled as
 201  * needed.  This allows parallel create operations in different
 202  * directories to retain fairly localized object ids which in turn
 203  * improves reblocking performance and layout.
 204  */
 205 #define OBJID_CACHE_SIZE        1024
 206 #define OBJID_CACHE_BULK_BITS   10              /* 10 bits (1024)       */
 207 #define OBJID_CACHE_BULK        (32 * 32)       /* two level (1024)     */
 208 #define OBJID_CACHE_BULK_MASK   (OBJID_CACHE_BULK - 1)
 209 #define OBJID_CACHE_BULK_MASK64 ((u_int64_t)(OBJID_CACHE_BULK - 1))
 210
 211 typedef struct hammer_objid_cache {
 212         TAILQ_ENTRY(hammer_objid_cache) entry;
 213         struct hammer_inode             *dip;
 214         hammer_tid_t                    base_tid;
 215         int                             count;
 216         u_int32_t                       bm0;
 217         u_int32_t                       bm1[32];
 218 } *hammer_objid_cache_t;
 219
 220 /*
 221  * Associate an inode with a B-Tree node to cache search start positions
 222  */
 223 typedef struct hammer_node_cache {
 224         TAILQ_ENTRY(hammer_node_cache)  entry;
 225         struct hammer_node              *node;
 226         struct hammer_inode             *ip;
 227 } *hammer_node_cache_t;
 228
 229 TAILQ_HEAD(hammer_node_cache_list, hammer_node_cache);
 230
 231 /*
 232  * Structure used to organize flush groups.  Flush groups must be
 233  * organized into chunks in order to avoid blowing out the UNDO FIFO.
 234  * Without this a 'sync' could end up flushing 50,000 inodes in a single
 235  * transaction.
 236  */
 237 struct hammer_flush_group {
 238         TAILQ_ENTRY(hammer_flush_group) flush_entry;
 239         TAILQ_HEAD(, hammer_inode)      flush_list;
 240         int                             unused01;       /* inode load */
 241         int                             total_count;    /* record load */
 242         int                             running;        /* group is running */
 243         int                             closed;
 244         int                             refs;
 245 };
 246
 247 typedef struct hammer_flush_group *hammer_flush_group_t;
 248
 249 TAILQ_HEAD(hammer_flush_group_list, hammer_flush_group);
 250
 251 /*
 252  * Structure used to represent an inode in-memory.
 253  *
 254  * The record and data associated with an inode may be out of sync with
 255  * the disk (xDIRTY flags), or not even on the disk at all (ONDISK flag
 256  * clear).
 257  *
 258  * An inode may also hold a cache of unsynchronized records, used for
 259  * database and directories only.  Unsynchronized regular file data is
 260  * stored in the buffer cache.
 261  *
 262  * NOTE: A file which is created and destroyed within the initial
 263  * synchronization period can wind up not doing any disk I/O at all.
 264  *
 265  * Finally, an inode may cache numerous disk-referencing B-Tree cursors.
 266  */
 267 struct hammer_ino_rb_tree;
 268 struct hammer_inode;
 269 RB_HEAD(hammer_ino_rb_tree, hammer_inode);
 270 RB_PROTOTYPEX(hammer_ino_rb_tree, INFO, hammer_inode, rb_node,
 271               hammer_ino_rb_compare, hammer_inode_info_t);
 272
 273 struct hammer_rec_rb_tree;
 274 struct hammer_record;
 275 RB_HEAD(hammer_rec_rb_tree, hammer_record);
 276 RB_PROTOTYPEX(hammer_rec_rb_tree, INFO, hammer_record, rb_node,
 277               hammer_rec_rb_compare, hammer_btree_leaf_elm_t);
 278
 279 TAILQ_HEAD(hammer_node_list, hammer_node);
 280
 281 struct hammer_inode {
 282         RB_ENTRY(hammer_inode)  rb_node;
 283         hammer_inode_state_t    flush_state;
 284         hammer_flush_group_t    flush_group;
 285         TAILQ_ENTRY(hammer_inode) flush_entry;
 286         struct hammer_record_list target_list;  /* target of dependant recs */
 287         int64_t                 obj_id;         /* (key) object identifier */
 288         hammer_tid_t            obj_asof;       /* (key) snapshot or 0 */
 289         u_int32_t               obj_localization; /* (key) pseudo-fs */
 290         struct hammer_mount     *hmp;
 291         hammer_objid_cache_t    objid_cache;
 292         int                     flags;
 293         int                     error;          /* flush error */
 294         int                     cursor_ip_refs; /* sanity */
 295         int                     rsv_recs;
 296         struct vnode            *vp;
 297         hammer_pseudofs_inmem_t pfsm;
 298         struct lockf            advlock;
 299         struct hammer_lock      lock;           /* sync copy interlock */
 300         off_t                   trunc_off;
 301         struct hammer_btree_leaf_elm ino_leaf;  /* in-memory cache */
 302         struct hammer_inode_data ino_data;      /* in-memory cache */
 303         struct hammer_rec_rb_tree rec_tree;     /* in-memory cache */
 304         int                     rec_generation;
 305         struct hammer_node_cache cache[4];      /* search initiate cache */
 306
 307         /*
 308          * When a demark is created to synchronize an inode to
 309          * disk, certain fields are copied so the front-end VOPs
 310          * can continue to run in parallel with the synchronization
 311          * occuring in the background.
 312          */
 313         int             sync_flags;             /* to-sync flags cache */
 314         off_t           sync_trunc_off;         /* to-sync truncation */
 315         off_t           save_trunc_off;         /* write optimization */
 316         struct hammer_btree_leaf_elm sync_ino_leaf; /* to-sync cache */
 317         struct hammer_inode_data sync_ino_data; /* to-sync cache */
 318 };
 319
 320 typedef struct hammer_inode *hammer_inode_t;
 321
 322 #define VTOI(vp)        ((struct hammer_inode *)(vp)->v_data)
 323
 324 #define HAMMER_INODE_DDIRTY     0x0001  /* in-memory ino_data is dirty */
 325                                         /* (not including atime/mtime) */
 326 #define HAMMER_INODE_RSV_INODES 0x0002  /* hmp->rsv_inodes bumped */
 327 #define HAMMER_INODE_CONN_DOWN  0x0004  /* include in downward recursion */
 328 #define HAMMER_INODE_XDIRTY     0x0008  /* in-memory records */
 329 #define HAMMER_INODE_ONDISK     0x0010  /* inode is on-disk (else not yet) */
 330 #define HAMMER_INODE_FLUSH      0x0020  /* flush on last ref */
 331 #define HAMMER_INODE_DELETED    0x0080  /* inode delete (backend) */
 332 #define HAMMER_INODE_DELONDISK  0x0100  /* delete synchronized to disk */
 333 #define HAMMER_INODE_RO         0x0200  /* read-only (because of as-of) */
 334 #define HAMMER_INODE_VHELD      0x0400  /* vnode held on sync */
 335 #define HAMMER_INODE_DONDISK    0x0800  /* data records may be on disk */
 336 #define HAMMER_INODE_BUFS       0x1000  /* dirty high level bps present */
 337 #define HAMMER_INODE_REFLUSH    0x2000  /* flush on dependancy / reflush */
 338 #define HAMMER_INODE_RECLAIM    0x4000  /* trying to reclaim */
 339 #define HAMMER_INODE_FLUSHW     0x8000  /* Someone waiting for flush */
 340
 341 #define HAMMER_INODE_TRUNCATED  0x00010000
 342 #define HAMMER_INODE_DELETING   0x00020000 /* inode delete request (frontend)*/
 343 #define HAMMER_INODE_RESIGNAL   0x00040000 /* re-signal on re-flush */
 344 #define HAMMER_INODE_ATIME      0x00100000 /* in-memory atime modified */
 345 #define HAMMER_INODE_MTIME      0x00200000 /* in-memory mtime modified */
 346 #define HAMMER_INODE_WOULDBLOCK 0x00400000 /* re-issue to new flush group */
 347 #define HAMMER_INODE_DUMMY      0x00800000 /* dummy inode covering bad file */
 348
 349 #define HAMMER_INODE_MODMASK    (HAMMER_INODE_DDIRTY|                       \
 350                                  HAMMER_INODE_XDIRTY|HAMMER_INODE_BUFS|     \
 351                                  HAMMER_INODE_ATIME|HAMMER_INODE_MTIME|     \
 352                                  HAMMER_INODE_TRUNCATED|HAMMER_INODE_DELETING)
 353
 354 #define HAMMER_INODE_MODMASK_NOXDIRTY \
 355                                 (HAMMER_INODE_MODMASK & ~HAMMER_INODE_XDIRTY)
 356
 357 #define HAMMER_FLUSH_GROUP_SIZE 64
 358
 359 #define HAMMER_FLUSH_SIGNAL     0x0001
 360 #define HAMMER_FLUSH_RECURSION  0x0002
 361
 362 /*
 363  * Used by the inode reclaim code to pipeline reclaims and avoid
 364  * blowing out kernel memory or letting the flusher get too far
 365  * behind.  The reclaim wakes up when count reaches 0 or the
 366  * timer expires.
