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 */
 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 */
 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 struct worklist {
 630         LIST_ENTRY(worklist) node;
 631 };
 632
 633 TAILQ_HEAD(hammer_io_list, hammer_io);
 634 typedef struct hammer_io_list *hammer_io_list_t;
 635
 636 struct hammer_io {
 637         struct worklist         worklist;
 638         struct hammer_lock      lock;
 639         enum hammer_io_type     type;
 640         struct hammer_mount     *hmp;
 641         struct hammer_volume    *volume;
 642         RB_ENTRY(hammer_io)     rb_node;     /* if modified */
 643         TAILQ_ENTRY(hammer_io)  iorun_entry; /* iorun_list */
 644         struct hammer_mod_rb_tree *mod_root;
 645         struct buf              *bp;
 646         int64_t                 offset;    /* zone-2 offset */
 647         int                     bytes;     /* buffer cache buffer size */
 648         int                     modify_refs;
 649
 650         /*
 651          * These can be modified at any time by the backend while holding
 652          * io_token, due to bio_done and hammer_io_complete() callbacks.
 653          */
 654         u_int           running : 1;    /* bp write IO in progress */
 655         u_int           waiting : 1;    /* someone is waiting on us */
 656         u_int           ioerror : 1;    /* abort on io-error */
 657         u_int           unusedA : 29;
 658
 659         /*
 660          * These can only be modified by the frontend while holding
 661          * fs_token, or by the backend while holding the io interlocked
 662          * with no references (which will block the frontend when it
 663          * tries to reference it).
 664          *
 665          * WARNING! SMP RACES will create havoc if the callbacks ever tried
 666          *          to modify any of these outside the above restrictions.
 667          */
 668         u_int           modified : 1;   /* bp's data was modified */
 669         u_int           released : 1;   /* bp released (w/ B_LOCKED set) */
 670         u_int           waitdep : 1;    /* flush waits for dependancies */
 671         u_int           recovered : 1;  /* has recovery ref */
 672         u_int           waitmod : 1;    /* waiting for modify_refs */
 673         u_int           reclaim : 1;    /* reclaim requested */
 674         u_int           gencrc : 1;     /* crc needs to be generated */
 675         u_int           unusedB : 25;
 676 };
 677
 678 typedef struct hammer_io *hammer_io_t;
 679
 680 #define HAMMER_CLUSTER_SIZE     (64 * 1024)
 681 #if HAMMER_CLUSTER_SIZE > MAXBSIZE
 682 #undef  HAMMER_CLUSTER_SIZE
 683 #define HAMMER_CLUSTER_SIZE     MAXBSIZE
 684 #endif
 685
 686 /*
 687  * In-memory volume representing on-disk buffer
 688  */
 689 struct hammer_volume {
 690         struct hammer_io io;
 691         RB_ENTRY(hammer_volume) rb_node;
 692         struct hammer_volume_ondisk *ondisk;
 693         int32_t vol_no;
 694         hammer_off_t maxbuf_off; /* Maximum buffer offset (zone-2) */
 695         char    *vol_name;
 696         struct vnode *devvp;
 697         int     vol_flags;
 698 };
 699
 700 typedef struct hammer_volume *hammer_volume_t;
 701
 702 /*
 703  * In-memory buffer representing an on-disk buffer.
 704  */
 705 struct hammer_buffer {
 706         struct hammer_io io;
 707         RB_ENTRY(hammer_buffer) rb_node;
 708         void *ondisk;
 709         hammer_off_t zoneX_offset;
 710         hammer_off_t zone2_offset;
 711         struct hammer_reserve *resv;
 712         struct hammer_node_list clist;
 713 };
 714
 715 typedef struct hammer_buffer *hammer_buffer_t;
 716
 717 #define HAMMER_ITOB(iop) ((hammer_buffer_t)(iop))
 718
 719 /*
 720  * In-memory B-Tree node, representing an on-disk B-Tree node.
 721  *
 722  * This is a hang-on structure which is backed by a hammer_buffer,
 723  * and used for fine-grained locking of B-Tree nodes in order to
 724  * properly control lock ordering.
 725  */
 726 struct hammer_node {
 727         struct hammer_lock      lock;           /* node-by-node lock */
 728         TAILQ_ENTRY(hammer_node) entry;         /* per-buffer linkage */
 729         RB_ENTRY(hammer_node)   rb_node;        /* per-mount linkage */
 730         hammer_off_t            node_offset;    /* full offset spec */
 731         struct hammer_mount     *hmp;
 732         struct hammer_buffer    *buffer;        /* backing buffer */
 733         hammer_node_ondisk_t    ondisk;         /* ptr to on-disk structure */
 734         TAILQ_HEAD(, hammer_cursor) cursor_list;  /* deadlock recovery */
 735         struct hammer_node_cache_list cache_list; /* passive caches */
 736         int                     flags;
 737 #if 0
 738         int                     cursor_exclreq_count;
 739 #endif
 740 };
 741
 742 #define HAMMER_NODE_DELETED     0x0001
 743 #define HAMMER_NODE_FLUSH       0x0002
 744 #define HAMMER_NODE_CRCGOOD     0x0004
 745 #define HAMMER_NODE_NEEDSCRC    0x0008
 746 #define HAMMER_NODE_NEEDSMIRROR 0x0010
 747 #define HAMMER_NODE_CRCBAD      0x0020
 748 #define HAMMER_NODE_NONLINEAR   0x0040          /* linear heuristic */
 749
 750 #define HAMMER_NODE_CRCANY      (HAMMER_NODE_CRCGOOD | HAMMER_NODE_CRCBAD)
 751
 752 typedef struct hammer_node      *hammer_node_t;
 753
 754 /*
 755  * List of locked nodes.  This structure is used to lock potentially large
 756  * numbers of nodes as an aid for complex B-Tree operations.
