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
  35 #ifndef VFS_HAMMER_HAMMER_H_
  36 #define VFS_HAMMER_HAMMER_H_
  37
  38 /*
  39  * This header file contains structures used internally by the HAMMERFS
  40  * implementation.  See hammer_disk.h for on-disk structures.
  41  */
  42
  43 #include <sys/param.h>
  44 #ifdef _KERNEL
  45 #include <sys/kernel.h>
  46 #include <sys/systm.h>
  47 #endif
  48 #include <sys/conf.h>
  49 #include <sys/tree.h>
  50 #include <sys/malloc.h>
  51 #include <sys/mount.h>
  52 #include <sys/vnode.h>
  53 #include <sys/proc.h>
  54 #include <sys/priv.h>
  55 #include <sys/dirent.h>
  56 #include <sys/stat.h>
  57 #include <sys/fcntl.h>
  58 #include <sys/lockf.h>
  59 #include <sys/file.h>
  60 #include <sys/event.h>
  61 #include <sys/buf.h>
  62 #include <sys/queue.h>
  63 #include <sys/ktr.h>
  64 #include <sys/limits.h>
  65 #include <sys/sysctl.h>
  66 #include <vm/swap_pager.h>
  67 #include <vm/vm_extern.h>
  68
  69 #include "hammer_disk.h"
  70 #include "hammer_mount.h"
  71 #include "hammer_ioctl.h"
  72
  73 #if defined(_KERNEL) || defined(_KERNEL_STRUCTURES)
  74
  75 MALLOC_DECLARE(M_HAMMER);
  76
  77 /*
  78  * Kernel trace
  79  */
  80 #if !defined(KTR_HAMMER)
  81 #define KTR_HAMMER      KTR_ALL
  82 #endif
  83 /* KTR_INFO_MASTER_EXTERN(hammer); */
  84
  85 /*
  86  * Misc structures
  87  */
  88 struct hammer_mount;
  89 struct hammer_inode;
  90 struct hammer_volume;
  91 struct hammer_buffer;
  92 struct hammer_node;
  93 struct hammer_undo;
  94 struct hammer_reserve;
  95 struct hammer_io;
  96
  97 /*
  98  * Key structure used for custom RB tree inode lookups.  This prototypes
  99  * the function hammer_ino_rb_tree_RB_LOOKUP_INFO(root, info).
 100  */
 101 typedef struct hammer_inode_info {
 102         int64_t         obj_id;         /* (key) object identifier */
 103         hammer_tid_t    obj_asof;       /* (key) snapshot transid or 0 */
 104         uint32_t        obj_localization; /* (key) pseudo-fs id for upper 16 bits */
 105         union {
 106                 struct hammer_btree_leaf_elm *leaf;
 107         } u;
 108 } *hammer_inode_info_t;
 109
 110 typedef enum hammer_transaction_type {
 111         HAMMER_TRANS_RO,
 112         HAMMER_TRANS_STD,
 113         HAMMER_TRANS_FLS
 114 } hammer_transaction_type_t;
 115
 116 /*
 117  * HAMMER Transaction tracking
 118  */
 119 struct hammer_transaction {
 120         hammer_transaction_type_t type;
 121         struct hammer_mount *hmp;
 122         hammer_tid_t    tid;
 123         uint64_t        time;
 124         uint32_t        time32;
 125         int             sync_lock_refs;
 126         int             flags;
 127         struct hammer_volume *rootvol;
 128 };
 129
 130 typedef struct hammer_transaction *hammer_transaction_t;
 131
 132 #define HAMMER_TRANSF_NEWINODE  0x0001
 133 #define HAMMER_TRANSF_DIDIO     0x0002
 134 #define HAMMER_TRANSF_CRCDOM    0x0004  /* EDOM on CRC error, less critical */
 135
 136 /*
 137  * HAMMER locks
 138  */
 139 struct hammer_lock {
 140         volatile u_int  refs;           /* active references */
 141         volatile u_int  lockval;        /* lock count and control bits */
 142         struct thread   *lowner;        /* owner if exclusively held */
 143         struct thread   *rowner;        /* owner if exclusively held */
 144 };
 145
 146 #define HAMMER_REFS_LOCKED      0x40000000      /* transition check */
 147 #define HAMMER_REFS_WANTED      0x20000000      /* transition check */
 148 #define HAMMER_REFS_CHECK       0x10000000      /* transition check */
 149
 150 #define HAMMER_REFS_FLAGS       (HAMMER_REFS_LOCKED | \
 151                                  HAMMER_REFS_WANTED | \
 152                                  HAMMER_REFS_CHECK)
 153
 154 #define HAMMER_LOCKF_EXCLUSIVE  0x40000000
 155 #define HAMMER_LOCKF_WANTED     0x20000000
 156
 157 static __inline int
 158 hammer_notlocked(struct hammer_lock *lock)
 159 {
 160         return(lock->lockval == 0);
 161 }
 162
 163 static __inline int
 164 hammer_islocked(struct hammer_lock *lock)
 165 {
 166         return(lock->lockval != 0);
 167 }
 168
 169 /*
 170  * Returns the number of refs on the object.
 171  */
 172 static __inline int
 173 hammer_isactive(struct hammer_lock *lock)
 174 {
 175         return(lock->refs & ~HAMMER_REFS_FLAGS);
 176 }
 177
 178 static __inline int
 179 hammer_oneref(struct hammer_lock *lock)
 180 {
 181         return((lock->refs & ~HAMMER_REFS_FLAGS) == 1);
 182 }
 183
 184 static __inline int
 185 hammer_norefs(struct hammer_lock *lock)
 186 {
 187         return((lock->refs & ~HAMMER_REFS_FLAGS) == 0);
 188 }
 189
 190 static __inline int
 191 hammer_norefsorlock(struct hammer_lock *lock)
 192 {
 193         return(lock->refs == 0);
 194 }
 195
 196 static __inline int
 197 hammer_refsorlock(struct hammer_lock *lock)
 198 {
 199         return(lock->refs != 0);
 200 }
 201
 202 /*
 203  * Return if we specifically own the lock exclusively.
 204  */
 205 static __inline int
 206 hammer_lock_excl_owned(struct hammer_lock *lock, thread_t td)
 207 {
 208         if ((lock->lockval & HAMMER_LOCKF_EXCLUSIVE) &&
 209             lock->lowner == td) {
 210                 return(1);
 211         }
 212         return(0);
 213 }
 214
 215 /*
 216  * Flush state, used by various structures
 217  */
 218 typedef enum hammer_inode_state {
 219         HAMMER_FST_IDLE,
 220         HAMMER_FST_SETUP,
 221         HAMMER_FST_FLUSH
 222 } hammer_inode_state_t;
 223
 224 /*
 225  * Pseudo-filesystem extended data tracking
 226  */
 227 struct hammer_pseudofs_inmem;
 228 RB_HEAD(hammer_pfs_rb_tree, hammer_pseudofs_inmem);
 229 RB_PROTOTYPE2(hammer_pfs_rb_tree, hammer_pseudofs_inmem, rb_node,
 230               hammer_pfs_rb_compare, uint32_t);
 231
 232 struct hammer_pseudofs_inmem {
 233         RB_ENTRY(hammer_pseudofs_inmem) rb_node;
 234         struct hammer_lock      lock;
 235         uint32_t                localization;
 236         hammer_tid_t            create_tid;
 237         int                     flags;
 238         udev_t                  fsid_udev;
 239         struct hammer_pseudofs_data pfsd;
 240 };
 241
 242 typedef struct hammer_pseudofs_inmem *hammer_pseudofs_inmem_t;
 243
 244 /*
 245  * Cache object ids.  A fixed number of objid cache structures are
 246  * created to reserve object id's for newly created files in multiples
 247  * of 100,000, localized to a particular directory, and recycled as
 248  * needed.  This allows parallel create operations in different
 249  * directories to retain fairly localized object ids which in turn
 250  * improves reblocking performance and layout.
 251  */
 252 #define OBJID_CACHE_SIZE        2048
 253 #define OBJID_CACHE_BULK_BITS   10              /* 10 bits (1024)       */
 254 #define OBJID_CACHE_BULK        (32 * 32)       /* two level (1024)     */
 255 #define OBJID_CACHE_BULK_MASK   (OBJID_CACHE_BULK - 1)
 256 #define OBJID_CACHE_BULK_MASK64 ((uint64_t)(OBJID_CACHE_BULK - 1))
 257
 258 typedef struct hammer_objid_cache {
 259         TAILQ_ENTRY(hammer_objid_cache) entry;
 260         struct hammer_inode             *dip;
 261         hammer_tid_t                    base_tid;
 262         int                             count;
 263         uint32_t                        bm0;
 264         uint32_t                        bm1[32];
 265 } *hammer_objid_cache_t;
 266
 267 /*
 268  * Associate an inode with a B-Tree node to cache search start positions
 269  */
 270 typedef struct hammer_node_cache {
 271         TAILQ_ENTRY(hammer_node_cache)  entry;
 272         struct hammer_node              *node;
 273         struct hammer_inode             *ip;
 274 } *hammer_node_cache_t;
 275
 276 TAILQ_HEAD(hammer_node_cache_list, hammer_node_cache);
 277
 278 /*
 279  * Live dedup cache
 280  */
 281 struct hammer_dedup_cache;
 282 RB_HEAD(hammer_dedup_crc_rb_tree, hammer_dedup_cache);
 283 RB_PROTOTYPE2(hammer_dedup_crc_rb_tree, hammer_dedup_cache, crc_entry,
 284                 hammer_dedup_crc_rb_compare, hammer_crc_t);
 285
 286 RB_HEAD(hammer_dedup_off_rb_tree, hammer_dedup_cache);
 287 RB_PROTOTYPE2(hammer_dedup_off_rb_tree, hammer_dedup_cache, off_entry,
 288                 hammer_dedup_off_rb_compare, hammer_off_t);
 289
 290 #define DEDUP_CACHE_SIZE        4096 /* XXX make it a dynamic tunable */
 291
 292 typedef struct hammer_dedup_cache {
 293         RB_ENTRY(hammer_dedup_cache) crc_entry;
 294         RB_ENTRY(hammer_dedup_cache) off_entry;
 295         TAILQ_ENTRY(hammer_dedup_cache) lru_entry;
 296         struct hammer_mount *hmp;
 297         int64_t obj_id;
 298         uint32_t localization;
 299         off_t file_offset;
 300         int bytes;
 301         hammer_off_t data_offset;
 302         hammer_crc_t crc;
 303 } *hammer_dedup_cache_t;
 304
 305 /*
 306  * Structure used to organize flush groups.  Flush groups must be
 307  * organized into chunks in order to avoid blowing out the UNDO FIFO.
 308  * Without this a 'sync' could end up flushing 50,000 inodes in a single
 309  * transaction.