 367  */
 368 struct hammer_reclaim {
 369         TAILQ_ENTRY(hammer_reclaim) entry;
 370         int     count;
 371 };
 372
 373 #define HAMMER_RECLAIM_FLUSH    2000
 374 #define HAMMER_RECLAIM_WAIT     4000
 375
 376 /*
 377  * Structure used to represent an unsynchronized record in-memory.  These
 378  * records typically represent directory entries.  Only non-historical
 379  * records are kept in-memory.
 380  *
 381  * Records are organized as a per-inode RB-Tree.  If the inode is not
 382  * on disk then neither are any records and the in-memory record tree
 383  * represents the entire contents of the inode.  If the inode is on disk
 384  * then the on-disk B-Tree is scanned in parallel with the in-memory
 385  * RB-Tree to synthesize the current state of the file.
 386  *
 387  * Records are also used to enforce the ordering of directory create/delete
 388  * operations.  A new inode will not be flushed to disk unless its related
 389  * directory entry is also being flushed at the same time.  A directory entry
 390  * will not be removed unless its related inode is also being removed at the
 391  * same time.
 392  */
 393 typedef enum hammer_record_type {
 394         HAMMER_MEM_RECORD_GENERAL,      /* misc record */
 395         HAMMER_MEM_RECORD_INODE,        /* inode record */
 396         HAMMER_MEM_RECORD_ADD,          /* positive memory cache record */
 397         HAMMER_MEM_RECORD_DEL,          /* negative delete-on-disk record */
 398         HAMMER_MEM_RECORD_DATA          /* bulk-data record w/on-disk ref */
 399 } hammer_record_type_t;
 400
 401 struct hammer_record {
 402         RB_ENTRY(hammer_record)         rb_node;
 403         TAILQ_ENTRY(hammer_record)      target_entry;
 404         hammer_inode_state_t            flush_state;
 405         hammer_flush_group_t            flush_group;
 406         hammer_record_type_t            type;
 407         struct hammer_lock              lock;
 408         struct hammer_reserve           *resv;
 409         struct hammer_inode             *ip;
 410         struct hammer_inode             *target_ip;
 411         struct hammer_btree_leaf_elm    leaf;
 412         union hammer_data_ondisk        *data;
 413         int                             flags;
 414         hammer_off_t                    zone2_offset;   /* direct-write only */
 415 };
 416
 417 typedef struct hammer_record *hammer_record_t;
 418
 419 /*
 420  * Record flags.  Note that FE can only be set by the frontend if the
 421  * record has not been interlocked by the backend w/ BE.
 422  */
 423 #define HAMMER_RECF_ALLOCDATA           0x0001
 424 #define HAMMER_RECF_ONRBTREE            0x0002
 425 #define HAMMER_RECF_DELETED_FE          0x0004  /* deleted (frontend) */
 426 #define HAMMER_RECF_DELETED_BE          0x0008  /* deleted (backend) */
 427 #define HAMMER_RECF_COMMITTED           0x0010  /* committed to the B-Tree */
 428 #define HAMMER_RECF_INTERLOCK_BE        0x0020  /* backend interlock */
 429 #define HAMMER_RECF_WANTED              0x0040  /* wanted by the frontend */
 430 #define HAMMER_RECF_CONVERT_DELETE      0x0100  /* special case */
 431 #define HAMMER_RECF_DIRECT_IO           0x0200  /* related direct I/O running*/
 432 #define HAMMER_RECF_DIRECT_WAIT         0x0400  /* related direct I/O running*/
 433 #define HAMMER_RECF_DIRECT_INVAL        0x0800  /* buffer alias invalidation */
 434
 435 /*
 436  * hammer_delete_at_cursor() flags
 437  */
 438 #define HAMMER_DELETE_ADJUST            0x0001
 439 #define HAMMER_DELETE_DESTROY           0x0002
 440
 441 /*
 442  * In-memory structures representing on-disk structures.
 443  */
 444 struct hammer_volume;
 445 struct hammer_buffer;
 446 struct hammer_node;
 447 struct hammer_undo;
 448 struct hammer_reserve;
 449
 450 RB_HEAD(hammer_vol_rb_tree, hammer_volume);
 451 RB_HEAD(hammer_buf_rb_tree, hammer_buffer);
 452 RB_HEAD(hammer_nod_rb_tree, hammer_node);
 453 RB_HEAD(hammer_und_rb_tree, hammer_undo);
 454 RB_HEAD(hammer_res_rb_tree, hammer_reserve);
 455
 456 RB_PROTOTYPE2(hammer_vol_rb_tree, hammer_volume, rb_node,
 457               hammer_vol_rb_compare, int32_t);
 458 RB_PROTOTYPE2(hammer_buf_rb_tree, hammer_buffer, rb_node,
 459               hammer_buf_rb_compare, hammer_off_t);
 460 RB_PROTOTYPE2(hammer_nod_rb_tree, hammer_node, rb_node,
 461               hammer_nod_rb_compare, hammer_off_t);
 462 RB_PROTOTYPE2(hammer_und_rb_tree, hammer_undo, rb_node,
 463               hammer_und_rb_compare, hammer_off_t);
 464 RB_PROTOTYPE2(hammer_res_rb_tree, hammer_reserve, rb_node,
 465               hammer_res_rb_compare, hammer_off_t);
 466
 467 /*
 468  * IO management - embedded at the head of various in-memory structures
 469  *
 470  * VOLUME       - hammer_volume containing meta-data
 471  * META_BUFFER  - hammer_buffer containing meta-data
 472  * DATA_BUFFER  - hammer_buffer containing pure-data
 473  *
 474  * Dirty volume headers and dirty meta-data buffers are locked until the
 475  * flusher can sequence them out.  Dirty pure-data buffers can be written.
 476  * Clean buffers can be passively released.