 757  */
 758 struct hammer_node_lock;
 759 TAILQ_HEAD(hammer_node_lock_list, hammer_node_lock);
 760
 761 struct hammer_node_lock {
 762         TAILQ_ENTRY(hammer_node_lock) entry;
 763         struct hammer_node_lock_list  list;
 764         struct hammer_node_lock       *parent;
 765         hammer_node_t   node;
 766         hammer_node_ondisk_t copy;      /* copy of on-disk data */
 767         int             index;          /* index of this node in parent */
 768         int             count;          /* count children */
 769         int             flags;
 770 };
 771
 772 typedef struct hammer_node_lock *hammer_node_lock_t;
 773
 774 #define HAMMER_NODE_LOCK_UPDATED        0x0001
 775 #define HAMMER_NODE_LOCK_LCACHE         0x0002
 776
 777 /*
 778  * Common I/O management structure - embedded in in-memory structures
 779  * which are backed by filesystem buffers.
 780  */
 781 union hammer_io_structure {
 782         struct hammer_io        io;
 783         struct hammer_volume    volume;
 784         struct hammer_buffer    buffer;
 785 };
 786
 787 typedef union hammer_io_structure *hammer_io_structure_t;
 788
 789 /*
 790  * The reserve structure prevents the blockmap from allocating
 791  * out of a reserved big-block.  Such reservations are used by
 792  * the direct-write mechanism.
 793  *
 794  * The structure is also used to hold off on reallocations of
 795  * big-blocks from the freemap until flush dependancies have
 796  * been dealt with.
 797  */
 798 struct hammer_reserve {
 799         RB_ENTRY(hammer_reserve) rb_node;
 800         TAILQ_ENTRY(hammer_reserve) delay_entry;
 801         int             flush_group;
 802         int             flags;
 803         int             refs;
 804         int             zone;
 805         int             append_off;
 806         int32_t         bytes_free;
 807         hammer_off_t    zone_offset;
 808 };
 809
 810 typedef struct hammer_reserve *hammer_reserve_t;
 811
 812 #define HAMMER_RESF_ONDELAY     0x0001
 813 #define HAMMER_RESF_LAYER2FREE  0x0002
 814
 815 #include "hammer_cursor.h"
 816
 817 /*
 818  * The undo structure tracks recent undos to avoid laying down duplicate
 819  * undos within a flush group, saving us a significant amount of overhead.
 820  *
 821  * This is strictly a heuristic.
 822  */
 823 #define HAMMER_MAX_UNDOS                1024
 824 #define HAMMER_MAX_FLUSHERS             4
 825
 826 struct hammer_undo {
 827         RB_ENTRY(hammer_undo)   rb_node;
 828         TAILQ_ENTRY(hammer_undo) lru_entry;
 829         hammer_off_t            offset;
 830         int                     bytes;
 831 };
 832
 833 typedef struct hammer_undo *hammer_undo_t;
 834
 835 struct hammer_flusher_info;
 836 TAILQ_HEAD(hammer_flusher_info_list, hammer_flusher_info);
 837
 838 struct hammer_flusher {
 839         int             signal;         /* flusher thread sequencer */
 840         int             done;           /* last completed flush group */
 841         int             next;           /* next unallocated flg seqno */
 842         int             group_lock;     /* lock sequencing of the next flush */
 843         int             exiting;        /* request master exit */
 844         thread_t        td;             /* master flusher thread */
 845         hammer_tid_t    tid;            /* last flushed transaction id */
 846         int             finalize_want;          /* serialize finalization */
 847         struct hammer_lock finalize_lock;       /* serialize finalization */
 848         struct hammer_transaction trans;        /* shared transaction */
 849         struct hammer_flusher_info_list run_list;
 850         struct hammer_flusher_info_list ready_list;
 851 };
 852
 853 #define HAMMER_FLUSH_UNDOS_RELAXED      0
 854 #define HAMMER_FLUSH_UNDOS_FORCED       1
 855 #define HAMMER_FLUSH_UNDOS_AUTO         2
 856 /*
 857  * Internal hammer mount data structure
 858  */
 859 struct hammer_mount {
 860         struct mount *mp;
 861         struct hammer_ino_rb_tree rb_inos_root;
 862         struct hammer_redo_rb_tree rb_redo_root;
 863         struct hammer_vol_rb_tree rb_vols_root;
 864         struct hammer_nod_rb_tree rb_nods_root;
 865         struct hammer_und_rb_tree rb_undo_root;
 866         struct hammer_res_rb_tree rb_resv_root;
 867         struct hammer_buf_rb_tree rb_bufs_root;
 868         struct hammer_pfs_rb_tree rb_pfsm_root;
 869
 870         struct hammer_dedup_crc_rb_tree rb_dedup_crc_root;
 871         struct hammer_dedup_off_rb_tree rb_dedup_off_root;
 872
 873         struct hammer_volume *rootvol;
 874         struct hammer_base_elm root_btree_beg;
 875         struct hammer_base_elm root_btree_end;
 876
 877         struct malloc_type      *m_misc;
 878         struct malloc_type      *m_inodes;
 879
 880         int     flags;          /* HAMMER_MOUNT_xxx flags */
 881         int     hflags;
 882         int     ronly;
 883         int     nvolumes;
 884         int     master_id;      /* -1 or 0-15 for mirroring */