 310  */
 311 RB_HEAD(hammer_fls_rb_tree, hammer_inode);
 312 RB_PROTOTYPE(hammer_fls_rb_tree, hammer_inode, rb_flsnode,
 313               hammer_ino_rb_compare);
 314
 315 struct hammer_flush_group {
 316         TAILQ_ENTRY(hammer_flush_group) flush_entry;
 317         struct hammer_fls_rb_tree       flush_tree;
 318         int                             seq;            /* our seq no */
 319         int                             total_count;    /* record load */
 320         int                             running;        /* group is running */
 321         int                             closed;
 322         int                             refs;
 323 };
 324
 325 typedef struct hammer_flush_group *hammer_flush_group_t;
 326
 327 TAILQ_HEAD(hammer_flush_group_list, hammer_flush_group);
 328
 329 /*
 330  * Structure used to represent an inode in-memory.
 331  *
 332  * The record and data associated with an inode may be out of sync with
 333  * the disk (xDIRTY flags), or not even on the disk at all (ONDISK flag
 334  * clear).
 335  *
 336  * An inode may also hold a cache of unsynchronized records, used for
 337  * database and directories only.  Unsynchronized regular file data is
 338  * stored in the buffer cache.
 339  *
 340  * NOTE: A file which is created and destroyed within the initial
 341  * synchronization period can wind up not doing any disk I/O at all.
 342  *
 343  * Finally, an inode may cache numerous disk-referencing B-Tree cursors.
 344  */
 345 RB_HEAD(hammer_ino_rb_tree, hammer_inode);
 346 RB_PROTOTYPEX(hammer_ino_rb_tree, INFO, hammer_inode, rb_node,
 347               hammer_ino_rb_compare, hammer_inode_info_t);
 348
 349 RB_HEAD(hammer_redo_rb_tree, hammer_inode);
 350 RB_PROTOTYPE2(hammer_redo_rb_tree, hammer_inode, rb_redonode,
 351               hammer_redo_rb_compare, hammer_off_t);
 352
 353 struct hammer_record;
 354 RB_HEAD(hammer_rec_rb_tree, hammer_record);
 355 RB_PROTOTYPEX(hammer_rec_rb_tree, INFO, hammer_record, rb_node,
 356               hammer_rec_rb_compare, hammer_btree_leaf_elm_t);
 357
 358 TAILQ_HEAD(hammer_record_list, hammer_record);
 359 TAILQ_HEAD(hammer_node_list, hammer_node);
 360
 361 struct hammer_inode {
 362         RB_ENTRY(hammer_inode)  rb_node;
 363         hammer_inode_state_t    flush_state;
 364         hammer_flush_group_t    flush_group;
 365         RB_ENTRY(hammer_inode)  rb_flsnode;     /* when on flush list */
 366         RB_ENTRY(hammer_inode)  rb_redonode;    /* when INODE_RDIRTY is set */
 367         struct hammer_record_list target_list;  /* target of dependant recs */
 368         int64_t                 obj_id;         /* (key) object identifier */
 369         hammer_tid_t            obj_asof;       /* (key) snapshot or 0 */
 370         uint32_t                obj_localization; /* (key) pseudo-fs id for upper 16 bits */
 371         struct hammer_mount     *hmp;
 372         hammer_objid_cache_t    objid_cache;
 373         int                     flags;
 374         int                     error;          /* flush error */
 375         int                     cursor_ip_refs; /* sanity */
 376 #if 0
 377         int                     cursor_exclreq_count;
 378 #endif
 379         int                     rsv_recs;
 380         struct vnode            *vp;
 381         hammer_pseudofs_inmem_t pfsm;
 382         struct lockf            advlock;
 383         struct hammer_lock      lock;           /* sync copy interlock */
 384         off_t                   trunc_off;
 385         struct hammer_btree_leaf_elm ino_leaf;  /* in-memory cache */
 386         struct hammer_inode_data ino_data;      /* in-memory cache */
 387         struct hammer_rec_rb_tree rec_tree;     /* in-memory cache */
 388         int                     rec_generation;
 389
 390         /*
 391          * search initiate cache
 392          * cache[0] - this inode
 393          * cache[1] - related data, the content depends on situations
 394          * cache[2] - for dip to cache ip to shortcut B-Tree search
 395          * cache[3] - related data copied from dip to a new ip's cache[1]
 396          */
 397         struct hammer_node_cache cache[4];
 398
 399         /*
 400          * When a demark is created to synchronize an inode to
 401          * disk, certain fields are copied so the front-end VOPs
 402          * can continue to run in parallel with the synchronization
 403          * occuring in the background.
 404          */
 405         int             sync_flags;             /* to-sync flags cache */
 406         off_t           sync_trunc_off;         /* to-sync truncation */
 407         off_t           save_trunc_off;         /* write optimization */
 408         struct hammer_btree_leaf_elm sync_ino_leaf; /* to-sync cache */
 409         struct hammer_inode_data sync_ino_data; /* to-sync cache */
 410         size_t          redo_count;
 411
 412         /*
 413          * Track the earliest offset in the UNDO/REDO FIFO containing
 414          * REDO records.  This is staged to the backend during flush
 415          * sequences.  While the inode is staged redo_fifo_next is used
 416          * to track the earliest offset for rotation into redo_fifo_start
 417          * on completion of the flush.
 418          */
 419         hammer_off_t    redo_fifo_start;
 420         hammer_off_t    redo_fifo_next;
 421 };
 422
 423 typedef struct hammer_inode *hammer_inode_t;
 424
 425 #define VTOI(vp)        ((struct hammer_inode *)(vp)->v_data)
 426
 427 /*
 428  * NOTE: DDIRTY does not include atime or mtime and does not include
 429  *       write-append size changes.  SDIRTY handles write-append size
 430  *       changes.
 431  *
 432  *       REDO indicates that REDO logging is active, creating a definitive
 433  *       stream of REDO records in the UNDO/REDO log for writes and
 434  *       truncations, including boundary records when/if REDO is turned off.
 435  *       REDO is typically enabled by fsync() and turned off if excessive
 436  *       writes without an fsync() occurs.
 437  *
 438  *       RDIRTY indicates that REDO records were laid down in the UNDO/REDO
 439  *       FIFO (even if REDO is turned off some might still be active) and
 440  *       still being tracked for this inode.  See hammer_redo.c
 441  */
 442 #define HAMMER_INODE_DDIRTY     0x0001  /* in-memory ino_data is dirty */
 443                                         /* (not including atime/mtime) */
 444 #define HAMMER_INODE_RSV_INODES 0x0002  /* hmp->rsv_inodes bumped */
 445 #define HAMMER_INODE_CONN_DOWN  0x0004  /* include in downward recursion */
 446 #define HAMMER_INODE_XDIRTY     0x0008  /* in-memory records */
 447 #define HAMMER_INODE_ONDISK     0x0010  /* inode is on-disk (else not yet) */
 448 #define HAMMER_INODE_FLUSH      0x0020  /* flush on last ref */
 449 #define HAMMER_INODE_DELETED    0x0080  /* inode delete (backend) */
 450 #define HAMMER_INODE_DELONDISK  0x0100  /* delete synchronized to disk */
 451 #define HAMMER_INODE_RO         0x0200  /* read-only (because of as-of) */
 452 #define HAMMER_INODE_RECSW      0x0400  /* waiting on data record flush */
 453 #define HAMMER_INODE_DONDISK    0x0800  /* data records may be on disk */
 454 #define HAMMER_INODE_BUFS       0x1000  /* dirty high level bps present */
 455 #define HAMMER_INODE_REFLUSH    0x2000  /* flush on dependancy / reflush */
 456 #define HAMMER_INODE_RECLAIM    0x4000  /* trying to reclaim */
 457 #define HAMMER_INODE_FLUSHW     0x8000  /* someone waiting for flush */
 458
 459 #define HAMMER_INODE_TRUNCATED  0x00010000
 460 #define HAMMER_INODE_DELETING   0x00020000 /* inode delete request (frontend)*/
 461 #define HAMMER_INODE_RESIGNAL   0x00040000 /* re-signal on re-flush */
 462 #define HAMMER_INODE_ATIME      0x00100000 /* in-memory atime modified */
 463 #define HAMMER_INODE_MTIME      0x00200000 /* in-memory mtime modified */
 464 #define HAMMER_INODE_WOULDBLOCK 0x00400000 /* re-issue to new flush group */
 465 #define HAMMER_INODE_DUMMY      0x00800000 /* dummy inode covering bad file */
 466 #define HAMMER_INODE_SDIRTY     0x01000000 /* in-memory ino_data.size is dirty*/
 467 #define HAMMER_INODE_REDO       0x02000000 /* REDO logging active */
 468 #define HAMMER_INODE_RDIRTY     0x04000000 /* REDO records active in fifo */
 469 #define HAMMER_INODE_SLAVEFLUSH 0x08000000 /* being flushed by slave */
 470
 471 #define HAMMER_INODE_MODMASK    (HAMMER_INODE_DDIRTY|HAMMER_INODE_SDIRTY|   \
 472                                  HAMMER_INODE_XDIRTY|HAMMER_INODE_BUFS|     \
 473                                  HAMMER_INODE_ATIME|HAMMER_INODE_MTIME|     \
 474                                  HAMMER_INODE_TRUNCATED|HAMMER_INODE_DELETING)
 475
 476 #define HAMMER_INODE_MODMASK_NOXDIRTY   \
 477                                 (HAMMER_INODE_MODMASK & ~HAMMER_INODE_XDIRTY)
 478
 479 #define HAMMER_INODE_MODMASK_NOREDO     \
 480                                 (HAMMER_INODE_DDIRTY|                       \
 481                                  HAMMER_INODE_XDIRTY|                       \
 482                                  HAMMER_INODE_TRUNCATED|HAMMER_INODE_DELETING)
 483
 484 #define HAMMER_FLUSH_SIGNAL     0x0001
 485 #define HAMMER_FLUSH_RECURSION  0x0002
 486
 487 /*
 488  * Used by the inode reclaim code to pipeline reclaims and avoid
 489  * blowing out kernel memory or letting the flusher get too far
 490  * behind.  The reclaim wakes up when count reaches 0 or the
 491  * timer expires.
 492  */
 493 struct hammer_reclaim {
 494         TAILQ_ENTRY(hammer_reclaim) entry;
 495         int     count;
 496 };
 497
 498 /*
 499  * Track who is creating the greatest burden on the
 500  * inode cache.
 501  */
 502 struct hammer_inostats {
 503         pid_t           pid;    /* track user process */
 504         int             ltick;  /* last tick */
 505         int             count;  /* count (degenerates) */
 506 };
 507
 508 #define HAMMER_INOSTATS_HSIZE   32
 509 #define HAMMER_INOSTATS_HMASK   (HAMMER_INOSTATS_HSIZE - 1)
 510
 511 /*
 512  * Structure used to represent an unsynchronized record in-memory.  These
 513  * records typically represent directory entries.  Only non-historical
 514  * records are kept in-memory.