 477  */
 478 typedef enum hammer_io_type {
 479         HAMMER_STRUCTURE_VOLUME,
 480         HAMMER_STRUCTURE_META_BUFFER,
 481         HAMMER_STRUCTURE_UNDO_BUFFER,
 482         HAMMER_STRUCTURE_DATA_BUFFER
 483 } hammer_io_type_t;
 484
 485 union hammer_io_structure;
 486 struct hammer_io;
 487
 488 struct worklist {
 489         LIST_ENTRY(worklist) node;
 490 };
 491
 492 TAILQ_HEAD(hammer_io_list, hammer_io);
 493 typedef struct hammer_io_list *hammer_io_list_t;
 494
 495 struct hammer_io {
 496         struct worklist         worklist;
 497         struct hammer_lock      lock;
 498         enum hammer_io_type     type;
 499         struct hammer_mount     *hmp;
 500         struct hammer_volume    *volume;
 501         TAILQ_ENTRY(hammer_io)  mod_entry; /* list entry if modified */
 502         hammer_io_list_t        mod_list;
 503         struct buf              *bp;
 504         int64_t                 offset;    /* zone-2 offset */
 505         int                     bytes;     /* buffer cache buffer size */
 506         int                     loading;   /* loading/unloading interlock */
 507         int                     modify_refs;
 508
 509         u_int           modified : 1;   /* bp's data was modified */
 510         u_int           released : 1;   /* bp released (w/ B_LOCKED set) */
 511         u_int           running : 1;    /* bp write IO in progress */
 512         u_int           waiting : 1;    /* someone is waiting on us */
 513         u_int           validated : 1;  /* ondisk has been validated */
 514         u_int           waitdep : 1;    /* flush waits for dependancies */
 515         u_int           recovered : 1;  /* has recovery ref */
 516         u_int           waitmod : 1;    /* waiting for modify_refs */
 517         u_int           reclaim : 1;    /* reclaim requested */
 518         u_int           gencrc : 1;     /* crc needs to be generated */
 519         u_int           ioerror : 1;    /* abort on io-error */
 520 };
 521
 522 typedef struct hammer_io *hammer_io_t;
 523
 524 #define HAMMER_CLUSTER_SIZE     (64 * 1024)
 525 #if HAMMER_CLUSTER_SIZE > MAXBSIZE
 526 #undef  HAMMER_CLUSTER_SIZE
 527 #define HAMMER_CLUSTER_SIZE     MAXBSIZE
 528 #endif
 529 #define HAMMER_CLUSTER_BUFS     (HAMMER_CLUSTER_SIZE / HAMMER_BUFSIZE)
 530
 531 /*
 532  * In-memory volume representing on-disk buffer
 533  */
 534 struct hammer_volume {
 535         struct hammer_io io;
 536         RB_ENTRY(hammer_volume) rb_node;
 537         struct hammer_volume_ondisk *ondisk;
 538         int32_t vol_no;
 539         int64_t nblocks;        /* note: special calculation for statfs */
 540         int64_t buffer_base;    /* base offset of buffer 0 */
 541         hammer_off_t maxbuf_off; /* Maximum buffer offset (zone-2) */
 542         hammer_off_t maxraw_off; /* Maximum raw offset for device */
 543         char    *vol_name;
 544         struct vnode *devvp;
 545         int     vol_flags;
 546 };
 547
 548 typedef struct hammer_volume *hammer_volume_t;
 549
 550 /*
 551  * In-memory buffer (other then volume, super-cluster, or cluster),
 552  * representing an on-disk buffer.
 553  */
 554 struct hammer_buffer {
 555         struct hammer_io io;
 556         RB_ENTRY(hammer_buffer) rb_node;
 557         void *ondisk;
 558         hammer_off_t zoneX_offset;
 559         hammer_off_t zone2_offset;
 560         struct hammer_reserve *resv;
 561         struct hammer_node_list clist;
 562 };
 563
 564 typedef struct hammer_buffer *hammer_buffer_t;
 565
 566 /*
 567  * In-memory B-Tree node, representing an on-disk B-Tree node.
 568  *
 569  * This is a hang-on structure which is backed by a hammer_buffer,
 570  * indexed by a hammer_cluster, and used for fine-grained locking of
 571  * B-Tree nodes in order to properly control lock ordering.  A hammer_buffer
 572  * can contain multiple nodes representing wildly disassociated portions
 573  * of the B-Tree so locking cannot be done on a buffer-by-buffer basis.
 574  *
 575  * This structure uses a cluster-relative index to reduce the number
 576  * of layers required to access it, and also because all on-disk B-Tree
 577  * references are cluster-relative offsets.
 578  */
 579 struct hammer_node {
 580         struct hammer_lock      lock;           /* node-by-node lock */
 581         TAILQ_ENTRY(hammer_node) entry;         /* per-buffer linkage */
 582         RB_ENTRY(hammer_node)   rb_node;        /* per-cluster linkage */
 583         hammer_off_t            node_offset;    /* full offset spec */
 584         struct hammer_mount     *hmp;
 585         struct hammer_buffer    *buffer;        /* backing buffer */
 586         hammer_node_ondisk_t    ondisk;         /* ptr to on-disk structure */
 587         TAILQ_HEAD(, hammer_cursor) cursor_list;  /* deadlock recovery */
 588         struct hammer_node_cache_list cache_list; /* passive caches */
 589         int                     flags;
 590         int                     loading;        /* load interlock */
 591 };
 592
 593 #define HAMMER_NODE_DELETED     0x0001
 594 #define HAMMER_NODE_FLUSH       0x0002
 595 #define HAMMER_NODE_CRCGOOD     0x0004
 596 #define HAMMER_NODE_NEEDSCRC    0x0008
 597 #define HAMMER_NODE_NEEDSMIRROR 0x0010
 598 #define HAMMER_NODE_CRCBAD      0x0020
 599
 600 #define HAMMER_NODE_CRCANY      (HAMMER_NODE_CRCGOOD | HAMMER_NODE_CRCBAD)
 601
 602 typedef struct hammer_node      *hammer_node_t;
 603
 604 /*
 605  * List of locked nodes.  This structure is used to lock potentially large
 606  * numbers of nodes as an aid for complex B-Tree operations.
 607  */
 608 struct hammer_node_lock;
 609 TAILQ_HEAD(hammer_node_lock_list, hammer_node_lock);
 610
 611 struct hammer_node_lock {
 612         TAILQ_ENTRY(hammer_node_lock) entry;
 613         struct hammer_node_lock_list  list;
 614         struct hammer_node_lock       *parent;
 615         hammer_node_t   node;
 616         hammer_node_ondisk_t copy;      /* copy of on-disk data */
 617         int             index;          /* index of this node in parent */
 618         int             count;          /* count children */
 619         int             flags;
 620 };
 621
 622 typedef struct hammer_node_lock *hammer_node_lock_t;
 623
 624 #define HAMMER_NODE_LOCK_UPDATED        0x0001
 625
 626 /*
 627  * Common I/O management structure - embedded in in-memory structures
 628  * which are backed by filesystem buffers.
 629  */
 630 union hammer_io_structure {
 631         struct hammer_io        io;
 632         struct hammer_volume    volume;
 633         struct hammer_buffer    buffer;
 634 };
 635
 636 typedef union hammer_io_structure *hammer_io_structure_t;
 637
 638 /*
 639  * The reserve structure prevents the blockmap from allocating
 640  * out of a reserved bigblock.  Such reservations are used by
 641  * the direct-write mechanism.
 642  *
 643  * The structure is also used to hold off on reallocations of
 644  * big blocks from the freemap until flush dependancies have
 645  * been dealt with.
 646  */
 647 struct hammer_reserve {
 648         RB_ENTRY(hammer_reserve) rb_node;
 649         TAILQ_ENTRY(hammer_reserve) delay_entry;
 650         int             flush_group;
 651         int             flags;
 652         int             refs;
 653         int             zone;
 654         int             append_off;
 655         hammer_off_t    zone_offset;
 656 };
 657
 658 typedef struct hammer_reserve *hammer_reserve_t;
 659
 660 #define HAMMER_RESF_ONDELAY     0x0001
 661 #define HAMMER_RESF_LAYER2FREE  0x0002
 662
 663 #include "hammer_cursor.h"
 664
 665 /*
 666  * The undo structure tracks recent undos to avoid laying down duplicate
 667  * undos within a flush group, saving us a significant amount of overhead.
 668  *
 669  * This is strictly a heuristic.