 885         int     version;        /* hammer filesystem version to use */
 886         int     rsv_inodes;     /* reserved space due to dirty inodes */
 887         int64_t rsv_databytes;  /* reserved space due to record data */
 888         int     rsv_recs;       /* reserved space due to dirty records */
 889         int     rsv_fromdelay;  /* big-blocks reserved due to flush delay */
 890         int     undo_rec_limit; /* based on size of undo area */
 891
 892         int     volume_to_remove; /* volume that is currently being removed */
 893
 894         int     count_inodes;   /* total number of inodes */
 895         int     count_iqueued;  /* inodes queued to flusher */
 896         int     count_reclaims; /* inodes pending reclaim by flusher */
 897
 898         struct hammer_flusher flusher;
 899
 900         u_int   check_interrupt;
 901         u_int   check_yield;
 902         uuid_t  fsid;
 903         struct hammer_mod_rb_tree volu_root;    /* dirty undo buffers */
 904         struct hammer_mod_rb_tree undo_root;    /* dirty undo buffers */
 905         struct hammer_mod_rb_tree data_root;    /* dirty data buffers */
 906         struct hammer_mod_rb_tree meta_root;    /* dirty meta bufs    */
 907         struct hammer_mod_rb_tree lose_root;    /* loose buffers      */
 908         long    locked_dirty_space;             /* meta/volu count    */
 909         long    io_running_space;               /* io_token */
 910         int     objid_cache_count;
 911         int     dedup_cache_count;
 912         int     error;                          /* critical I/O error */
 913         struct krate    krate;                  /* rate limited kprintf */
 914         struct krate    kdiag;                  /* rate limited kprintf */
 915         hammer_tid_t    asof;                   /* snapshot mount */
 916         hammer_tid_t    next_tid;
 917         hammer_tid_t    flush_tid1;             /* flusher tid sequencing */
 918         hammer_tid_t    flush_tid2;             /* flusher tid sequencing */
 919         int64_t copy_stat_freebigblocks;        /* number of free big-blocks */
 920         uint32_t        undo_seqno;             /* UNDO/REDO FIFO seqno */
 921         uint32_t        recover_stage2_seqno;   /* REDO recovery seqno */
 922         hammer_off_t    recover_stage2_offset;  /* REDO recovery offset */
 923
 924         struct netexport export;
 925         struct hammer_lock sync_lock;
 926         struct hammer_lock free_lock;
 927         struct hammer_lock undo_lock;
 928         struct hammer_lock blkmap_lock;
 929         struct hammer_lock snapshot_lock;
 930         struct hammer_lock volume_lock;
 931         struct hammer_blockmap  blockmap[HAMMER_MAX_ZONES];
 932         struct hammer_undo      undos[HAMMER_MAX_UNDOS];
 933         int                     undo_alloc;
 934         TAILQ_HEAD(, hammer_undo)  undo_lru_list;
 935         TAILQ_HEAD(, hammer_reserve) delay_list;
 936         struct hammer_flush_group_list  flush_group_list;
 937         hammer_flush_group_t    fill_flush_group;
 938         hammer_flush_group_t    next_flush_group;
 939         TAILQ_HEAD(, hammer_objid_cache) objid_cache_list;
 940         TAILQ_HEAD(, hammer_dedup_cache) dedup_lru_list;
 941         hammer_dedup_cache_t    dedup_free_cache;
 942         TAILQ_HEAD(, hammer_reclaim) reclaim_list;
 943         TAILQ_HEAD(, hammer_io) iorun_list;
 944
 945         struct lwkt_token       fs_token;       /* high level */
 946         struct lwkt_token       io_token;       /* low level (IO callback) */
 947
 948         struct hammer_inostats  inostats[HAMMER_INOSTATS_HSIZE];
 949         uint64_t volume_map[4];  /* 256 bits bitfield */
 950 };
 951
 952 typedef struct hammer_mount     *hammer_mount_t;
 953
 954 #define HAMMER_MOUNT_CRITICAL_ERROR     0x0001
 955 #define HAMMER_MOUNT_FLUSH_RECOVERY     0x0002
 956 #define HAMMER_MOUNT_REDO_SYNC          0x0004
 957 #define HAMMER_MOUNT_REDO_RECOVERY_REQ  0x0008
 958 #define HAMMER_MOUNT_REDO_RECOVERY_RUN  0x0010
 959
 960 #define HAMMER_VOLUME_NUMBER_ADD(hmp, vol)                      \
 961         (hmp)->volume_map[(vol)->vol_no >> 6] |=                \
 962         ((uint64_t)1 << ((vol)->vol_no & ((1 << 6) - 1)))
 963
 964 #define HAMMER_VOLUME_NUMBER_DEL(hmp, vol)                      \
 965         (hmp)->volume_map[(vol)->vol_no >> 6] &=                \
 966         ~((uint64_t)1 << ((vol)->vol_no & ((1 << 6) - 1)))
 967
 968 #define HAMMER_VOLUME_NUMBER_IS_SET(hmp, n)                     \
 969         (((hmp)->volume_map[(n) >> 6] &                         \
 970         ((uint64_t)1 << ((n) & ((1 << 6) - 1)))) != 0)
 971
 972 #define HAMMER_VOLUME_NUMBER_FOREACH(hmp, n)                    \
 973         for (n = 0; n < HAMMER_MAX_VOLUMES; n++)                \
 974                 if (HAMMER_VOLUME_NUMBER_IS_SET(hmp, n))
 975
 976 struct hammer_sync_info {
 977         int error;
 978         int waitfor;
 979 };
 980
 981 /*
 982  * Minium buffer cache bufs required to rebalance the B-Tree.