 515  *
 516  * Records are organized as a per-inode RB-Tree.  If the inode is not
 517  * on disk then neither are any records and the in-memory record tree
 518  * represents the entire contents of the inode.  If the inode is on disk
 519  * then the on-disk B-Tree is scanned in parallel with the in-memory
 520  * RB-Tree to synthesize the current state of the file.
 521  *
 522  * Records are also used to enforce the ordering of directory create/delete
 523  * operations.  A new inode will not be flushed to disk unless its related
 524  * directory entry is also being flushed at the same time.  A directory entry
 525  * will not be removed unless its related inode is also being removed at the
 526  * same time.
 527  */
 528 typedef enum hammer_record_type {
 529         HAMMER_MEM_RECORD_GENERAL,      /* misc record */
 530         HAMMER_MEM_RECORD_INODE,        /* inode record */
 531         HAMMER_MEM_RECORD_ADD,          /* positive memory cache record */
 532         HAMMER_MEM_RECORD_DEL,          /* negative delete-on-disk record */
 533         HAMMER_MEM_RECORD_DATA          /* bulk-data record w/on-disk ref */
 534 } hammer_record_type_t;
 535
 536 struct hammer_record {
 537         RB_ENTRY(hammer_record)         rb_node;
 538         TAILQ_ENTRY(hammer_record)      target_entry;
 539         hammer_inode_state_t            flush_state;
 540         hammer_flush_group_t            flush_group;
 541         hammer_record_type_t            type;
 542         struct hammer_lock              lock;
 543         struct hammer_reserve           *resv;
 544         struct hammer_inode             *ip;
 545         struct hammer_inode             *target_ip;
 546         struct hammer_btree_leaf_elm    leaf;
 547         union hammer_data_ondisk        *data;
 548         int                             flags;
 549         int                             gflags;
 550         hammer_off_t                    zone2_offset;   /* direct-write only */
 551 };
 552
 553 typedef struct hammer_record *hammer_record_t;
 554
 555 /*
 556  * Record flags.  Note that FE can only be set by the frontend if the
 557  * record has not been interlocked by the backend w/ BE.
 558  */
 559 #define HAMMER_RECF_ALLOCDATA           0x0001
 560 #define HAMMER_RECF_ONRBTREE            0x0002
 561 #define HAMMER_RECF_DELETED_FE          0x0004  /* deleted (frontend) */
 562 #define HAMMER_RECF_DELETED_BE          0x0008  /* deleted (backend) */
 563 #define HAMMER_RECF_COMMITTED           0x0010  /* committed to the B-Tree */
 564 #define HAMMER_RECF_INTERLOCK_BE        0x0020  /* backend interlock */
 565 #define HAMMER_RECF_WANTED              0x0040  /* wanted by the frontend */
 566 #define HAMMER_RECF_DEDUPED             0x0080  /* will be live-dedup'ed */
 567 #define HAMMER_RECF_CONVERT_DELETE      0x0100  /* special case */
 568 #define HAMMER_RECF_REDO                0x1000  /* REDO was laid down */
 569
 570 /*
 571  * These flags must be separate to deal with SMP races
 572  */
 573 #define HAMMER_RECG_DIRECT_IO           0x0001  /* related direct I/O running*/
 574 #define HAMMER_RECG_DIRECT_WAIT         0x0002  /* related direct I/O running*/
 575 #define HAMMER_RECG_DIRECT_INVAL        0x0004  /* buffer alias invalidation */
 576 /*
 577  * hammer_create_at_cursor() and hammer_delete_at_cursor() flags.
 578  */
 579 #define HAMMER_CREATE_MODE_UMIRROR      0x0001
 580 #define HAMMER_CREATE_MODE_SYS          0x0002
 581
 582 #define HAMMER_DELETE_ADJUST            0x0001
 583 #define HAMMER_DELETE_DESTROY           0x0002
 584
 585 /*
 586  * In-memory structures representing on-disk structures.
 587  */
 588 RB_HEAD(hammer_vol_rb_tree, hammer_volume);
 589 RB_HEAD(hammer_buf_rb_tree, hammer_buffer);
 590 RB_HEAD(hammer_nod_rb_tree, hammer_node);
 591 RB_HEAD(hammer_und_rb_tree, hammer_undo);
 592 RB_HEAD(hammer_res_rb_tree, hammer_reserve);
 593 RB_HEAD(hammer_mod_rb_tree, hammer_io);
 594
 595 RB_PROTOTYPE2(hammer_vol_rb_tree, hammer_volume, rb_node,
 596               hammer_vol_rb_compare, int32_t);
 597 RB_PROTOTYPE2(hammer_buf_rb_tree, hammer_buffer, rb_node,
 598               hammer_buf_rb_compare, hammer_off_t);
 599 RB_PROTOTYPE2(hammer_nod_rb_tree, hammer_node, rb_node,
 600               hammer_nod_rb_compare, hammer_off_t);
 601 RB_PROTOTYPE2(hammer_und_rb_tree, hammer_undo, rb_node,
 602               hammer_und_rb_compare, hammer_off_t);
 603 RB_PROTOTYPE2(hammer_res_rb_tree, hammer_reserve, rb_node,
 604               hammer_res_rb_compare, hammer_off_t);
 605 RB_PROTOTYPE2(hammer_mod_rb_tree, hammer_io, rb_node,
 606               hammer_mod_rb_compare, hammer_off_t);
 607
 608 /*
 609  * IO management - embedded at the head of various in-memory structures
 610  *
 611  * VOLUME       - hammer_volume containing meta-data
 612  * META_BUFFER  - hammer_buffer containing meta-data
 613  * UNDO_BUFFER  - hammer_buffer containing undo-data
 614  * DATA_BUFFER  - hammer_buffer containing pure-data
 615  * DUMMY        - hammer_buffer not containing valid data
 616  *
 617  * Dirty volume headers and dirty meta-data buffers are locked until the
 618  * flusher can sequence them out.  Dirty pure-data buffers can be written.
 619  * Clean buffers can be passively released.
 620  */
 621 typedef enum hammer_io_type {
 622         HAMMER_STRUCTURE_VOLUME,
 623         HAMMER_STRUCTURE_META_BUFFER,
 624         HAMMER_STRUCTURE_UNDO_BUFFER,
 625         HAMMER_STRUCTURE_DATA_BUFFER,
 626         HAMMER_STRUCTURE_DUMMY
 627 } hammer_io_type_t;
 628
 629 /*
 630  * XXX: struct hammer_io can't directly embed LIST_ENTRY() at offset 0,
 631  * since a list head in struct buf expects a struct called worklist for
 632  * list entries.  HAMMER needs to define and use struct worklist.
 633  */
 634 struct worklist {
 635         LIST_ENTRY(worklist) node;
 636 };
 637
 638 TAILQ_HEAD(hammer_io_list, hammer_io);
 639 typedef struct hammer_io_list *hammer_io_list_t;
 640
 641 struct hammer_io {
 642         struct worklist         worklist; /* must be at offset 0 */
 643         struct hammer_lock      lock;
 644         enum hammer_io_type     type;
 645         struct hammer_mount     *hmp;
 646         struct hammer_volume    *volume;
 647         RB_ENTRY(hammer_io)     rb_node;     /* if modified */
 648         TAILQ_ENTRY(hammer_io)  iorun_entry; /* iorun_list */
 649         struct hammer_mod_rb_tree *mod_root;
 650         struct buf              *bp;
 651         int64_t                 offset;    /* volume offset */
 652         int                     bytes;     /* buffer cache buffer size */
 653         int                     modify_refs;
 654
 655         /*
 656          * These can be modified at any time by the backend while holding
 657          * io_token, due to bio_done and hammer_io_complete() callbacks.
 658          */
 659         u_int           running : 1;    /* bp write IO in progress */
 660         u_int           waiting : 1;    /* someone is waiting on us */
 661         u_int           ioerror : 1;    /* abort on io-error */
 662         u_int           unusedA : 29;
 663
 664         /*
 665          * These can only be modified by the frontend while holding
 666          * fs_token, or by the backend while holding the io interlocked
 667          * with no references (which will block the frontend when it
 668          * tries to reference it).
 669          *
 670          * WARNING! SMP RACES will create havoc if the callbacks ever tried
 671          *          to modify any of these outside the above restrictions.
 672          */
 673         u_int           modified : 1;   /* bp's data was modified */
 674         u_int           released : 1;   /* bp released (w/ B_LOCKED set) */
 675         u_int           waitdep : 1;    /* flush waits for dependancies */
 676         u_int           recovered : 1;  /* has recovery ref */
 677         u_int           waitmod : 1;    /* waiting for modify_refs */
 678         u_int           reclaim : 1;    /* reclaim requested */
 679         u_int           gencrc : 1;     /* crc needs to be generated */
 680         u_int           unusedB : 25;
 681 };
 682
 683 typedef struct hammer_io *hammer_io_t;
 684
 685 #define HAMMER_CLUSTER_SIZE     (64 * 1024)
 686 #if HAMMER_CLUSTER_SIZE > MAXBSIZE
 687 #undef  HAMMER_CLUSTER_SIZE
 688 #define HAMMER_CLUSTER_SIZE     MAXBSIZE
 689 #endif
 690
 691 /*
 692  * In-memory volume representing on-disk buffer
 693  */
 694 struct hammer_volume {
 695         struct hammer_io io; /* must be at offset 0 */
 696         RB_ENTRY(hammer_volume) rb_node;
 697         struct hammer_volume_ondisk *ondisk;
 698         int32_t vol_no;
 699         hammer_off_t maxbuf_off; /* Maximum buffer offset (zone-2) */
 700         char    *vol_name;
 701         struct vnode *devvp;
 702         int     vol_flags;
 703 };
 704
 705 typedef struct hammer_volume *hammer_volume_t;
 706
 707 #define HAMMER_ITOV(iop) ((hammer_volume_t)(iop))
 708
 709 /*
 710  * In-memory buffer representing an on-disk buffer.
 711  */
 712 struct hammer_buffer {
 713         struct hammer_io io; /* must be at offset 0 */
 714         RB_ENTRY(hammer_buffer) rb_node;
 715         void *ondisk;
 716         hammer_off_t zoneX_offset;
 717         hammer_off_t zone2_offset;
 718         struct hammer_reserve *resv;
 719         struct hammer_node_list clist;
 720 };
 721
 722 typedef struct hammer_buffer *hammer_buffer_t;
 723
 724 #define HAMMER_ITOB(iop) ((hammer_buffer_t)(iop))
 725
 726 /*
 727  * In-memory B-Tree node, representing an on-disk B-Tree node.
 728  *
 729  * This is a hang-on structure which is backed by a hammer_buffer,
 730  * and used for fine-grained locking of B-Tree nodes in order to
 731  * properly control lock ordering.