 670  */
 671 #define HAMMER_MAX_UNDOS                1024
 672 #define HAMMER_MAX_FLUSHERS             4
 673
 674 struct hammer_undo {
 675         RB_ENTRY(hammer_undo)   rb_node;
 676         TAILQ_ENTRY(hammer_undo) lru_entry;
 677         hammer_off_t            offset;
 678         int                     bytes;
 679 };
 680
 681 typedef struct hammer_undo *hammer_undo_t;
 682
 683 struct hammer_flusher_info;
 684 TAILQ_HEAD(hammer_flusher_info_list, hammer_flusher_info);
 685
 686 struct hammer_flusher {
 687         int             signal;         /* flusher thread sequencer */
 688         int             act;            /* currently active flush group */
 689         int             done;           /* set to act when complete */
 690         int             next;           /* next flush group */
 691         int             group_lock;     /* lock sequencing of the next flush */
 692         int             exiting;        /* request master exit */
 693         thread_t        td;             /* master flusher thread */
 694         hammer_tid_t    tid;            /* last flushed transaction id */
 695         int             finalize_want;          /* serialize finalization */
 696         struct hammer_lock finalize_lock;       /* serialize finalization */
 697         struct hammer_transaction trans;        /* shared transaction */
 698         struct hammer_flusher_info_list run_list;
 699         struct hammer_flusher_info_list ready_list;
 700 };
 701
 702 /*
 703  * Internal hammer mount data structure
 704  */
 705 struct hammer_mount {
 706         struct mount *mp;
 707         /*struct vnode *rootvp;*/
 708         struct hammer_ino_rb_tree rb_inos_root;
 709         struct hammer_vol_rb_tree rb_vols_root;
 710         struct hammer_nod_rb_tree rb_nods_root;
 711         struct hammer_und_rb_tree rb_undo_root;
 712         struct hammer_res_rb_tree rb_resv_root;
 713         struct hammer_buf_rb_tree rb_bufs_root;
 714         struct hammer_pfs_rb_tree rb_pfsm_root;
 715         struct hammer_volume *rootvol;
 716         struct hammer_base_elm root_btree_beg;
 717         struct hammer_base_elm root_btree_end;
 718
 719         struct malloc_type      *m_misc;
 720         struct malloc_type      *m_inodes;
 721
 722         int     flags;          /* HAMMER_MOUNT_xxx flags */
 723         int     hflags;
 724         int     ronly;
 725         int     nvolumes;
 726         int     volume_iterator;
 727         int     master_id;      /* -1 or 0-15 - clustering and mirroring */
 728         int     version;        /* hammer filesystem version to use */
 729         int     rsv_inodes;     /* reserved space due to dirty inodes */
 730         int64_t rsv_databytes;  /* reserved space due to record data */
 731         int     rsv_recs;       /* reserved space due to dirty records */
 732         int     rsv_fromdelay;  /* bigblocks reserved due to flush delay */
 733         int     undo_rec_limit; /* based on size of undo area */
 734         int     last_newrecords;
 735         int     count_newrecords;
 736
 737         int     inode_reclaims; /* inodes pending reclaim by flusher */
 738         int     count_inodes;   /* total number of inodes */
 739         int     count_iqueued;  /* inodes queued to flusher */
 740
 741         struct hammer_flusher flusher;
 742
 743         u_int   check_interrupt;
 744         uuid_t  fsid;
 745         struct hammer_io_list volu_list;        /* dirty undo buffers */
 746         struct hammer_io_list undo_list;        /* dirty undo buffers */
 747         struct hammer_io_list data_list;        /* dirty data buffers */
 748         struct hammer_io_list alt_data_list;    /* dirty data buffers */
 749         struct hammer_io_list meta_list;        /* dirty meta bufs    */
 750         struct hammer_io_list lose_list;        /* loose buffers      */
 751         int     locked_dirty_space;             /* meta/volu count    */
 752         int     io_running_space;
 753         int     objid_cache_count;
 754         int     error;                          /* critical I/O error */
 755         struct krate    krate;                  /* rate limited kprintf */
 756         hammer_tid_t    asof;                   /* snapshot mount */
 757         hammer_tid_t    next_tid;
 758         hammer_tid_t    flush_tid1;             /* flusher tid sequencing */
 759         hammer_tid_t    flush_tid2;             /* flusher tid sequencing */
 760         int64_t copy_stat_freebigblocks;        /* number of free bigblocks */
 761
 762         struct netexport export;
 763         struct hammer_lock sync_lock;
 764         struct hammer_lock free_lock;
 765         struct hammer_lock undo_lock;
 766         struct hammer_lock blkmap_lock;
 767         struct hammer_blockmap  blockmap[HAMMER_MAX_ZONES];
 768         struct hammer_undo      undos[HAMMER_MAX_UNDOS];
 769         int                     undo_alloc;
 770         TAILQ_HEAD(, hammer_undo)  undo_lru_list;
 771         TAILQ_HEAD(, hammer_reserve) delay_list;
 772         struct hammer_flush_group_list  flush_group_list;
 773         hammer_flush_group_t    next_flush_group;
 774         TAILQ_HEAD(, hammer_objid_cache) objid_cache_list;
 775         TAILQ_HEAD(, hammer_reclaim) reclaim_list;
 776 };
 777
 778 typedef struct hammer_mount     *hammer_mount_t;
 779
 780 #define HAMMER_MOUNT_CRITICAL_ERROR     0x0001
 781 #define HAMMER_MOUNT_FLUSH_RECOVERY     0x0002
 782
 783 struct hammer_sync_info {
 784         int error;
 785         int waitfor;
 786 };
 787
 788 #endif
 789
 790 /*
 791  * checkspace slop (8MB chunks), higher numbers are more conservative.