 983  * This is because we must hold the children and the children's children
 984  * locked.  Even this might not be enough if things are horribly out
 985  * of balance.
 986  */
 987 #define HAMMER_REBALANCE_MIN_BUFS       \
 988         (HAMMER_BTREE_LEAF_ELMS * HAMMER_BTREE_LEAF_ELMS)
 989
 990 #endif  /* _KERNEL || _KERNEL_STRUCTURES */
 991
 992 #if defined(_KERNEL)
 993 /*
 994  * checkspace slop (8MB chunks), higher numbers are more conservative.
 995  */
 996 #define HAMMER_CHKSPC_REBLOCK   25
 997 #define HAMMER_CHKSPC_MIRROR    20
 998 #define HAMMER_CHKSPC_WRITE     20
 999 #define HAMMER_CHKSPC_CREATE    20
1000 #define HAMMER_CHKSPC_REMOVE    10
1001 #define HAMMER_CHKSPC_EMERGENCY 0
1002
1003 extern struct vop_ops hammer_vnode_vops;
1004 extern struct vop_ops hammer_spec_vops;
1005 extern struct vop_ops hammer_fifo_vops;
1006 extern struct bio_ops hammer_bioops;
1007
1008 extern int hammer_debug_io;
1009 extern int hammer_debug_general;
1010 extern int hammer_debug_inode;
1011 extern int hammer_debug_locks;
1012 extern int hammer_debug_btree;
1013 extern int hammer_debug_tid;
1014 extern int hammer_debug_recover;
1015 extern int hammer_debug_recover_faults;
1016 extern int hammer_debug_critical;
1017 extern int hammer_cluster_enable;
1018 extern int hammer_live_dedup;
1019 extern int hammer_tdmux_ticks;
1020 extern int hammer_count_fsyncs;
1021 extern int hammer_count_inodes;
1022 extern int hammer_count_iqueued;
1023 extern int hammer_count_reclaims;
1024 extern int hammer_count_records;
1025 extern int hammer_count_record_datas;
1026 extern int hammer_count_volumes;
1027 extern int hammer_count_buffers;
1028 extern int hammer_count_nodes;
1029 extern int64_t hammer_count_extra_space_used;
1030 extern int64_t hammer_stats_btree_lookups;
1031 extern int64_t hammer_stats_btree_searches;
1032 extern int64_t hammer_stats_btree_inserts;
1033 extern int64_t hammer_stats_btree_deletes;
1034 extern int64_t hammer_stats_btree_elements;
1035 extern int64_t hammer_stats_btree_splits;
1036 extern int64_t hammer_stats_btree_iterations;
1037 extern int64_t hammer_stats_btree_root_iterations;
1038 extern int64_t hammer_stats_record_iterations;
1039 extern int64_t hammer_stats_file_read;
1040 extern int64_t hammer_stats_file_write;
1041 extern int64_t hammer_stats_file_iopsr;
1042 extern int64_t hammer_stats_file_iopsw;
1043 extern int64_t hammer_stats_disk_read;
1044 extern int64_t hammer_stats_disk_write;
1045 extern int64_t hammer_stats_inode_flushes;
1046 extern int64_t hammer_stats_commits;
1047 extern int64_t hammer_stats_undo;
1048 extern int64_t hammer_stats_redo;
1049 extern long hammer_count_dirtybufspace;
1050 extern int hammer_count_refedbufs;
1051 extern int hammer_count_reservations;
1052 extern long hammer_count_io_running_read;
1053 extern long hammer_count_io_running_write;
1054 extern int hammer_count_io_locked;
1055 extern long hammer_limit_dirtybufspace;
1056 extern int hammer_limit_recs;
1057 extern int hammer_limit_inode_recs;
1058 extern int hammer_limit_reclaims;
1059 extern int hammer_live_dedup_cache_size;
1060 extern int hammer_limit_redo;
1061 extern int hammer_bio_count;
1062 extern int hammer_verify_zone;
1063 extern int hammer_verify_data;
1064 extern int hammer_write_mode;
1065 extern int hammer_double_buffer;
1066 extern int hammer_btree_full_undo;
1067 extern int hammer_yield_check;
1068 extern int hammer_fsync_mode;
1069 extern int hammer_autoflush;
1070 extern int64_t hammer_contention_count;
1071
1072 extern int64_t hammer_live_dedup_vnode_bcmps;
1073 extern int64_t hammer_live_dedup_device_bcmps;
1074 extern int64_t hammer_live_dedup_findblk_failures;
1075 extern int64_t hammer_live_dedup_bmap_saves;
1076
1077 void    hammer_critical_error(hammer_mount_t hmp, hammer_inode_t ip,
1078                         int error, const char *msg);
1079 int     hammer_vop_inactive(struct vop_inactive_args *);
1080 int     hammer_vop_reclaim(struct vop_reclaim_args *);
1081 int     hammer_get_vnode(struct hammer_inode *ip, struct vnode **vpp);
1082 struct hammer_inode *hammer_get_inode(hammer_transaction_t trans,
1083                         hammer_inode_t dip, int64_t obj_id,
1084                         hammer_tid_t asof, uint32_t localization,