 732  */
 733 struct hammer_node {
 734         struct hammer_lock      lock;           /* node-by-node lock */
 735         TAILQ_ENTRY(hammer_node) entry;         /* per-buffer linkage */
 736         RB_ENTRY(hammer_node)   rb_node;        /* per-mount linkage */
 737         hammer_off_t            node_offset;    /* full offset spec */
 738         struct hammer_mount     *hmp;
 739         struct hammer_buffer    *buffer;        /* backing buffer */
 740         hammer_node_ondisk_t    ondisk;         /* ptr to on-disk structure */
 741         TAILQ_HEAD(, hammer_cursor) cursor_list;  /* deadlock recovery */
 742         struct hammer_node_cache_list cache_list; /* passive caches */
 743         int                     flags;
 744 #if 0
 745         int                     cursor_exclreq_count;
 746 #endif
 747 };
 748
 749 #define HAMMER_NODE_DELETED     0x0001
 750 #define HAMMER_NODE_FLUSH       0x0002
 751 #define HAMMER_NODE_CRCGOOD     0x0004
 752 #define HAMMER_NODE_NEEDSCRC    0x0008
 753 #define HAMMER_NODE_NEEDSMIRROR 0x0010
 754 #define HAMMER_NODE_CRCBAD      0x0020
 755 #define HAMMER_NODE_NONLINEAR   0x0040          /* linear heuristic */
 756
 757 #define HAMMER_NODE_CRCANY      (HAMMER_NODE_CRCGOOD | HAMMER_NODE_CRCBAD)
 758
 759 typedef struct hammer_node      *hammer_node_t;
 760
 761 /*
 762  * List of locked nodes.  This structure is used to lock potentially large
 763  * numbers of nodes as an aid for complex B-Tree operations.
 764  */
 765 struct hammer_node_lock;
 766 TAILQ_HEAD(hammer_node_lock_list, hammer_node_lock);
 767
 768 struct hammer_node_lock {
 769         TAILQ_ENTRY(hammer_node_lock) entry;
 770         struct hammer_node_lock_list  list;
 771         struct hammer_node_lock       *parent;
 772         hammer_node_t   node;
 773         hammer_node_ondisk_t copy;      /* copy of on-disk data */
 774         int             index;          /* index of this node in parent */
 775         int             count;          /* count children */
 776         int             flags;
 777 };
 778
 779 typedef struct hammer_node_lock *hammer_node_lock_t;
 780
 781 #define HAMMER_NODE_LOCK_UPDATED        0x0001
 782 #define HAMMER_NODE_LOCK_LCACHE         0x0002
 783
 784 /*
 785  * The reserve structure prevents the blockmap from allocating
 786  * out of a reserved big-block.  Such reservations are used by
 787  * the direct-write mechanism.
 788  *
 789  * The structure is also used to hold off on reallocations of
 790  * big-blocks from the freemap until flush dependancies have
 791  * been dealt with.
 792  */
 793 struct hammer_reserve {
 794         RB_ENTRY(hammer_reserve) rb_node;
 795         TAILQ_ENTRY(hammer_reserve) delay_entry;
 796         int             flg_no;
 797         int             flags;
 798         int             refs;
 799         int             zone;
 800         int             append_off;
 801         int32_t         bytes_free;
 802         hammer_off_t    zone_offset;
 803 };
 804
 805 typedef struct hammer_reserve *hammer_reserve_t;
 806
 807 #define HAMMER_RESF_ONDELAY     0x0001
 808 #define HAMMER_RESF_LAYER2FREE  0x0002
 809
 810 #include "hammer_cursor.h"
 811
 812 /*
 813  * The undo structure tracks recent undos to avoid laying down duplicate
 814  * undos within a flush group, saving us a significant amount of overhead.
 815  *
 816  * This is strictly a heuristic.
 817  */
 818 #define HAMMER_MAX_UNDOS                1024
 819 #define HAMMER_MAX_FLUSHERS             4
 820
 821 struct hammer_undo {
 822         RB_ENTRY(hammer_undo)   rb_node;
 823         TAILQ_ENTRY(hammer_undo) lru_entry;
 824         hammer_off_t            offset;
 825         int                     bytes;
 826 };
 827
 828 typedef struct hammer_undo *hammer_undo_t;
 829
 830 struct hammer_flusher_info;
 831 TAILQ_HEAD(hammer_flusher_info_list, hammer_flusher_info);
 832
 833 struct hammer_flusher {
 834         int             signal;         /* flusher thread sequencer */
 835         int             done;           /* last completed flush group */
 836         int             next;           /* next unallocated flg seqno */
 837         int             group_lock;     /* lock sequencing of the next flush */
 838         int             exiting;        /* request master exit */
 839         thread_t        td;             /* master flusher thread */
 840         hammer_tid_t    tid;            /* last flushed transaction id */
 841         int             finalize_want;          /* serialize finalization */
 842         struct hammer_lock finalize_lock;       /* serialize finalization */
 843         struct hammer_transaction trans;        /* shared transaction */
 844         struct hammer_flusher_info_list run_list;
 845         struct hammer_flusher_info_list ready_list;
 846 };
 847
 848 #define HAMMER_FLUSH_UNDOS_RELAXED      0
 849 #define HAMMER_FLUSH_UNDOS_FORCED       1
 850 #define HAMMER_FLUSH_UNDOS_AUTO         2
 851 /*
 852  * Internal hammer mount data structure
 853  */
 854 struct hammer_mount {
 855         struct mount *mp;
 856         struct hammer_ino_rb_tree rb_inos_root;
 857         struct hammer_redo_rb_tree rb_redo_root;
 858         struct hammer_vol_rb_tree rb_vols_root;
 859         struct hammer_nod_rb_tree rb_nods_root;
 860         struct hammer_und_rb_tree rb_undo_root;
 861         struct hammer_res_rb_tree rb_resv_root;
 862         struct hammer_buf_rb_tree rb_bufs_root;
 863         struct hammer_pfs_rb_tree rb_pfsm_root;
 864
 865         struct hammer_dedup_crc_rb_tree rb_dedup_crc_root;
 866         struct hammer_dedup_off_rb_tree rb_dedup_off_root;
 867
 868         struct hammer_volume *rootvol;
 869         struct hammer_base_elm root_btree_beg;
 870         struct hammer_base_elm root_btree_end;
 871
 872         struct malloc_type      *m_misc;
 873         struct malloc_type      *m_inodes;
 874
 875         int     flags;          /* HAMMER_MOUNT_xxx flags */
 876         int     hflags;
 877         int     ronly;
 878         int     nvolumes;
 879         int     master_id;      /* -1 or 0-15 for mirroring */
 880         int     version;        /* hammer filesystem version to use */
 881         int     rsv_inodes;     /* reserved space due to dirty inodes */
 882         int64_t rsv_databytes;  /* reserved space due to record data */
 883         int     rsv_recs;       /* reserved space due to dirty records */
 884         int     rsv_fromdelay;  /* big-blocks reserved due to flush delay */
 885         int     undo_rec_limit; /* based on size of undo area */
 886
 887         int     volume_to_remove; /* volume that is currently being removed */
 888
 889         int     count_inodes;   /* total number of inodes */
 890         int     count_iqueued;  /* inodes queued to flusher */
 891         int     count_reclaims; /* inodes pending reclaim by flusher */
 892
 893         struct hammer_flusher flusher;
 894
 895         u_int   check_interrupt;
 896         u_int   check_yield;
 897         uuid_t  fsid;
 898         struct hammer_mod_rb_tree volu_root;    /* dirty undo buffers */
 899         struct hammer_mod_rb_tree undo_root;    /* dirty undo buffers */
 900         struct hammer_mod_rb_tree data_root;    /* dirty data buffers */
 901         struct hammer_mod_rb_tree meta_root;    /* dirty meta bufs    */
 902         struct hammer_mod_rb_tree lose_root;    /* loose buffers      */
 903         long    locked_dirty_space;             /* meta/volu count    */
 904         long    io_running_space;               /* io_token */
 905         int     objid_cache_count;
 906         int     dedup_cache_count;
 907         int     error;                          /* critical I/O error */
 908         struct krate    krate;                  /* rate limited kprintf */
 909         struct krate    kdiag;                  /* rate limited kprintf */
 910         hammer_tid_t    asof;                   /* snapshot mount */
 911         hammer_tid_t    next_tid;
 912         hammer_tid_t    flush_tid1;             /* flusher tid sequencing */
 913         hammer_tid_t    flush_tid2;             /* flusher tid sequencing */
 914         int64_t copy_stat_freebigblocks;        /* number of free big-blocks */
 915         uint32_t        undo_seqno;             /* UNDO/REDO FIFO seqno */
 916         uint32_t        recover_stage2_seqno;   /* REDO recovery seqno */
 917         hammer_off_t    recover_stage2_offset;  /* REDO recovery offset */
 918
 919         struct netexport export;
 920         struct hammer_lock sync_lock;
 921         struct hammer_lock undo_lock;
 922         struct hammer_lock blkmap_lock;
 923         struct hammer_lock snapshot_lock;
 924         struct hammer_lock volume_lock;
 925         struct hammer_blockmap  blockmap[HAMMER_MAX_ZONES];
 926         struct hammer_undo      undos[HAMMER_MAX_UNDOS];
 927         int                     undo_alloc;
 928         TAILQ_HEAD(, hammer_undo)  undo_lru_list;
 929         TAILQ_HEAD(, hammer_reserve) delay_list;
 930         struct hammer_flush_group_list  flush_group_list;
 931         hammer_flush_group_t    fill_flush_group;
 932         hammer_flush_group_t    next_flush_group;
 933         TAILQ_HEAD(, hammer_objid_cache) objid_cache_list;
 934         TAILQ_HEAD(, hammer_dedup_cache) dedup_lru_list;
 935         hammer_dedup_cache_t    dedup_free_cache;
 936         TAILQ_HEAD(, hammer_reclaim) reclaim_list;
 937         TAILQ_HEAD(, hammer_io) iorun_list;
 938
 939         struct lwkt_token       fs_token;       /* high level */
 940         struct lwkt_token       io_token;       /* low level (IO callback) */
 941
 942         struct hammer_inostats  inostats[HAMMER_INOSTATS_HSIZE];
 943         uint64_t volume_map[4];  /* 256 bits bitfield */
 944 };
 945
 946 typedef struct hammer_mount     *hammer_mount_t;
 947
 948 #define HAMMER_MOUNT_CRITICAL_ERROR     0x0001
 949 #define HAMMER_MOUNT_FLUSH_RECOVERY     0x0002
 950 #define HAMMER_MOUNT_REDO_SYNC          0x0004
 951 #define HAMMER_MOUNT_REDO_RECOVERY_REQ  0x0008
 952 #define HAMMER_MOUNT_REDO_RECOVERY_RUN  0x0010
 953
 954 #define HAMMER_VOLUME_NUMBER_ADD(hmp, vol)                      \
 955         (hmp)->volume_map[(vol)->vol_no >> 6] |=                \
 956         ((uint64_t)1 << ((vol)->vol_no & ((1 << 6) - 1)))
 957
 958 #define HAMMER_VOLUME_NUMBER_DEL(hmp, vol)                      \
 959         (hmp)->volume_map[(vol)->vol_no >> 6] &=                \
 960         ~((uint64_t)1 << ((vol)->vol_no & ((1 << 6) - 1)))
 961
 962 #define HAMMER_VOLUME_NUMBER_IS_SET(hmp, n)                     \
 963         (((hmp)->volume_map[(n) >> 6] &                         \
 964         ((uint64_t)1 << ((n) & ((1 << 6) - 1)))) != 0)
 965
 966 #define HAMMER_VOLUME_NUMBER_FOREACH(hmp, n)                    \
 967         for (n = 0; n < HAMMER_MAX_VOLUMES; n++)                \
 968                 if (HAMMER_VOLUME_NUMBER_IS_SET(hmp, n))
 969
 970 /*
 971  * Minium buffer cache bufs required to rebalance the B-Tree.