 792  */
 793 #define HAMMER_CHKSPC_REBLOCK   25
 794 #define HAMMER_CHKSPC_MIRROR    20
 795 #define HAMMER_CHKSPC_WRITE     20
 796 #define HAMMER_CHKSPC_CREATE    20
 797 #define HAMMER_CHKSPC_REMOVE    10
 798 #define HAMMER_CHKSPC_EMERGENCY 0
 799
 800 #if defined(_KERNEL)
 801
 802 extern struct vop_ops hammer_vnode_vops;
 803 extern struct vop_ops hammer_spec_vops;
 804 extern struct vop_ops hammer_fifo_vops;
 805 extern struct bio_ops hammer_bioops;
 806
 807 extern int hammer_debug_io;
 808 extern int hammer_debug_general;
 809 extern int hammer_debug_debug;
 810 extern int hammer_debug_inode;
 811 extern int hammer_debug_locks;
 812 extern int hammer_debug_btree;
 813 extern int hammer_debug_tid;
 814 extern int hammer_debug_recover;
 815 extern int hammer_debug_recover_faults;
 816 extern int hammer_cluster_enable;
 817 extern int hammer_count_fsyncs;
 818 extern int hammer_count_inodes;
 819 extern int hammer_count_iqueued;
 820 extern int hammer_count_reclaiming;
 821 extern int hammer_count_records;
 822 extern int hammer_count_record_datas;
 823 extern int hammer_count_volumes;
 824 extern int hammer_count_buffers;
 825 extern int hammer_count_nodes;
 826 extern int64_t hammer_count_extra_space_used;
 827 extern int64_t hammer_stats_btree_lookups;
 828 extern int64_t hammer_stats_btree_searches;
 829 extern int64_t hammer_stats_btree_inserts;
 830 extern int64_t hammer_stats_btree_deletes;
 831 extern int64_t hammer_stats_btree_elements;
 832 extern int64_t hammer_stats_btree_splits;
 833 extern int64_t hammer_stats_btree_iterations;
 834 extern int64_t hammer_stats_btree_root_iterations;
 835 extern int64_t hammer_stats_record_iterations;
 836 extern int64_t hammer_stats_file_read;
 837 extern int64_t hammer_stats_file_write;
 838 extern int64_t hammer_stats_file_iopsr;
 839 extern int64_t hammer_stats_file_iopsw;
 840 extern int64_t hammer_stats_disk_read;
 841 extern int64_t hammer_stats_disk_write;
 842 extern int64_t hammer_stats_inode_flushes;
 843 extern int64_t hammer_stats_commits;
 844 extern int hammer_count_dirtybufspace;
 845 extern int hammer_count_refedbufs;
 846 extern int hammer_count_reservations;
 847 extern int hammer_count_io_running_read;
 848 extern int hammer_count_io_running_write;
 849 extern int hammer_count_io_locked;
 850 extern int hammer_limit_dirtybufspace;
 851 extern int hammer_limit_recs;
 852 extern int hammer_bio_count;
 853 extern int hammer_verify_zone;
 854 extern int hammer_verify_data;
 855 extern int hammer_write_mode;
 856 extern int hammer_autoflush;
 857 extern int64_t hammer_contention_count;
 858
 859 void    hammer_critical_error(hammer_mount_t hmp, hammer_inode_t ip,
 860                         int error, const char *msg);
 861 int     hammer_vop_inactive(struct vop_inactive_args *);
 862 int     hammer_vop_reclaim(struct vop_reclaim_args *);
 863 int     hammer_get_vnode(struct hammer_inode *ip, struct vnode **vpp);
 864 struct hammer_inode *hammer_get_inode(hammer_transaction_t trans,
 865                         hammer_inode_t dip, int64_t obj_id,
 866                         hammer_tid_t asof, u_int32_t localization,
 867                         int flags, int *errorp);
 868 struct hammer_inode *hammer_get_dummy_inode(hammer_transaction_t trans,
 869                         hammer_inode_t dip, int64_t obj_id,
 870                         hammer_tid_t asof, u_int32_t localization,
 871                         int flags, int *errorp);
 872 struct hammer_inode *hammer_find_inode(hammer_transaction_t trans,
 873                         int64_t obj_id, hammer_tid_t asof,
 874                         u_int32_t localization);
 875 void    hammer_scan_inode_snapshots(hammer_mount_t hmp,
 876                         hammer_inode_info_t iinfo,
 877                         int (*callback)(hammer_inode_t ip, void *data),
 878                         void *data);
 879 void    hammer_put_inode(struct hammer_inode *ip);
 880 void    hammer_put_inode_ref(struct hammer_inode *ip);
 881 void    hammer_inode_waitreclaims(hammer_mount_t hmp);
 882 void    hammer_inode_waithard(hammer_mount_t hmp);
 883
 884 int     hammer_unload_volume(hammer_volume_t volume, void *data __unused);
 885 int     hammer_adjust_volume_mode(hammer_volume_t volume, void *data __unused);
 886
 887 int     hammer_unload_buffer(hammer_buffer_t buffer, void *data __unused);
 888 int     hammer_install_volume(hammer_mount_t hmp, const char *volname,
 889                         struct vnode *devvp);
 890 int     hammer_mountcheck_volumes(hammer_mount_t hmp);
 891
 892 int     hammer_mem_add(hammer_record_t record);
 893 int     hammer_ip_lookup(hammer_cursor_t cursor);
 894 int     hammer_ip_first(hammer_cursor_t cursor);
 895 int     hammer_ip_next(hammer_cursor_t cursor);
 896 int     hammer_ip_resolve_data(hammer_cursor_t cursor);
 897 int     hammer_ip_delete_record(hammer_cursor_t cursor, hammer_inode_t ip,
 898                         hammer_tid_t tid);
 899 int     hammer_delete_at_cursor(hammer_cursor_t cursor, int delete_flags,
 900                         hammer_tid_t delete_tid, u_int32_t delete_ts,
 901                         int track, int64_t *stat_bytes);
 902 int     hammer_ip_check_directory_empty(hammer_transaction_t trans,
 903                         hammer_inode_t ip);
 904 int     hammer_sync_hmp(hammer_mount_t hmp, int waitfor);
 905 int     hammer_queue_inodes_flusher(hammer_mount_t hmp, int waitfor);
 906
 907 hammer_record_t
 908         hammer_alloc_mem_record(hammer_inode_t ip, int data_len);
 909 void    hammer_flush_record_done(hammer_record_t record, int error);
 910 void    hammer_wait_mem_record_ident(hammer_record_t record, const char *ident);
 911 void    hammer_rel_mem_record(hammer_record_t record);
 912
 913 int     hammer_cursor_up(hammer_cursor_t cursor);
 914 int     hammer_cursor_up_locked(hammer_cursor_t cursor);
 915 int     hammer_cursor_down(hammer_cursor_t cursor);
 916 int     hammer_cursor_upgrade(hammer_cursor_t cursor);
 917 int     hammer_cursor_upgrade_node(hammer_cursor_t cursor);
 918 void    hammer_cursor_downgrade(hammer_cursor_t cursor);
 919 int     hammer_cursor_seek(hammer_cursor_t cursor, hammer_node_t node,
 920                         int index);
 921 void    hammer_lock_ex_ident(struct hammer_lock *lock, const char *ident);
 922 int     hammer_lock_ex_try(struct hammer_lock *lock);
 923 void    hammer_lock_sh(struct hammer_lock *lock);
 924 int     hammer_lock_sh_try(struct hammer_lock *lock);
 925 int     hammer_lock_upgrade(struct hammer_lock *lock);
 926 void    hammer_lock_downgrade(struct hammer_lock *lock);
 927 int     hammer_lock_status(struct hammer_lock *lock);
 928 void    hammer_unlock(struct hammer_lock *lock);
 929 void    hammer_ref(struct hammer_lock *lock);
 930 void    hammer_unref(struct hammer_lock *lock);
 931
 932 void    hammer_sync_lock_ex(hammer_transaction_t trans);
 933 void    hammer_sync_lock_sh(hammer_transaction_t trans);
 934 int     hammer_sync_lock_sh_try(hammer_transaction_t trans);
 935 void    hammer_sync_unlock(hammer_transaction_t trans);
 936
 937 u_int32_t hammer_to_unix_xid(uuid_t *uuid);
 938 void hammer_guid_to_uuid(uuid_t *uuid, u_int32_t guid);
 939 void    hammer_time_to_timespec(u_int64_t xtime, struct timespec *ts);
 940 u_int64_t hammer_timespec_to_time(struct timespec *ts);