1085                         int flags, int *errorp);
1086 struct hammer_inode *hammer_get_dummy_inode(hammer_transaction_t trans,
1087                         hammer_inode_t dip, int64_t obj_id,
1088                         hammer_tid_t asof, uint32_t localization,
1089                         int flags, int *errorp);
1090 struct hammer_inode *hammer_find_inode(hammer_transaction_t trans,
1091                         int64_t obj_id, hammer_tid_t asof,
1092                         uint32_t localization);
1093 void    hammer_scan_inode_snapshots(hammer_mount_t hmp,
1094                         hammer_inode_info_t iinfo,
1095                         int (*callback)(hammer_inode_t ip, void *data),
1096                         void *data);
1097 void    hammer_put_inode(struct hammer_inode *ip);
1098 void    hammer_put_inode_ref(struct hammer_inode *ip);
1099 void    hammer_inode_waitreclaims(hammer_transaction_t trans);
1100 void    hammer_inode_dirty(struct hammer_inode *ip);
1101
1102 int     hammer_unload_volume(hammer_volume_t volume, void *data);
1103 int     hammer_adjust_volume_mode(hammer_volume_t volume, void *data __unused);
1104
1105 int     hammer_unload_buffer(hammer_buffer_t buffer, void *data);
1106 int     hammer_install_volume(hammer_mount_t hmp, const char *volname,
1107                         struct vnode *devvp, void *data);
1108 int     hammer_mountcheck_volumes(hammer_mount_t hmp);
1109 int     hammer_get_installed_volumes(hammer_mount_t hmp);
1110
1111 int     hammer_mem_add(hammer_record_t record);
1112 int     hammer_ip_lookup(hammer_cursor_t cursor);
1113 int     hammer_ip_first(hammer_cursor_t cursor);
1114 int     hammer_ip_next(hammer_cursor_t cursor);
1115 int     hammer_ip_resolve_data(hammer_cursor_t cursor);
1116 int     hammer_ip_delete_record(hammer_cursor_t cursor, hammer_inode_t ip,
1117                         hammer_tid_t tid);
1118 int     hammer_create_at_cursor(hammer_cursor_t cursor,
1119                         hammer_btree_leaf_elm_t leaf, void *udata, int mode);
1120 int     hammer_delete_at_cursor(hammer_cursor_t cursor, int delete_flags,
1121                         hammer_tid_t delete_tid, uint32_t delete_ts,
1122                         int track, int64_t *stat_bytes);
1123 int     hammer_ip_check_directory_empty(hammer_transaction_t trans,
1124                         hammer_inode_t ip);
1125 int     hammer_sync_hmp(hammer_mount_t hmp, int waitfor);
1126 int     hammer_queue_inodes_flusher(hammer_mount_t hmp, int waitfor);
1127
1128 hammer_record_t
1129         hammer_alloc_mem_record(hammer_inode_t ip, int data_len);
1130 void    hammer_flush_record_done(hammer_record_t record, int error);
1131 void    hammer_wait_mem_record_ident(hammer_record_t record, const char *ident);
1132 void    hammer_rel_mem_record(hammer_record_t record);
1133
1134 int     hammer_cursor_up(hammer_cursor_t cursor);
1135 int     hammer_cursor_up_locked(hammer_cursor_t cursor);
1136 int     hammer_cursor_down(hammer_cursor_t cursor);
1137 int     hammer_cursor_upgrade(hammer_cursor_t cursor);
1138 int     hammer_cursor_upgrade_node(hammer_cursor_t cursor);
1139 void    hammer_cursor_downgrade(hammer_cursor_t cursor);
1140 int     hammer_cursor_upgrade2(hammer_cursor_t c1, hammer_cursor_t c2);
1141 void    hammer_cursor_downgrade2(hammer_cursor_t c1, hammer_cursor_t c2);
1142 int     hammer_cursor_seek(hammer_cursor_t cursor, hammer_node_t node,
1143                         int index);
1144 void    hammer_lock_ex_ident(struct hammer_lock *lock, const char *ident);
1145 int     hammer_lock_ex_try(struct hammer_lock *lock);
1146 void    hammer_lock_sh(struct hammer_lock *lock);
1147 int     hammer_lock_sh_try(struct hammer_lock *lock);
1148 int     hammer_lock_upgrade(struct hammer_lock *lock, int shcount);
1149 void    hammer_lock_downgrade(struct hammer_lock *lock, int shcount);
1150 int     hammer_lock_status(struct hammer_lock *lock);
1151 void    hammer_unlock(struct hammer_lock *lock);
1152 void    hammer_ref(struct hammer_lock *lock);
1153 int     hammer_ref_interlock(struct hammer_lock *lock);
1154 int     hammer_ref_interlock_true(struct hammer_lock *lock);
1155 void    hammer_ref_interlock_done(struct hammer_lock *lock);