 972  * This is because we must hold the children and the children's children
 973  * locked.  Even this might not be enough if things are horribly out
 974  * of balance.
 975  */
 976 #define HAMMER_REBALANCE_MIN_BUFS       \
 977         (HAMMER_BTREE_LEAF_ELMS * HAMMER_BTREE_LEAF_ELMS)
 978
 979 #endif  /* _KERNEL || _KERNEL_STRUCTURES */
 980
 981 #if defined(_KERNEL)
 982 /*
 983  * checkspace slop (8MB chunks), higher numbers are more conservative.
 984  */
 985 #define HAMMER_CHKSPC_REBLOCK   25
 986 #define HAMMER_CHKSPC_MIRROR    20
 987 #define HAMMER_CHKSPC_WRITE     20
 988 #define HAMMER_CHKSPC_CREATE    20
 989 #define HAMMER_CHKSPC_REMOVE    10
 990 #define HAMMER_CHKSPC_EMERGENCY 0
 991
 992 extern struct vop_ops hammer_vnode_vops;
 993 extern struct vop_ops hammer_spec_vops;
 994 extern struct vop_ops hammer_fifo_vops;
 995
 996 extern int hammer_debug_io;
 997 extern int hammer_debug_general;
 998 extern int hammer_debug_inode;
 999 extern int hammer_debug_locks;
1000 extern int hammer_debug_btree;
1001 extern int hammer_debug_tid;
1002 extern int hammer_debug_recover;
1003 extern int hammer_debug_critical;
1004 extern int hammer_cluster_enable;
1005 extern int hammer_live_dedup;
1006 extern int hammer_tdmux_ticks;
1007 extern int hammer_count_fsyncs;
1008 extern int hammer_count_inodes;
1009 extern int hammer_count_iqueued;
1010 extern int hammer_count_reclaims;
1011 extern int hammer_count_records;
1012 extern int hammer_count_record_datas;
1013 extern int hammer_count_volumes;
1014 extern int hammer_count_buffers;
1015 extern int hammer_count_nodes;
1016 extern int64_t hammer_count_extra_space_used;
1017 extern int64_t hammer_stats_btree_lookups;
1018 extern int64_t hammer_stats_btree_searches;
1019 extern int64_t hammer_stats_btree_inserts;
1020 extern int64_t hammer_stats_btree_deletes;
1021 extern int64_t hammer_stats_btree_elements;
1022 extern int64_t hammer_stats_btree_splits;
1023 extern int64_t hammer_stats_btree_iterations;
1024 extern int64_t hammer_stats_btree_root_iterations;
1025 extern int64_t hammer_stats_record_iterations;
1026 extern int64_t hammer_stats_file_read;
1027 extern int64_t hammer_stats_file_write;
1028 extern int64_t hammer_stats_file_iopsr;
1029 extern int64_t hammer_stats_file_iopsw;
1030 extern int64_t hammer_stats_disk_read;
1031 extern int64_t hammer_stats_disk_write;
1032 extern int64_t hammer_stats_inode_flushes;
1033 extern int64_t hammer_stats_commits;
1034 extern int64_t hammer_stats_undo;
1035 extern int64_t hammer_stats_redo;
1036 extern long hammer_count_dirtybufspace;
1037 extern int hammer_count_refedbufs;
1038 extern int hammer_count_reservations;
1039 extern long hammer_count_io_running_read;
1040 extern long hammer_count_io_running_write;
1041 extern int hammer_count_io_locked;
1042 extern long hammer_limit_dirtybufspace;
1043 extern int hammer_limit_recs;
1044 extern int hammer_limit_inode_recs;
1045 extern int hammer_limit_reclaims;
1046 extern int hammer_live_dedup_cache_size;
1047 extern int hammer_limit_redo;
1048 extern int hammer_verify_zone;
1049 extern int hammer_verify_data;
1050 extern int hammer_double_buffer;
1051 extern int hammer_btree_full_undo;
1052 extern int hammer_yield_check;
1053 extern int hammer_fsync_mode;
1054 extern int hammer_autoflush;
1055 extern int64_t hammer_contention_count;
1056
1057 extern int64_t hammer_live_dedup_vnode_bcmps;
1058 extern int64_t hammer_live_dedup_device_bcmps;
1059 extern int64_t hammer_live_dedup_findblk_failures;
1060 extern int64_t hammer_live_dedup_bmap_saves;
1061
1062 void    hammer_critical_error(hammer_mount_t hmp, hammer_inode_t ip,
1063                         int error, const char *msg);
1064 int     hammer_vop_inactive(struct vop_inactive_args *);
1065 int     hammer_vop_reclaim(struct vop_reclaim_args *);
1066 int     hammer_get_vnode(struct hammer_inode *ip, struct vnode **vpp);
1067 struct hammer_inode *hammer_get_inode(hammer_transaction_t trans,
1068                         hammer_inode_t dip, int64_t obj_id,
1069                         hammer_tid_t asof, uint32_t localization,
1070                         int flags, int *errorp);
1071 struct hammer_inode *hammer_get_dummy_inode(hammer_transaction_t trans,
1072                         hammer_inode_t dip, int64_t obj_id,
1073                         hammer_tid_t asof, uint32_t localization,
1074                         int flags, int *errorp);
1075 struct hammer_inode *hammer_find_inode(hammer_transaction_t trans,
1076                         int64_t obj_id, hammer_tid_t asof,
1077                         uint32_t localization);
1078 void    hammer_scan_inode_snapshots(hammer_mount_t hmp,
1079                         hammer_inode_info_t iinfo,
1080                         int (*callback)(hammer_inode_t ip, void *data),
1081                         void *data);
1082 void    hammer_put_inode(struct hammer_inode *ip);
1083 void    hammer_put_inode_ref(struct hammer_inode *ip);
1084 void    hammer_inode_waitreclaims(hammer_transaction_t trans);
1085 void    hammer_inode_dirty(struct hammer_inode *ip);
1086
1087 int     hammer_unload_volume(hammer_volume_t volume, void *data);
1088 int     hammer_adjust_volume_mode(hammer_volume_t volume, void *data __unused);
1089
1090 int     hammer_unload_buffer(hammer_buffer_t buffer, void *data);
1091 int     hammer_install_volume(hammer_mount_t hmp, const char *volname,
1092                         struct vnode *devvp, void *data);
1093 int     hammer_mountcheck_volumes(hammer_mount_t hmp);
1094 int     hammer_get_installed_volumes(hammer_mount_t hmp);
1095
1096 int     hammer_mem_add(hammer_record_t record);
1097 int     hammer_ip_lookup(hammer_cursor_t cursor);
1098 int     hammer_ip_first(hammer_cursor_t cursor);
1099 int     hammer_ip_next(hammer_cursor_t cursor);
1100 int     hammer_ip_resolve_data(hammer_cursor_t cursor);
1101 int     hammer_ip_delete_record(hammer_cursor_t cursor, hammer_inode_t ip,
1102                         hammer_tid_t tid);
1103 int     hammer_create_at_cursor(hammer_cursor_t cursor,
1104                         hammer_btree_leaf_elm_t leaf, void *udata, int mode);
1105 int     hammer_delete_at_cursor(hammer_cursor_t cursor, int delete_flags,
1106                         hammer_tid_t delete_tid, uint32_t delete_ts,
1107                         int track, int64_t *stat_bytes);
1108 int     hammer_ip_check_directory_empty(hammer_transaction_t trans,
1109                         hammer_inode_t ip);
1110 int     hammer_sync_hmp(hammer_mount_t hmp, int waitfor);
1111 int     hammer_queue_inodes_flusher(hammer_mount_t hmp, int waitfor);
1112
1113 hammer_record_t
1114         hammer_alloc_mem_record(hammer_inode_t ip, int data_len);
1115 void    hammer_flush_record_done(hammer_record_t record, int error);
1116 void    hammer_wait_mem_record_ident(hammer_record_t record, const char *ident);
1117 void    hammer_rel_mem_record(hammer_record_t record);
1118
1119 int     hammer_cursor_up(hammer_cursor_t cursor);
1120 int     hammer_cursor_up_locked(hammer_cursor_t cursor);
1121 int     hammer_cursor_down(hammer_cursor_t cursor);
1122 int     hammer_cursor_upgrade(hammer_cursor_t cursor);
1123 int     hammer_cursor_upgrade_node(hammer_cursor_t cursor);
1124 void    hammer_cursor_downgrade(hammer_cursor_t cursor);
1125 int     hammer_cursor_upgrade2(hammer_cursor_t c1, hammer_cursor_t c2);
1126 void    hammer_cursor_downgrade2(hammer_cursor_t c1, hammer_cursor_t c2);
1127 int     hammer_cursor_seek(hammer_cursor_t cursor, hammer_node_t node,
1128                         int index);
1129 void    hammer_lock_ex_ident(struct hammer_lock *lock, const char *ident);
1130 int     hammer_lock_ex_try(struct hammer_lock *lock);
1131 void    hammer_lock_sh(struct hammer_lock *lock);
1132 int     hammer_lock_sh_try(struct hammer_lock *lock);
1133 int     hammer_lock_upgrade(struct hammer_lock *lock, int shcount);
1134 void    hammer_lock_downgrade(struct hammer_lock *lock, int shcount);
1135 int     hammer_lock_status(struct hammer_lock *lock);
1136 void    hammer_unlock(struct hammer_lock *lock);
1137 void    hammer_ref(struct hammer_lock *lock);
1138 int     hammer_ref_interlock(struct hammer_lock *lock);
1139 int     hammer_ref_interlock_true(struct hammer_lock *lock);
1140 void    hammer_ref_interlock_done(struct hammer_lock *lock);
1141 void    hammer_rel(struct hammer_lock *lock);
1142 int     hammer_rel_interlock(struct hammer_lock *lock, int locked);
1143 void    hammer_rel_interlock_done(struct hammer_lock *lock, int orig_locked);
1144 int     hammer_get_interlock(struct hammer_lock *lock);
1145 int     hammer_try_interlock_norefs(struct hammer_lock *lock);
1146 void    hammer_put_interlock(struct hammer_lock *lock, int error);
1147
1148 void    hammer_sync_lock_ex(hammer_transaction_t trans);
1149 void    hammer_sync_lock_sh(hammer_transaction_t trans);
1150 int     hammer_sync_lock_sh_try(hammer_transaction_t trans);
1151 void    hammer_sync_unlock(hammer_transaction_t trans);
1152
1153 uint32_t hammer_to_unix_xid(uuid_t *uuid);
1154 void hammer_guid_to_uuid(uuid_t *uuid, uint32_t guid);
1155 void    hammer_time_to_timespec(uint64_t xtime, struct timespec *ts);
1156 uint64_t hammer_timespec_to_time(struct timespec *ts);
1157 int     hammer_str_to_tid(const char *str, int *ispfsp,
1158                         hammer_tid_t *tidp, uint32_t *localizationp);
1159 hammer_tid_t hammer_alloc_objid(hammer_mount_t hmp, hammer_inode_t dip,
1160                         int64_t namekey);