 941 int     hammer_str_to_tid(const char *str, int *ispfsp,
 942                         hammer_tid_t *tidp, u_int32_t *localizationp);
 943 int     hammer_is_atatext(const char *name, int len);
 944 hammer_tid_t hammer_alloc_objid(hammer_mount_t hmp, hammer_inode_t dip,
 945                         int64_t namekey);
 946 void hammer_clear_objid(hammer_inode_t dip);
 947 void hammer_destroy_objid_cache(hammer_mount_t hmp);
 948
 949 int hammer_enter_undo_history(hammer_mount_t hmp, hammer_off_t offset,
 950                         int bytes);
 951 void hammer_clear_undo_history(hammer_mount_t hmp);
 952 enum vtype hammer_get_vnode_type(u_int8_t obj_type);
 953 int hammer_get_dtype(u_int8_t obj_type);
 954 u_int8_t hammer_get_obj_type(enum vtype vtype);
 955 int64_t hammer_directory_namekey(hammer_inode_t dip, const void *name, int len,
 956                         u_int32_t *max_iterationsp);
 957 int     hammer_nohistory(hammer_inode_t ip);
 958
 959 int     hammer_init_cursor(hammer_transaction_t trans, hammer_cursor_t cursor,
 960                         hammer_node_cache_t cache, hammer_inode_t ip);
 961 void    hammer_normalize_cursor(hammer_cursor_t cursor);
 962 void    hammer_done_cursor(hammer_cursor_t cursor);
 963 int     hammer_recover_cursor(hammer_cursor_t cursor);
 964 void    hammer_unlock_cursor(hammer_cursor_t cursor);
 965 int     hammer_lock_cursor(hammer_cursor_t cursor);
 966 hammer_cursor_t hammer_push_cursor(hammer_cursor_t ocursor);
 967 void    hammer_pop_cursor(hammer_cursor_t ocursor, hammer_cursor_t ncursor);
 968
 969 void    hammer_cursor_replaced_node(hammer_node_t onode, hammer_node_t nnode);
 970 void    hammer_cursor_removed_node(hammer_node_t onode, hammer_node_t parent,
 971                         int index);
 972 void    hammer_cursor_split_node(hammer_node_t onode, hammer_node_t nnode,
 973                         int index);
 974 void    hammer_cursor_moved_element(hammer_node_t onode, hammer_node_t nnode,
 975                         int oindex, int nindex);
 976 void    hammer_cursor_parent_changed(hammer_node_t node, hammer_node_t oparent,
 977                         hammer_node_t nparent, int nindex);
 978 void    hammer_cursor_inserted_element(hammer_node_t node, int index);
 979 void    hammer_cursor_deleted_element(hammer_node_t node, int index);
 980
 981 int     hammer_btree_lookup(hammer_cursor_t cursor);
 982 int     hammer_btree_first(hammer_cursor_t cursor);
 983 int     hammer_btree_last(hammer_cursor_t cursor);
 984 int     hammer_btree_extract(hammer_cursor_t cursor, int flags);
 985 int     hammer_btree_iterate(hammer_cursor_t cursor);
 986 int     hammer_btree_iterate_reverse(hammer_cursor_t cursor);
 987 int     hammer_btree_insert(hammer_cursor_t cursor,
 988                             hammer_btree_leaf_elm_t elm, int *doprop);
 989 int     hammer_btree_delete(hammer_cursor_t cursor);
 990 void    hammer_btree_do_propagation(hammer_cursor_t cursor,
 991                             hammer_pseudofs_inmem_t pfsm,
 992                             hammer_btree_leaf_elm_t leaf);
 993 int     hammer_btree_cmp(hammer_base_elm_t key1, hammer_base_elm_t key2);
 994 int     hammer_btree_chkts(hammer_tid_t ts, hammer_base_elm_t key);
 995 int     hammer_btree_correct_rhb(hammer_cursor_t cursor, hammer_tid_t tid);
 996 int     hammer_btree_correct_lhb(hammer_cursor_t cursor, hammer_tid_t tid);
 997
 998 int     btree_set_parent(hammer_transaction_t trans, hammer_node_t node,
 999                         hammer_btree_elm_t elm);
1000 void    hammer_node_lock_init(hammer_node_lock_t parent, hammer_node_t node);
1001 int     hammer_btree_lock_children(hammer_cursor_t cursor, int depth,
1002                         hammer_node_lock_t parent);
1003 void    hammer_btree_lock_copy(hammer_cursor_t cursor,
1004                         hammer_node_lock_t parent);
1005 int     hammer_btree_sync_copy(hammer_cursor_t cursor,
1006                         hammer_node_lock_t parent);
1007 void    hammer_btree_unlock_children(hammer_cursor_t cursor,
1008                         hammer_node_lock_t parent);
1009 int     hammer_btree_search_node(hammer_base_elm_t elm, hammer_node_ondisk_t node);
1010 hammer_node_t hammer_btree_get_parent(hammer_transaction_t trans,
1011                         hammer_node_t node, int *parent_indexp,
1012                         int *errorp, int try_exclusive);
1013
1014 void    hammer_print_btree_node(hammer_node_ondisk_t ondisk);
1015 void    hammer_print_btree_elm(hammer_btree_elm_t elm, u_int8_t type, int i);
1016
1017 void    *hammer_bread(struct hammer_mount *hmp, hammer_off_t off,
1018                         int *errorp, struct hammer_buffer **bufferp);
1019 void    *hammer_bnew(struct hammer_mount *hmp, hammer_off_t off,
1020                         int *errorp, struct hammer_buffer **bufferp);
1021 void    *hammer_bread_ext(struct hammer_mount *hmp, hammer_off_t off, int bytes,
1022                         int *errorp, struct hammer_buffer **bufferp);
1023 void    *hammer_bnew_ext(struct hammer_mount *hmp, hammer_off_t off, int bytes,
1024                         int *errorp, struct hammer_buffer **bufferp);
1025
1026 hammer_volume_t hammer_get_root_volume(hammer_mount_t hmp, int *errorp);
1027
1028 hammer_volume_t hammer_get_volume(hammer_mount_t hmp,
1029                         int32_t vol_no, int *errorp);
1030 hammer_buffer_t hammer_get_buffer(hammer_mount_t hmp, hammer_off_t buf_offset,
1031                         int bytes, int isnew, int *errorp);
1032 void            hammer_sync_buffers(hammer_mount_t hmp,
1033                         hammer_off_t base_offset, int bytes);
1034 int             hammer_del_buffers(hammer_mount_t hmp,
1035                         hammer_off_t base_offset,
1036                         hammer_off_t zone2_offset, int bytes,
1037                         int report_conflicts);
1038
1039 int             hammer_ref_volume(hammer_volume_t volume);
1040 int             hammer_ref_buffer(hammer_buffer_t buffer);
1041 void            hammer_flush_buffer_nodes(hammer_buffer_t buffer);
1042
1043 void            hammer_rel_volume(hammer_volume_t volume, int flush);
1044 void            hammer_rel_buffer(hammer_buffer_t buffer, int flush);
1045
1046 int             hammer_vfs_export(struct mount *mp, int op,
1047                         const struct export_args *export);
1048 hammer_node_t   hammer_get_node(hammer_transaction_t trans,
1049                         hammer_off_t node_offset, int isnew, int *errorp);
1050 void            hammer_ref_node(hammer_node_t node);
1051 hammer_node_t   hammer_ref_node_safe(hammer_transaction_t trans,
1052                         hammer_node_cache_t cache, int *errorp);
1053 void            hammer_rel_node(hammer_node_t node);
1054 void            hammer_delete_node(hammer_transaction_t trans,
1055                         hammer_node_t node);
1056 void            hammer_cache_node(hammer_node_cache_t cache,
1057                         hammer_node_t node);
1058 void            hammer_uncache_node(hammer_node_cache_t cache);
1059 void            hammer_flush_node(hammer_node_t node);
1060
1061 void hammer_dup_buffer(struct hammer_buffer **bufferp,
1062                         struct hammer_buffer *buffer);
1063 hammer_node_t hammer_alloc_btree(hammer_transaction_t trans,