1156 void    hammer_rel(struct hammer_lock *lock);
1157 int     hammer_rel_interlock(struct hammer_lock *lock, int locked);
1158 void    hammer_rel_interlock_done(struct hammer_lock *lock, int orig_locked);
1159 int     hammer_get_interlock(struct hammer_lock *lock);
1160 int     hammer_try_interlock_norefs(struct hammer_lock *lock);
1161 void    hammer_put_interlock(struct hammer_lock *lock, int error);
1162
1163 void    hammer_sync_lock_ex(hammer_transaction_t trans);
1164 void    hammer_sync_lock_sh(hammer_transaction_t trans);
1165 int     hammer_sync_lock_sh_try(hammer_transaction_t trans);
1166 void    hammer_sync_unlock(hammer_transaction_t trans);
1167
1168 uint32_t hammer_to_unix_xid(uuid_t *uuid);
1169 void hammer_guid_to_uuid(uuid_t *uuid, uint32_t guid);
1170 void    hammer_time_to_timespec(uint64_t xtime, struct timespec *ts);
1171 uint64_t hammer_timespec_to_time(struct timespec *ts);
1172 int     hammer_str_to_tid(const char *str, int *ispfsp,
1173                         hammer_tid_t *tidp, uint32_t *localizationp);
1174 hammer_tid_t hammer_alloc_objid(hammer_mount_t hmp, hammer_inode_t dip,
1175                         int64_t namekey);
1176 void hammer_clear_objid(hammer_inode_t dip);
1177 void hammer_destroy_objid_cache(hammer_mount_t hmp);
1178
1179 int hammer_dedup_crc_rb_compare(hammer_dedup_cache_t dc1,
1180                         hammer_dedup_cache_t dc2);
1181 int hammer_dedup_off_rb_compare(hammer_dedup_cache_t dc1,
1182                         hammer_dedup_cache_t dc2);
1183 hammer_dedup_cache_t hammer_dedup_cache_add(hammer_inode_t ip,
1184                         hammer_btree_leaf_elm_t leaf);
1185 hammer_dedup_cache_t hammer_dedup_cache_lookup(hammer_mount_t hmp,
1186                         hammer_crc_t crc);
1187 void hammer_dedup_cache_inval(hammer_mount_t hmp, hammer_off_t base_offset);
1188 void hammer_destroy_dedup_cache(hammer_mount_t hmp);
1189 void hammer_dump_dedup_cache(hammer_mount_t hmp);
1190 int hammer_dedup_validate(hammer_dedup_cache_t dcp, int zone, int bytes,
1191                         void *data);
1192
1193 int hammer_enter_undo_history(hammer_mount_t hmp, hammer_off_t offset,
1194                         int bytes);
1195 void hammer_clear_undo_history(hammer_mount_t hmp);
1196 enum vtype hammer_get_vnode_type(uint8_t obj_type);
1197 int hammer_get_dtype(uint8_t obj_type);
1198 uint8_t hammer_get_obj_type(enum vtype vtype);
1199 int64_t hammer_directory_namekey(hammer_inode_t dip, const void *name, int len,
1200                         uint32_t *max_iterationsp);
1201 int     hammer_nohistory(hammer_inode_t ip);
1202
1203 int     hammer_init_cursor(hammer_transaction_t trans, hammer_cursor_t cursor,
1204                         hammer_node_cache_t cache, hammer_inode_t ip);
1205 void    hammer_normalize_cursor(hammer_cursor_t cursor);
1206 void    hammer_done_cursor(hammer_cursor_t cursor);
1207 int     hammer_recover_cursor(hammer_cursor_t cursor);
1208 void    hammer_unlock_cursor(hammer_cursor_t cursor);
1209 int     hammer_lock_cursor(hammer_cursor_t cursor);
1210 hammer_cursor_t hammer_push_cursor(hammer_cursor_t ocursor);
1211 void    hammer_pop_cursor(hammer_cursor_t ocursor, hammer_cursor_t ncursor);
1212
1213 void    hammer_cursor_replaced_node(hammer_node_t onode, hammer_node_t nnode);
1214 void    hammer_cursor_removed_node(hammer_node_t onode, hammer_node_t parent,
1215                         int index);
1216 void    hammer_cursor_split_node(hammer_node_t onode, hammer_node_t nnode,
1217                         int index);
1218 void    hammer_cursor_moved_element(hammer_node_t oparent, int pindex,
1219                         hammer_node_t onode, int oindex,
1220                         hammer_node_t nnode, int nindex);
1221 void    hammer_cursor_parent_changed(hammer_node_t node, hammer_node_t oparent,
1222                         hammer_node_t nparent, int nindex);
1223 void    hammer_cursor_inserted_element(hammer_node_t node, int index);
1224 void    hammer_cursor_deleted_element(hammer_node_t node, int index);
1225 void    hammer_cursor_invalidate_cache(hammer_cursor_t cursor);
1226
1227 int     hammer_btree_lookup(hammer_cursor_t cursor);