1161 void hammer_clear_objid(hammer_inode_t dip);
1162 void hammer_destroy_objid_cache(hammer_mount_t hmp);
1163
1164 int hammer_dedup_crc_rb_compare(hammer_dedup_cache_t dc1,
1165                         hammer_dedup_cache_t dc2);
1166 int hammer_dedup_off_rb_compare(hammer_dedup_cache_t dc1,
1167                         hammer_dedup_cache_t dc2);
1168 hammer_dedup_cache_t hammer_dedup_cache_add(hammer_inode_t ip,
1169                         hammer_btree_leaf_elm_t leaf);
1170 hammer_dedup_cache_t hammer_dedup_cache_lookup(hammer_mount_t hmp,
1171                         hammer_crc_t crc);
1172 void hammer_dedup_cache_inval(hammer_mount_t hmp, hammer_off_t base_offset);
1173 void hammer_destroy_dedup_cache(hammer_mount_t hmp);
1174 void hammer_dump_dedup_cache(hammer_mount_t hmp);
1175 int hammer_dedup_validate(hammer_dedup_cache_t dcp, int zone, int bytes,
1176                         void *data);
1177
1178 int hammer_enter_undo_history(hammer_mount_t hmp, hammer_off_t offset,
1179                         int bytes);
1180 void hammer_clear_undo_history(hammer_mount_t hmp);
1181 enum vtype hammer_get_vnode_type(uint8_t obj_type);
1182 int hammer_get_dtype(uint8_t obj_type);
1183 uint8_t hammer_get_obj_type(enum vtype vtype);
1184 int64_t hammer_directory_namekey(hammer_inode_t dip, const void *name, int len,
1185                         uint32_t *max_iterationsp);
1186 int     hammer_nohistory(hammer_inode_t ip);
1187
1188 int     hammer_init_cursor(hammer_transaction_t trans, hammer_cursor_t cursor,
1189                         hammer_node_cache_t cache, hammer_inode_t ip);
1190 void    hammer_normalize_cursor(hammer_cursor_t cursor);
1191 void    hammer_done_cursor(hammer_cursor_t cursor);
1192 int     hammer_recover_cursor(hammer_cursor_t cursor);
1193 void    hammer_unlock_cursor(hammer_cursor_t cursor);
1194 int     hammer_lock_cursor(hammer_cursor_t cursor);
1195 hammer_cursor_t hammer_push_cursor(hammer_cursor_t ocursor);
1196 void    hammer_pop_cursor(hammer_cursor_t ocursor, hammer_cursor_t ncursor);
1197
1198 void    hammer_cursor_replaced_node(hammer_node_t onode, hammer_node_t nnode);
1199 void    hammer_cursor_removed_node(hammer_node_t onode, hammer_node_t parent,
1200                         int index);
1201 void    hammer_cursor_split_node(hammer_node_t onode, hammer_node_t nnode,
1202                         int index);
1203 void    hammer_cursor_moved_element(hammer_node_t oparent, int pindex,
1204                         hammer_node_t onode, int oindex,
1205                         hammer_node_t nnode, int nindex);
1206 void    hammer_cursor_parent_changed(hammer_node_t node, hammer_node_t oparent,
1207                         hammer_node_t nparent, int nindex);
1208 void    hammer_cursor_inserted_element(hammer_node_t node, int index);
1209 void    hammer_cursor_deleted_element(hammer_node_t node, int index);
1210 void    hammer_cursor_invalidate_cache(hammer_cursor_t cursor);
1211
1212 int     hammer_btree_lookup(hammer_cursor_t cursor);
1213 int     hammer_btree_first(hammer_cursor_t cursor);
1214 int     hammer_btree_last(hammer_cursor_t cursor);
1215 int     hammer_btree_extract(hammer_cursor_t cursor, int flags);
1216 int     hammer_btree_iterate(hammer_cursor_t cursor);
1217 int     hammer_btree_iterate_reverse(hammer_cursor_t cursor);
1218 int     hammer_btree_insert(hammer_cursor_t cursor,
1219                             hammer_btree_leaf_elm_t elm, int *doprop);
1220 int     hammer_btree_delete(hammer_cursor_t cursor, int *ndelete);
1221 void    hammer_btree_do_propagation(hammer_cursor_t cursor,
1222                             hammer_pseudofs_inmem_t pfsm,
1223                             hammer_btree_leaf_elm_t leaf);
1224 int     hammer_btree_cmp(hammer_base_elm_t key1, hammer_base_elm_t key2);
1225 int     hammer_btree_chkts(hammer_tid_t ts, hammer_base_elm_t key);
1226 int     hammer_btree_correct_rhb(hammer_cursor_t cursor, hammer_tid_t tid);
1227 int     hammer_btree_correct_lhb(hammer_cursor_t cursor, hammer_tid_t tid);
1228
1229 int     btree_set_parent_of_child(hammer_transaction_t trans,
1230                         hammer_node_t node,
1231                         hammer_btree_elm_t elm);
1232 void    hammer_node_lock_init(hammer_node_lock_t parent, hammer_node_t node);
1233 void    hammer_btree_lcache_init(hammer_mount_t hmp, hammer_node_lock_t lcache,
1234                         int depth);
1235 void    hammer_btree_lcache_free(hammer_mount_t hmp, hammer_node_lock_t lcache);
1236 int     hammer_btree_lock_children(hammer_cursor_t cursor, int depth,
1237                         hammer_node_lock_t parent,
1238                         hammer_node_lock_t lcache);
1239 void    hammer_btree_lock_copy(hammer_cursor_t cursor,
1240                         hammer_node_lock_t parent);
1241 int     hammer_btree_sync_copy(hammer_cursor_t cursor,
1242                         hammer_node_lock_t parent);
1243 void    hammer_btree_unlock_children(hammer_mount_t hmp,
1244                         hammer_node_lock_t parent,
1245                         hammer_node_lock_t lcache);
1246 int     hammer_btree_search_node(hammer_base_elm_t elm, hammer_node_ondisk_t node);
1247 hammer_node_t hammer_btree_get_parent(hammer_transaction_t trans,
1248                         hammer_node_t node, int *parent_indexp,
1249                         int *errorp, int try_exclusive);
1250
1251 void    hammer_print_btree_node(hammer_node_ondisk_t ondisk);
1252 void    hammer_print_btree_elm(hammer_btree_elm_t elm);
1253
1254 void    *hammer_bread(hammer_mount_t hmp, hammer_off_t off,
1255                         int *errorp, struct hammer_buffer **bufferp);
1256 void    *hammer_bnew(hammer_mount_t hmp, hammer_off_t off,
1257                         int *errorp, struct hammer_buffer **bufferp);
1258 void    *hammer_bread_ext(hammer_mount_t hmp, hammer_off_t off, int bytes,
1259                         int *errorp, struct hammer_buffer **bufferp);
1260 void    *hammer_bnew_ext(hammer_mount_t hmp, hammer_off_t off, int bytes,
1261                         int *errorp, struct hammer_buffer **bufferp);
1262
1263 hammer_volume_t hammer_get_root_volume(hammer_mount_t hmp, int *errorp);
1264
1265 hammer_volume_t hammer_get_volume(hammer_mount_t hmp,
1266                         int32_t vol_no, int *errorp);
1267 hammer_buffer_t hammer_get_buffer(hammer_mount_t hmp, hammer_off_t buf_offset,
1268                         int bytes, int isnew, int *errorp);
1269 void            hammer_sync_buffers(hammer_mount_t hmp,
1270                         hammer_off_t base_offset, int bytes);
1271 int             hammer_del_buffers(hammer_mount_t hmp,
1272                         hammer_off_t base_offset,
1273                         hammer_off_t zone2_offset, int bytes,
1274                         int report_conflicts);
1275
1276 int             hammer_ref_volume(hammer_volume_t volume);
1277 int             hammer_ref_buffer(hammer_buffer_t buffer);
1278 void            hammer_flush_buffer_nodes(hammer_buffer_t buffer);
1279
1280 void            hammer_rel_volume(hammer_volume_t volume, int locked);
1281 void            hammer_rel_buffer(hammer_buffer_t buffer, int locked);
1282
1283 int             hammer_vfs_export(struct mount *mp, int op,
1284                         const struct export_args *export);
1285 hammer_node_t   hammer_get_node(hammer_transaction_t trans,
1286                         hammer_off_t node_offset, int isnew, int *errorp);
1287 void            hammer_ref_node(hammer_node_t node);
1288 hammer_node_t   hammer_ref_node_safe(hammer_transaction_t trans,
1289                         hammer_node_cache_t cache, int *errorp);
1290 void            hammer_rel_node(hammer_node_t node);
1291 void            hammer_delete_node(hammer_transaction_t trans,
1292                         hammer_node_t node);
1293 void            hammer_cache_node(hammer_node_cache_t cache,
1294                         hammer_node_t node);
1295 void            hammer_uncache_node(hammer_node_cache_t cache);
1296 void            hammer_flush_node(hammer_node_t node, int locked);
1297
1298 void hammer_dup_buffer(struct hammer_buffer **bufferp,
1299                         struct hammer_buffer *buffer);
1300 hammer_node_t hammer_alloc_btree(hammer_transaction_t trans,
1301                         hammer_off_t hint, int *errorp);
1302 void *hammer_alloc_data(hammer_transaction_t trans, int32_t data_len,
1303                         uint16_t rec_type, hammer_off_t *data_offsetp,
1304                         struct hammer_buffer **data_bufferp,
1305                         hammer_off_t hint, int *errorp);
1306
1307 int hammer_generate_undo(hammer_transaction_t trans,
1308                         hammer_off_t zone_offset, void *base, int len);
1309 int hammer_generate_redo(hammer_transaction_t trans, hammer_inode_t ip,
1310                         hammer_off_t file_offset, uint32_t flags,
1311                         void *base, int len);
1312 void hammer_generate_redo_sync(hammer_transaction_t trans);
1313 void hammer_redo_fifo_start_flush(hammer_inode_t ip);
1314 void hammer_redo_fifo_end_flush(hammer_inode_t ip);
1315
1316 void hammer_format_undo(void *base, uint32_t seqno);
1317 int hammer_upgrade_undo_4(hammer_transaction_t trans);
1318
1319 void hammer_put_volume(struct hammer_volume *volume, int flush);
1320 void hammer_put_buffer(struct hammer_buffer *buffer, int flush);
1321
1322 hammer_off_t hammer_freemap_alloc(hammer_transaction_t trans,