1064                         hammer_off_t hint, int *errorp);
1065 void *hammer_alloc_data(hammer_transaction_t trans, int32_t data_len,
1066                         u_int16_t rec_type, hammer_off_t *data_offsetp,
1067                         struct hammer_buffer **data_bufferp,
1068                         hammer_off_t hint, int *errorp);
1069
1070 int hammer_generate_undo(hammer_transaction_t trans, hammer_io_t io,
1071                         hammer_off_t zone1_offset, void *base, int len);
1072
1073 void hammer_put_volume(struct hammer_volume *volume, int flush);
1074 void hammer_put_buffer(struct hammer_buffer *buffer, int flush);
1075
1076 hammer_off_t hammer_freemap_alloc(hammer_transaction_t trans,
1077                         hammer_off_t owner, int *errorp);
1078 void hammer_freemap_free(hammer_transaction_t trans, hammer_off_t phys_offset,
1079                         hammer_off_t owner, int *errorp);
1080 int _hammer_checkspace(hammer_mount_t hmp, int slop, int64_t *resp);
1081 hammer_off_t hammer_blockmap_alloc(hammer_transaction_t trans, int zone,
1082                         int bytes, hammer_off_t hint, int *errorp);
1083 hammer_reserve_t hammer_blockmap_reserve(hammer_mount_t hmp, int zone,
1084                         int bytes, hammer_off_t *zone_offp, int *errorp);
1085 void hammer_blockmap_reserve_complete(hammer_mount_t hmp,
1086                         hammer_reserve_t resv);
1087 void hammer_reserve_clrdelay(hammer_mount_t hmp, hammer_reserve_t resv);
1088 void hammer_blockmap_free(hammer_transaction_t trans,
1089                         hammer_off_t bmap_off, int bytes);
1090 int hammer_blockmap_finalize(hammer_transaction_t trans,
1091                         hammer_reserve_t resv,
1092                         hammer_off_t bmap_off, int bytes);
1093 int hammer_blockmap_getfree(hammer_mount_t hmp, hammer_off_t bmap_off,
1094                         int *curp, int *errorp);
1095 hammer_off_t hammer_blockmap_lookup(hammer_mount_t hmp, hammer_off_t bmap_off,
1096                         int *errorp);
1097 hammer_off_t hammer_undo_lookup(hammer_mount_t hmp, hammer_off_t bmap_off,
1098                         int *errorp);
1099 int64_t hammer_undo_used(hammer_transaction_t trans);
1100 int64_t hammer_undo_space(hammer_transaction_t trans);
1101 int64_t hammer_undo_max(hammer_mount_t hmp);
1102
1103 void hammer_start_transaction(struct hammer_transaction *trans,
1104                               struct hammer_mount *hmp);
1105 void hammer_simple_transaction(struct hammer_transaction *trans,
1106                               struct hammer_mount *hmp);
1107 void hammer_start_transaction_fls(struct hammer_transaction *trans,
1108                                   struct hammer_mount *hmp);
1109 void hammer_done_transaction(struct hammer_transaction *trans);
1110
1111 void hammer_modify_inode(hammer_inode_t ip, int flags);
1112 void hammer_flush_inode(hammer_inode_t ip, int flags);
1113 void hammer_flush_inode_done(hammer_inode_t ip, int error);
1114 void hammer_wait_inode(hammer_inode_t ip);
1115
1116 int  hammer_create_inode(struct hammer_transaction *trans, struct vattr *vap,
1117                         struct ucred *cred, struct hammer_inode *dip,
1118                         const char *name, int namelen,
1119                         hammer_pseudofs_inmem_t pfsm,
1120                         struct hammer_inode **ipp);
1121 void hammer_rel_inode(hammer_inode_t ip, int flush);
1122 int hammer_reload_inode(hammer_inode_t ip, void *arg __unused);
1123 int hammer_ino_rb_compare(hammer_inode_t ip1, hammer_inode_t ip2);
1124 int hammer_destroy_inode_callback(hammer_inode_t ip, void *data __unused);
1125
1126 int hammer_sync_inode(hammer_transaction_t trans, hammer_inode_t ip);
1127 void hammer_test_inode(hammer_inode_t dip);
1128 void hammer_inode_unloadable_check(hammer_inode_t ip, int getvp);
1129
1130 int  hammer_ip_add_directory(struct hammer_transaction *trans,
1131                         hammer_inode_t dip, const char *name, int bytes,
1132                         hammer_inode_t nip);
1133 int  hammer_ip_del_directory(struct hammer_transaction *trans,
1134                         hammer_cursor_t cursor, hammer_inode_t dip,
1135                         hammer_inode_t ip);
1136 hammer_record_t hammer_ip_add_bulk(hammer_inode_t ip, off_t file_offset,
1137                         void *data, int bytes, int *errorp);
1138 int  hammer_ip_frontend_trunc(struct hammer_inode *ip, off_t file_size);
1139 int  hammer_ip_add_record(struct hammer_transaction *trans,
1140                         hammer_record_t record);
1141 int  hammer_ip_delete_range(hammer_cursor_t cursor, hammer_inode_t ip,
1142                         int64_t ran_beg, int64_t ran_end, int truncating);
1143 int  hammer_ip_delete_clean(hammer_cursor_t cursor, hammer_inode_t ip,
1144                         int *countp);
1145 int  hammer_ip_sync_data(hammer_cursor_t cursor, hammer_inode_t ip,
1146                         int64_t offset, void *data, int bytes);
1147 int  hammer_ip_sync_record(hammer_transaction_t trans, hammer_record_t rec);
1148 int  hammer_ip_sync_record_cursor(hammer_cursor_t cursor, hammer_record_t rec);
1149 hammer_pseudofs_inmem_t  hammer_load_pseudofs(hammer_transaction_t trans,
1150                         u_int32_t localization, int *errorp);
1151 int  hammer_mkroot_pseudofs(hammer_transaction_t trans, struct ucred *cred,
1152                         hammer_pseudofs_inmem_t pfsm);
1153 int  hammer_save_pseudofs(hammer_transaction_t trans,
1154                         hammer_pseudofs_inmem_t pfsm);
1155 int  hammer_unload_pseudofs(hammer_transaction_t trans, u_int32_t localization);
1156 void hammer_rel_pseudofs(hammer_mount_t hmp, hammer_pseudofs_inmem_t pfsm);
1157 int hammer_ioctl(hammer_inode_t ip, u_long com, caddr_t data, int fflag,
1158                         struct ucred *cred);
1159
1160 void hammer_io_init(hammer_io_t io, hammer_volume_t volume,
1161                         enum hammer_io_type type);
1162 int hammer_io_read(struct vnode *devvp, struct hammer_io *io,
1163                         hammer_off_t limit);
1164 int hammer_io_new(struct vnode *devvp, struct hammer_io *io);
1165 int hammer_io_inval(hammer_volume_t volume, hammer_off_t zone2_offset);
1166 struct buf *hammer_io_release(struct hammer_io *io, int flush);
1167 void hammer_io_flush(struct hammer_io *io);
1168 void hammer_io_wait(struct hammer_io *io);
1169 void hammer_io_waitdep(struct hammer_io *io);
1170 void hammer_io_wait_all(hammer_mount_t hmp, const char *ident);
1171 int hammer_io_direct_read(hammer_mount_t hmp, struct bio *bio,
1172                         hammer_btree_leaf_elm_t leaf);
1173 int hammer_io_direct_write(hammer_mount_t hmp, hammer_record_t record,
1174                         struct bio *bio);
1175 void hammer_io_direct_wait(hammer_record_t record);
1176 void hammer_io_direct_uncache(hammer_mount_t hmp, hammer_btree_leaf_elm_t leaf);
1177 void hammer_io_write_interlock(hammer_io_t io);
1178 void hammer_io_done_interlock(hammer_io_t io);
1179 void hammer_io_clear_modify(struct hammer_io *io, int inval);
1180 void hammer_io_clear_modlist(struct hammer_io *io);
1181 void hammer_io_flush_sync(hammer_mount_t hmp);
1182
1183 void hammer_modify_volume(hammer_transaction_t trans, hammer_volume_t volume,
1184                         void *base, int len);