1228 int     hammer_btree_first(hammer_cursor_t cursor);
1229 int     hammer_btree_last(hammer_cursor_t cursor);
1230 int     hammer_btree_extract(hammer_cursor_t cursor, int flags);
1231 int     hammer_btree_iterate(hammer_cursor_t cursor);
1232 int     hammer_btree_iterate_reverse(hammer_cursor_t cursor);
1233 int     hammer_btree_insert(hammer_cursor_t cursor,
1234                             hammer_btree_leaf_elm_t elm, int *doprop);
1235 int     hammer_btree_delete(hammer_cursor_t cursor, int *ndelete);
1236 void    hammer_btree_do_propagation(hammer_cursor_t cursor,
1237                             hammer_pseudofs_inmem_t pfsm,
1238                             hammer_btree_leaf_elm_t leaf);
1239 int     hammer_btree_cmp(hammer_base_elm_t key1, hammer_base_elm_t key2);
1240 int     hammer_btree_chkts(hammer_tid_t ts, hammer_base_elm_t key);
1241 int     hammer_btree_correct_rhb(hammer_cursor_t cursor, hammer_tid_t tid);
1242 int     hammer_btree_correct_lhb(hammer_cursor_t cursor, hammer_tid_t tid);
1243
1244 int     btree_set_parent_of_child(hammer_transaction_t trans,
1245                         hammer_node_t node,
1246                         hammer_btree_elm_t elm);
1247 void    hammer_node_lock_init(hammer_node_lock_t parent, hammer_node_t node);
1248 void    hammer_btree_lcache_init(hammer_mount_t hmp, hammer_node_lock_t lcache,
1249                         int depth);
1250 void    hammer_btree_lcache_free(hammer_mount_t hmp, hammer_node_lock_t lcache);
1251 int     hammer_btree_lock_children(hammer_cursor_t cursor, int depth,
1252                         hammer_node_lock_t parent,
1253                         hammer_node_lock_t lcache);
1254 void    hammer_btree_lock_copy(hammer_cursor_t cursor,
1255                         hammer_node_lock_t parent);
1256 int     hammer_btree_sync_copy(hammer_cursor_t cursor,
1257                         hammer_node_lock_t parent);
1258 void    hammer_btree_unlock_children(hammer_mount_t hmp,
1259                         hammer_node_lock_t parent,
1260                         hammer_node_lock_t lcache);
1261 int     hammer_btree_search_node(hammer_base_elm_t elm, hammer_node_ondisk_t node);
1262 hammer_node_t hammer_btree_get_parent(hammer_transaction_t trans,
1263                         hammer_node_t node, int *parent_indexp,
1264                         int *errorp, int try_exclusive);
1265
1266 void    hammer_print_btree_node(hammer_node_ondisk_t ondisk);
1267 void    hammer_print_btree_elm(hammer_btree_elm_t elm);
1268
1269 void    *hammer_bread(hammer_mount_t hmp, hammer_off_t off,
1270                         int *errorp, struct hammer_buffer **bufferp);
1271 void    *hammer_bnew(hammer_mount_t hmp, hammer_off_t off,
1272                         int *errorp, struct hammer_buffer **bufferp);
1273 void    *hammer_bread_ext(hammer_mount_t hmp, hammer_off_t off, int bytes,
1274                         int *errorp, struct hammer_buffer **bufferp);
1275 void    *hammer_bnew_ext(hammer_mount_t hmp, hammer_off_t off, int bytes,
1276                         int *errorp, struct hammer_buffer **bufferp);
1277
1278 hammer_volume_t hammer_get_root_volume(hammer_mount_t hmp, int *errorp);
1279
1280 hammer_volume_t hammer_get_volume(hammer_mount_t hmp,
1281                         int32_t vol_no, int *errorp);
1282 hammer_buffer_t hammer_get_buffer(hammer_mount_t hmp, hammer_off_t buf_offset,
1283                         int bytes, int isnew, int *errorp);
1284 void            hammer_sync_buffers(hammer_mount_t hmp,
1285                         hammer_off_t base_offset, int bytes);
1286 int             hammer_del_buffers(hammer_mount_t hmp,
1287                         hammer_off_t base_offset,
1288                         hammer_off_t zone2_offset, int bytes,
1289                         int report_conflicts);
1290
1291 int             hammer_ref_volume(hammer_volume_t volume);
1292 int             hammer_ref_buffer(hammer_buffer_t buffer);
1293 void            hammer_flush_buffer_nodes(hammer_buffer_t buffer);
1294
1295 void            hammer_rel_volume(hammer_volume_t volume, int locked);
1296 void            hammer_rel_buffer(hammer_buffer_t buffer, int locked);
1297
1298 int             hammer_vfs_export(struct mount *mp, int op,