1323                         hammer_off_t owner, int *errorp);
1324 void hammer_freemap_free(hammer_transaction_t trans, hammer_off_t phys_offset,
1325                         hammer_off_t owner, int *errorp);
1326 int _hammer_checkspace(hammer_mount_t hmp, int slop, int64_t *resp);
1327 hammer_off_t hammer_blockmap_alloc(hammer_transaction_t trans, int zone,
1328                         int bytes, hammer_off_t hint, int *errorp);
1329 hammer_reserve_t hammer_blockmap_reserve(hammer_mount_t hmp, int zone,
1330                         int bytes, hammer_off_t *zone_offp, int *errorp);
1331 hammer_reserve_t hammer_blockmap_reserve_dedup(hammer_mount_t hmp, int zone,
1332                         int bytes, hammer_off_t zone_offset, int *errorp);
1333 void hammer_blockmap_reserve_complete(hammer_mount_t hmp,
1334                         hammer_reserve_t resv);
1335 void hammer_reserve_clrdelay(hammer_mount_t hmp, hammer_reserve_t resv);
1336 void hammer_blockmap_free(hammer_transaction_t trans,
1337                         hammer_off_t zone_offset, int bytes);
1338 int hammer_blockmap_dedup(hammer_transaction_t trans,
1339                         hammer_off_t zone_offset, int bytes);
1340 int hammer_blockmap_finalize(hammer_transaction_t trans,
1341                         hammer_reserve_t resv,
1342                         hammer_off_t zone_offset, int bytes);
1343 int hammer_blockmap_getfree(hammer_mount_t hmp, hammer_off_t zone_offset,
1344                         int *curp, int *errorp);
1345 hammer_off_t hammer_blockmap_lookup_verify(hammer_mount_t hmp,
1346                         hammer_off_t zone_offset, int *errorp);
1347
1348 hammer_off_t hammer_undo_lookup(hammer_mount_t hmp, hammer_off_t zone_offset,
1349                         int *errorp);
1350 int64_t hammer_undo_used(hammer_transaction_t trans);
1351 int64_t hammer_undo_space(hammer_transaction_t trans);
1352 int64_t hammer_undo_max(hammer_mount_t hmp);
1353 int hammer_undo_reclaim(hammer_io_t io);
1354
1355 void hammer_start_transaction(struct hammer_transaction *trans,
1356                               struct hammer_mount *hmp);
1357 void hammer_simple_transaction(struct hammer_transaction *trans,
1358                               struct hammer_mount *hmp);
1359 void hammer_start_transaction_fls(struct hammer_transaction *trans,
1360                                   struct hammer_mount *hmp);
1361 void hammer_done_transaction(struct hammer_transaction *trans);
1362 hammer_tid_t hammer_alloc_tid(hammer_mount_t hmp, int count);
1363
1364 void hammer_modify_inode(hammer_transaction_t trans, hammer_inode_t ip, int flags);
1365 void hammer_flush_inode(hammer_inode_t ip, int flags);
1366 void hammer_flush_inode_done(hammer_inode_t ip, int error);
1367 void hammer_wait_inode(hammer_inode_t ip);
1368
1369 int  hammer_create_inode(struct hammer_transaction *trans, struct vattr *vap,
1370                         struct ucred *cred, struct hammer_inode *dip,
1371                         const char *name, int namelen,
1372                         hammer_pseudofs_inmem_t pfsm,
1373                         struct hammer_inode **ipp);
1374 void hammer_rel_inode(hammer_inode_t ip, int flush);
1375 int hammer_reload_inode(hammer_inode_t ip, void *arg __unused);
1376 int hammer_ino_rb_compare(hammer_inode_t ip1, hammer_inode_t ip2);
1377 int hammer_redo_rb_compare(hammer_inode_t ip1, hammer_inode_t ip2);
1378 int hammer_destroy_inode_callback(hammer_inode_t ip, void *data __unused);
1379
1380 int hammer_sync_inode(hammer_transaction_t trans, hammer_inode_t ip);
1381 void hammer_test_inode(hammer_inode_t dip);
1382 void hammer_inode_unloadable_check(hammer_inode_t ip, int getvp);
1383 int hammer_update_atime_quick(hammer_inode_t ip);
1384
1385 int  hammer_ip_add_directory(struct hammer_transaction *trans,
1386                         hammer_inode_t dip, const char *name, int bytes,
1387                         hammer_inode_t nip);
1388 int  hammer_ip_del_directory(struct hammer_transaction *trans,
1389                         hammer_cursor_t cursor, hammer_inode_t dip,
1390                         hammer_inode_t ip);
1391 void hammer_ip_replace_bulk(hammer_mount_t hmp, hammer_record_t record);
1392 hammer_record_t hammer_ip_add_bulk(hammer_inode_t ip, off_t file_offset,
1393                         void *data, int bytes, int *errorp);
1394 int  hammer_ip_frontend_trunc(struct hammer_inode *ip, off_t file_size);
1395 int  hammer_ip_add_record(struct hammer_transaction *trans,
1396                         hammer_record_t record);
1397 int  hammer_ip_delete_range(hammer_cursor_t cursor, hammer_inode_t ip,
1398                         int64_t ran_beg, int64_t ran_end, int truncating);
1399 int  hammer_ip_delete_clean(hammer_cursor_t cursor, hammer_inode_t ip,
1400                         int *countp);
1401 int  hammer_ip_sync_data(hammer_cursor_t cursor, hammer_inode_t ip,
1402                         int64_t offset, void *data, int bytes);
1403 int  hammer_ip_sync_record_cursor(hammer_cursor_t cursor, hammer_record_t rec);
1404 hammer_pseudofs_inmem_t  hammer_load_pseudofs(hammer_transaction_t trans,
1405                         uint32_t localization, int *errorp);
1406 int  hammer_mkroot_pseudofs(hammer_transaction_t trans, struct ucred *cred,
1407                         hammer_pseudofs_inmem_t pfsm, hammer_inode_t dip);
1408 int  hammer_save_pseudofs(hammer_transaction_t trans,
1409                         hammer_pseudofs_inmem_t pfsm);
1410 int  hammer_unload_pseudofs(hammer_transaction_t trans, uint32_t localization);
1411 void hammer_rel_pseudofs(hammer_mount_t hmp, hammer_pseudofs_inmem_t pfsm);
1412 int hammer_ioctl(hammer_inode_t ip, u_long com, caddr_t data, int fflag,
1413                         struct ucred *cred);
1414
1415 void hammer_io_init(hammer_io_t io, hammer_volume_t volume,
1416                         enum hammer_io_type type);
1417 int hammer_io_read(struct vnode *devvp, struct hammer_io *io, int limit);
1418 void hammer_io_advance(struct hammer_io *io);
1419 int hammer_io_new(struct vnode *devvp, struct hammer_io *io);
1420 int hammer_io_inval(hammer_volume_t volume, hammer_off_t zone2_offset);
1421 struct buf *hammer_io_release(struct hammer_io *io, int flush);
1422 void hammer_io_flush(struct hammer_io *io, int reclaim);
1423 void hammer_io_wait(struct hammer_io *io);
1424 void hammer_io_waitdep(struct hammer_io *io);
1425 void hammer_io_wait_all(hammer_mount_t hmp, const char *ident, int doflush);
1426 int hammer_io_direct_read(hammer_mount_t hmp, struct bio *bio,
1427                         hammer_btree_leaf_elm_t leaf);
1428 int hammer_io_indirect_read(hammer_mount_t hmp, struct bio *bio,
1429                         hammer_btree_leaf_elm_t leaf);
1430 int hammer_io_direct_write(hammer_mount_t hmp, struct bio *bio,
1431                         hammer_record_t record);
1432 void hammer_io_direct_wait(hammer_record_t record);
1433 void hammer_io_direct_uncache(hammer_mount_t hmp, hammer_btree_leaf_elm_t leaf);
1434 void hammer_io_write_interlock(hammer_io_t io);
1435 void hammer_io_done_interlock(hammer_io_t io);
1436 void hammer_io_clear_modify(struct hammer_io *io, int inval);
1437 void hammer_io_clear_modlist(struct hammer_io *io);
1438 void hammer_io_flush_sync(hammer_mount_t hmp);
1439 void hammer_io_clear_error(struct hammer_io *io);
1440 void hammer_io_clear_error_noassert(struct hammer_io *io);
1441 void hammer_io_notmeta(hammer_buffer_t buffer);
1442 void hammer_io_limit_backlog(hammer_mount_t hmp);
1443
1444 void hammer_modify_volume(hammer_transaction_t trans, hammer_volume_t volume,
1445                         void *base, int len);
1446 void hammer_modify_buffer(hammer_transaction_t trans, hammer_buffer_t buffer,
1447                         void *base, int len);
1448 void hammer_modify_volume_done(hammer_volume_t volume);
1449 void hammer_modify_buffer_done(hammer_buffer_t buffer);
1450
1451 int hammer_ioc_reblock(hammer_transaction_t trans, hammer_inode_t ip,
1452                         struct hammer_ioc_reblock *reblock);
1453 int hammer_ioc_rebalance(hammer_transaction_t trans, hammer_inode_t ip,
1454                         struct hammer_ioc_rebalance *rebal);
1455 int hammer_ioc_prune(hammer_transaction_t trans, hammer_inode_t ip,
1456                         struct hammer_ioc_prune *prune);
1457 int hammer_ioc_mirror_read(hammer_transaction_t trans, hammer_inode_t ip,
1458                         struct hammer_ioc_mirror_rw *mirror);
1459 int hammer_ioc_mirror_write(hammer_transaction_t trans, hammer_inode_t ip,
1460                         struct hammer_ioc_mirror_rw *mirror);
1461 int hammer_ioc_set_pseudofs(hammer_transaction_t trans, hammer_inode_t ip,
1462                         struct ucred *cred, struct hammer_ioc_pseudofs_rw *pfs);
1463 int hammer_ioc_get_pseudofs(hammer_transaction_t trans, hammer_inode_t ip,
1464                         struct hammer_ioc_pseudofs_rw *pfs);
1465 int hammer_ioc_destroy_pseudofs(hammer_transaction_t trans, hammer_inode_t ip,
1466                         struct hammer_ioc_pseudofs_rw *pfs);
1467 int hammer_ioc_downgrade_pseudofs(hammer_transaction_t trans, hammer_inode_t ip,
1468                         struct hammer_ioc_pseudofs_rw *pfs);
1469 int hammer_ioc_upgrade_pseudofs(hammer_transaction_t trans, hammer_inode_t ip,
1470                         struct hammer_ioc_pseudofs_rw *pfs);
1471 int hammer_ioc_wait_pseudofs(hammer_transaction_t trans, hammer_inode_t ip,