1185 void hammer_modify_buffer(hammer_transaction_t trans, hammer_buffer_t buffer,
1186                         void *base, int len);
1187 void hammer_modify_volume_done(hammer_volume_t volume);
1188 void hammer_modify_buffer_done(hammer_buffer_t buffer);
1189
1190 int hammer_ioc_reblock(hammer_transaction_t trans, hammer_inode_t ip,
1191                         struct hammer_ioc_reblock *reblock);
1192 int hammer_ioc_rebalance(hammer_transaction_t trans, hammer_inode_t ip,
1193                         struct hammer_ioc_rebalance *rebal);
1194 int hammer_ioc_prune(hammer_transaction_t trans, hammer_inode_t ip,
1195                         struct hammer_ioc_prune *prune);
1196 int hammer_ioc_mirror_read(hammer_transaction_t trans, hammer_inode_t ip,
1197                         struct hammer_ioc_mirror_rw *mirror);
1198 int hammer_ioc_mirror_write(hammer_transaction_t trans, hammer_inode_t ip,
1199                         struct hammer_ioc_mirror_rw *mirror);
1200 int hammer_ioc_set_pseudofs(hammer_transaction_t trans, hammer_inode_t ip,
1201                         struct ucred *cred, struct hammer_ioc_pseudofs_rw *pfs);
1202 int hammer_ioc_get_pseudofs(hammer_transaction_t trans, hammer_inode_t ip,
1203                         struct hammer_ioc_pseudofs_rw *pfs);
1204 int hammer_ioc_destroy_pseudofs(hammer_transaction_t trans, hammer_inode_t ip,
1205                         struct hammer_ioc_pseudofs_rw *pfs);
1206 int hammer_ioc_downgrade_pseudofs(hammer_transaction_t trans, hammer_inode_t ip,
1207                         struct hammer_ioc_pseudofs_rw *pfs);
1208 int hammer_ioc_upgrade_pseudofs(hammer_transaction_t trans, hammer_inode_t ip,
1209                         struct hammer_ioc_pseudofs_rw *pfs);
1210 int hammer_ioc_wait_pseudofs(hammer_transaction_t trans, hammer_inode_t ip,
1211                         struct hammer_ioc_pseudofs_rw *pfs);
1212
1213 int hammer_signal_check(hammer_mount_t hmp);
1214
1215 void hammer_flusher_create(hammer_mount_t hmp);
1216 void hammer_flusher_destroy(hammer_mount_t hmp);
1217 void hammer_flusher_sync(hammer_mount_t hmp);
1218 int  hammer_flusher_async(hammer_mount_t hmp, hammer_flush_group_t flg);
1219 int  hammer_flusher_async_one(hammer_mount_t hmp);
1220 void hammer_flusher_wait(hammer_mount_t hmp, int seq);
1221 void hammer_flusher_wait_next(hammer_mount_t hmp);
1222 int  hammer_flusher_meta_limit(hammer_mount_t hmp);
1223 int  hammer_flusher_meta_halflimit(hammer_mount_t hmp);
1224 int  hammer_flusher_undo_exhausted(hammer_transaction_t trans, int quarter);
1225 void hammer_flusher_clean_loose_ios(hammer_mount_t hmp);
1226 void hammer_flusher_finalize(hammer_transaction_t trans, int final);
1227 int  hammer_flusher_haswork(hammer_mount_t hmp);
1228
1229
1230 int hammer_recover(hammer_mount_t hmp, hammer_volume_t rootvol);
1231 void hammer_recover_flush_buffers(hammer_mount_t hmp,
1232                         hammer_volume_t root_volume, int final);
1233
1234 void hammer_crc_set_blockmap(hammer_blockmap_t blockmap);
1235 void hammer_crc_set_volume(hammer_volume_ondisk_t ondisk);
1236 void hammer_crc_set_leaf(void *data, hammer_btree_leaf_elm_t leaf);
1237
1238 int hammer_crc_test_blockmap(hammer_blockmap_t blockmap);
1239 int hammer_crc_test_volume(hammer_volume_ondisk_t ondisk);
1240 int hammer_crc_test_btree(hammer_node_ondisk_t ondisk);
1241 int hammer_crc_test_leaf(void *data, hammer_btree_leaf_elm_t leaf);
1242 void hkprintf(const char *ctl, ...);
1243 udev_t hammer_fsid_to_udev(uuid_t *uuid);
1244
1245
1246 int hammer_blocksize(int64_t file_offset);
1247 int64_t hammer_blockdemarc(int64_t file_offset1, int64_t file_offset2);
1248
1249 /*
1250  * Shortcut for _hammer_checkspace(), used all over the code.
1251  */
1252 static __inline int
1253 hammer_checkspace(hammer_mount_t hmp, int slop)
1254 {
1255         return(_hammer_checkspace(hmp, slop, NULL));
1256 }
1257
1258 #endif
1259
1260 static __inline void
1261 hammer_wait_mem_record(hammer_record_t record)
1262 {
1263         hammer_wait_mem_record_ident(record, "hmmwai");
1264 }
1265
1266 static __inline void
1267 hammer_lock_ex(struct hammer_lock *lock)
1268 {
1269         hammer_lock_ex_ident(lock, "hmrlck");
1270 }
1271
1272 /*
1273  * Indicate that a B-Tree node is being modified.
1274  */
1275 static __inline void
1276 hammer_modify_node_noundo(hammer_transaction_t trans, hammer_node_t node)
1277 {
1278         KKASSERT((node->flags & HAMMER_NODE_CRCBAD) == 0);
1279         hammer_modify_buffer(trans, node->buffer, NULL, 0);
1280 }
1281
1282 static __inline void
1283 hammer_modify_node_all(hammer_transaction_t trans, struct hammer_node *node)
1284 {
1285         KKASSERT((node->flags & HAMMER_NODE_CRCBAD) == 0);
1286         hammer_modify_buffer(trans, node->buffer,
1287                              node->ondisk, sizeof(*node->ondisk));
1288 }
1289
1290 static __inline void
1291 hammer_modify_node(hammer_transaction_t trans, hammer_node_t node,
1292                    void *base, int len)
1293 {
1294         hammer_crc_t *crcptr;
1295
1296         KKASSERT((char *)base >= (char *)node->ondisk &&
1297                  (char *)base + len <=
1298                     (char *)node->ondisk + sizeof(*node->ondisk));
1299         KKASSERT((node->flags & HAMMER_NODE_CRCBAD) == 0);
1300         hammer_modify_buffer(trans, node->buffer, base, len);
1301         crcptr = &node->ondisk->crc;
1302         hammer_modify_buffer(trans, node->buffer, crcptr, sizeof(hammer_crc_t));
1303         --node->buffer->io.modify_refs; /* only want one ref */
1304 }
1305
1306 /*
1307  * Indicate that the specified modifications have been completed.
1308  *
1309  * Do not try to generate the crc here, it's very expensive to do and a
1310  * sequence of insertions or deletions can result in many calls to this
1311  * function on the same node.
1312  */
1313 static __inline void
1314 hammer_modify_node_done(hammer_node_t node)
1315 {
1316         node->flags |= HAMMER_NODE_CRCGOOD;
1317         if ((node->flags & HAMMER_NODE_NEEDSCRC) == 0) {
1318                 node->flags |= HAMMER_NODE_NEEDSCRC;
1319                 node->buffer->io.gencrc = 1;
1320                 hammer_ref_node(node);
1321         }
1322         hammer_modify_buffer_done(node->buffer);
1323 }
1324
1325 #define hammer_modify_volume_field(trans, vol, field)           \
1326         hammer_modify_volume(trans, vol, &(vol)->ondisk->field, \
1327                              sizeof((vol)->ondisk->field))
1328
1329 #define hammer_modify_node_field(trans, node, field)            \
1330         hammer_modify_node(trans, node, &(node)->ondisk->field, \
1331                              sizeof((node)->ondisk->field))
1332
1333 /*
1334  * The HAMMER_INODE_CAP_DIR_LOCAL_INO capability is set on newly
1335  * created directories for HAMMER version 2 or greater and causes
1336  * directory entries to be placed the inode localization zone in
1337  * the B-Tree instead of the misc zone.
1338  *
1339  * This greatly improves localization between directory entries and
1340  * inodes
1341  */
1342 static __inline u_int32_t
1343 hammer_dir_localization(hammer_inode_t dip)
1344 {
1345         if (dip->ino_data.cap_flags & HAMMER_INODE_CAP_DIR_LOCAL_INO)
1346                 return(HAMMER_LOCALIZE_INODE);
1347         else
1348                 return(HAMMER_LOCALIZE_MISC);
1349 }