1472                         struct hammer_ioc_pseudofs_rw *pfs);
1473 int hammer_ioc_iterate_pseudofs(hammer_transaction_t trans, hammer_inode_t ip,
1474                         struct hammer_ioc_pfs_iterate *pi);
1475 int hammer_ioc_volume_add(hammer_transaction_t trans, hammer_inode_t ip,
1476                         struct hammer_ioc_volume *ioc);
1477 int hammer_ioc_volume_del(hammer_transaction_t trans, hammer_inode_t ip,
1478                         struct hammer_ioc_volume *ioc);
1479 int hammer_ioc_volume_list(hammer_transaction_t trans, hammer_inode_t ip,
1480                         struct hammer_ioc_volume_list *ioc);
1481 int hammer_ioc_dedup(hammer_transaction_t trans, hammer_inode_t ip,
1482                         struct hammer_ioc_dedup *dedup);
1483
1484 int hammer_signal_check(hammer_mount_t hmp);
1485
1486 void hammer_flusher_create(hammer_mount_t hmp);
1487 void hammer_flusher_destroy(hammer_mount_t hmp);
1488 void hammer_flusher_sync(hammer_mount_t hmp);
1489 int  hammer_flusher_async(hammer_mount_t hmp, hammer_flush_group_t flg);
1490 int  hammer_flusher_async_one(hammer_mount_t hmp);
1491 int hammer_flusher_running(hammer_mount_t hmp);
1492 void hammer_flusher_wait(hammer_mount_t hmp, int seq);
1493 void hammer_flusher_wait_next(hammer_mount_t hmp);
1494 int  hammer_flusher_meta_limit(hammer_mount_t hmp);
1495 int  hammer_flusher_meta_halflimit(hammer_mount_t hmp);
1496 int  hammer_flusher_undo_exhausted(hammer_transaction_t trans, int quarter);
1497 void hammer_flusher_clean_loose_ios(hammer_mount_t hmp);
1498 void hammer_flusher_finalize(hammer_transaction_t trans, int final);
1499 int  hammer_flusher_haswork(hammer_mount_t hmp);
1500 int  hammer_flush_dirty(hammer_mount_t hmp, int max_count);
1501 void hammer_flusher_flush_undos(hammer_mount_t hmp, int already_flushed);
1502
1503 int hammer_recover_stage1(hammer_mount_t hmp, hammer_volume_t rootvol);
1504 int hammer_recover_stage2(hammer_mount_t hmp, hammer_volume_t rootvol);
1505 void hammer_recover_flush_buffers(hammer_mount_t hmp,
1506                         hammer_volume_t root_volume, int final);
1507
1508 void hammer_crc_set_blockmap(hammer_blockmap_t blockmap);
1509 void hammer_crc_set_volume(hammer_volume_ondisk_t ondisk);
1510 void hammer_crc_set_leaf(void *data, hammer_btree_leaf_elm_t leaf);
1511
1512 int hammer_crc_test_blockmap(hammer_blockmap_t blockmap);
1513 int hammer_crc_test_volume(hammer_volume_ondisk_t ondisk);
1514 int hammer_crc_test_btree(hammer_node_ondisk_t ondisk);
1515 int hammer_crc_test_leaf(void *data, hammer_btree_leaf_elm_t leaf);
1516 udev_t hammer_fsid_to_udev(uuid_t *uuid);
1517
1518
1519 int hammer_blocksize(int64_t file_offset);
1520 int hammer_blockoff(int64_t file_offset);
1521 int64_t hammer_blockdemarc(int64_t file_offset1, int64_t file_offset2);
1522
1523 /*
1524  * Shortcut for _hammer_checkspace(), used all over the code.
1525  */
1526 static __inline int
1527 hammer_checkspace(hammer_mount_t hmp, int slop)
1528 {
1529         return(_hammer_checkspace(hmp, slop, NULL));
1530 }
1531
1532 static __inline void
1533 hammer_wait_mem_record(hammer_record_t record)
1534 {
1535         hammer_wait_mem_record_ident(record, "hmmwai");
1536 }
1537
1538 static __inline void
1539 hammer_lock_ex(struct hammer_lock *lock)
1540 {
1541         hammer_lock_ex_ident(lock, "hmrlck");
1542 }
1543
1544 static __inline void
1545 hammer_modify_volume_noundo(hammer_transaction_t trans, hammer_volume_t volume)
1546 {
1547         hammer_modify_volume(trans, volume, NULL, 0);
1548 }
1549
1550 static __inline void
1551 hammer_modify_buffer_noundo(hammer_transaction_t trans, hammer_buffer_t buffer)
1552 {
1553         hammer_modify_buffer(trans, buffer, NULL, 0);
1554 }
1555
1556 /*
1557  * Indicate that a B-Tree node is being modified.
1558  */
1559 static __inline void
1560 hammer_modify_node_noundo(hammer_transaction_t trans, hammer_node_t node)
1561 {
1562         KKASSERT((node->flags & HAMMER_NODE_CRCBAD) == 0);
1563         hammer_modify_buffer(trans, node->buffer, NULL, 0);
1564 }
1565
1566 static __inline void
1567 hammer_modify_node_all(hammer_transaction_t trans, struct hammer_node *node)
1568 {
1569         KKASSERT((node->flags & HAMMER_NODE_CRCBAD) == 0);
1570         hammer_modify_buffer(trans, node->buffer,
1571                              node->ondisk, sizeof(*node->ondisk));
1572 }
1573
1574 static __inline void
1575 hammer_modify_node(hammer_transaction_t trans, hammer_node_t node,
1576                    void *base, int len)
1577 {
1578         hammer_crc_t *crcptr;
1579
1580         KKASSERT((char *)base >= (char *)node->ondisk &&
1581                  (char *)base + len <=
1582                     (char *)node->ondisk + sizeof(*node->ondisk));
1583         KKASSERT((node->flags & HAMMER_NODE_CRCBAD) == 0);
1584
1585         if (hammer_btree_full_undo) {
1586                 hammer_modify_node_all(trans, node);
1587         } else {
1588                 hammer_modify_buffer(trans, node->buffer, base, len);
1589                 crcptr = &node->ondisk->crc;
1590                 hammer_modify_buffer(trans, node->buffer,
1591                                      crcptr, sizeof(hammer_crc_t));
1592                 --node->buffer->io.modify_refs; /* only want one ref */
1593         }
1594 }
1595
1596 /*
1597  * Indicate that the specified modifications have been completed.
1598  *
1599  * Do not try to generate the crc here, it's very expensive to do and a
1600  * sequence of insertions or deletions can result in many calls to this
1601  * function on the same node.
1602  */
1603 static __inline void
1604 hammer_modify_node_done(hammer_node_t node)
1605 {
1606         node->flags |= HAMMER_NODE_CRCGOOD;
1607         if ((node->flags & HAMMER_NODE_NEEDSCRC) == 0) {
1608                 node->flags |= HAMMER_NODE_NEEDSCRC;
1609                 node->buffer->io.gencrc = 1;
1610                 hammer_ref_node(node);
1611         }
1612         hammer_modify_buffer_done(node->buffer);
1613 }
1614
1615 /*
1616  * Lookup a blockmap offset.
1617  */
1618 static __inline hammer_off_t
1619 hammer_blockmap_lookup(hammer_mount_t hmp, hammer_off_t zone_offset,
1620                         int *errorp)
1621 {
1622 #if defined INVARIANTS
1623         int zone = HAMMER_ZONE_DECODE(zone_offset);
1624         KKASSERT(hammer_is_zone2_mapped_index(zone));
1625 #endif
1626
1627         /*
1628          * We can actually skip blockmap verify by default,
1629          * as normal blockmaps are now direct-mapped onto the freemap
1630          * and so represent zone-2 addresses.
1631          */
1632         if (hammer_verify_zone == 0) {
1633                 *errorp = 0;
1634                 return hammer_xlate_to_zone2(zone_offset);
1635         }
1636
1637         return hammer_blockmap_lookup_verify(hmp, zone_offset, errorp);
1638 }
1639
1640 #define hammer_modify_volume_field(trans, vol, field)           \
1641         hammer_modify_volume(trans, vol, &(vol)->ondisk->field, \
1642                              sizeof((vol)->ondisk->field))
1643
1644 #define hammer_modify_node_field(trans, node, field)            \
1645         hammer_modify_node(trans, node, &(node)->ondisk->field, \
1646                              sizeof((node)->ondisk->field))
1647
1648 /*
1649  * The HAMMER_INODE_CAP_DIR_LOCAL_INO capability is set on newly
1650  * created directories for HAMMER version 2 or greater and causes
1651  * directory entries to be placed the inode localization zone in
1652  * the B-Tree instead of the misc zone.
1653  *
1654  * This greatly improves localization between directory entries and
1655  * inodes
1656  */
1657 static __inline uint32_t
1658 hammer_dir_localization(hammer_inode_t dip)
1659 {
1660         return(HAMMER_DIR_INODE_LOCALIZATION(&dip->ino_data));
1661 }
1662
1663 static __inline
1664 hammer_io_t
1665 hammer_buf_peek_io(struct buf *bp)
1666 {
1667         return((hammer_io_t)LIST_FIRST(&bp->b_dep));
1668 }
1669
1670 static __inline
1671 void
1672 hammer_buf_attach_io(struct buf *bp, hammer_io_t io)
1673 {
1674         /* struct buf and struct hammer_io are 1:1 */
1675         KKASSERT(hammer_buf_peek_io(bp) == NULL);
1676         LIST_INSERT_HEAD(&bp->b_dep, &io->worklist, node);
1677 }
1678
1679 #define hkprintf(format, args...)                       \
1680         kprintf("HAMMER: "format,## args)
1681 #define hvkprintf(vol, format, args...)                 \
1682         kprintf("HAMMER(%s) "format, vol->ondisk->vol_name,## args)
1683 #define hmkprintf(hmp, format, args...)                 \
1684         kprintf("HAMMER(%s) "format, hmp->mp->mnt_stat.f_mntfromname,## args)
1685 #define hdkprintf(format, args...)                      \
1686         kprintf("%s: "format, __func__,## args)
1687
1688 #define hkrateprintf(rate , format, args...)            \
1689         krateprintf(rate, "HAMMER: "format,## args)
1690 #define hvkrateprintf(rate, vol, format, args...)       \
1691         krateprintf(rate, "HAMMER(%s) "format, vol->ondisk->vol_name,## args)
1692 #define hmkrateprintf(rate, hmp, format, args...)       \
1693         krateprintf(rate, "HAMMER(%s) "format, hmp->mp->mnt_stat.f_mntfromname,## args)
1694 #define hdkrateprintf(rate, format, args...)            \
1695         krateprintf(rate, "%s: "format, __func__,## args)
1696
1697 #define hpanic(format, args...)                         \
1698         panic("%s: "format, __func__,## args)
1699 #endif  /* _KERNEL */
1700
1701 #endif /* !VFS_HAMMER_HAMMER_H_ */