1299                         const struct export_args *export);
1300 hammer_node_t   hammer_get_node(hammer_transaction_t trans,
1301                         hammer_off_t node_offset, int isnew, int *errorp);
1302 void            hammer_ref_node(hammer_node_t node);
1303 hammer_node_t   hammer_ref_node_safe(hammer_transaction_t trans,
1304                         hammer_node_cache_t cache, int *errorp);
1305 void            hammer_rel_node(hammer_node_t node);
1306 void            hammer_delete_node(hammer_transaction_t trans,
1307                         hammer_node_t node);
1308 void            hammer_cache_node(hammer_node_cache_t cache,
1309                         hammer_node_t node);
1310 void            hammer_uncache_node(hammer_node_cache_t cache);
1311 void            hammer_flush_node(hammer_node_t node, int locked);
1312
1313 void hammer_dup_buffer(struct hammer_buffer **bufferp,
1314                         struct hammer_buffer *buffer);
1315 hammer_node_t hammer_alloc_btree(hammer_transaction_t trans,
1316                         hammer_off_t hint, int *errorp);
1317 void *hammer_alloc_data(hammer_transaction_t trans, int32_t data_len,
1318                         uint16_t rec_type, hammer_off_t *data_offsetp,
1319                         struct hammer_buffer **data_bufferp,
1320                         hammer_off_t hint, int *errorp);
1321
1322 int hammer_generate_undo(hammer_transaction_t trans,
1323                         hammer_off_t zone_offset, void *base, int len);
1324 int hammer_generate_redo(hammer_transaction_t trans, hammer_inode_t ip,
1325                         hammer_off_t file_offset, uint32_t flags,
1326                         void *base, int len);
1327 void hammer_generate_redo_sync(hammer_transaction_t trans);
1328 void hammer_redo_fifo_start_flush(hammer_inode_t ip);
1329 void hammer_redo_fifo_end_flush(hammer_inode_t ip);
1330
1331 void hammer_format_undo(void *base, uint32_t seqno);
1332 int hammer_upgrade_undo_4(hammer_transaction_t trans);
1333
1334 void hammer_put_volume(struct hammer_volume *volume, int flush);
1335 void hammer_put_buffer(struct hammer_buffer *buffer, int flush);
1336
1337 hammer_off_t hammer_freemap_alloc(hammer_transaction_t trans,
1338                         hammer_off_t owner, int *errorp);
1339 void hammer_freemap_free(hammer_transaction_t trans, hammer_off_t phys_offset,
1340                         hammer_off_t owner, int *errorp);
1341 int _hammer_checkspace(hammer_mount_t hmp, int slop, int64_t *resp);
1342 hammer_off_t hammer_blockmap_alloc(hammer_transaction_t trans, int zone,
1343                         int bytes, hammer_off_t hint, int *errorp);
1344 hammer_reserve_t hammer_blockmap_reserve(hammer_mount_t hmp, int zone,
1345                         int bytes, hammer_off_t *zone_offp, int *errorp);
1346 hammer_reserve_t hammer_blockmap_reserve_dedup(hammer_mount_t hmp, int zone,
1347                         int bytes, hammer_off_t zone_offset, int *errorp);
1348 void hammer_blockmap_reserve_complete(hammer_mount_t hmp,
1349                         hammer_reserve_t resv);
1350 void hammer_reserve_clrdelay(hammer_mount_t hmp, hammer_reserve_t resv);
1351 void hammer_blockmap_free(hammer_transaction_t trans,
1352                         hammer_off_t zone_offset, int bytes);
1353 int hammer_blockmap_dedup(hammer_transaction_t trans,
1354                         hammer_off_t zone_offset, int bytes);
1355 int hammer_blockmap_finalize(hammer_transaction_t trans,
1356                         hammer_reserve_t resv,
1357                         hammer_off_t zone_offset, int bytes);
1358 int hammer_blockmap_getfree(hammer_mount_t hmp, hammer_off_t zone_offset,
1359                         int *curp, int *errorp);
1360 hammer_off_t hammer_blockmap_lookup_verify(hammer_mount_t hmp,
1361                         hammer_off_t zone_offset, int *errorp);
1362
1363 hammer_off_t hammer_undo_lookup(hammer_mount_t hmp, hammer_off_t zone_offset,
1364                         int *errorp);
1365 int64_t hammer_undo_used(hammer_transaction_t trans);
1366 int64_t hammer_undo_space(hammer_transaction_t trans);
1367 int64_t hammer_undo_max(hammer_mount_t hmp);
1368 int hammer_undo_reclaim(hammer_io_t io);
1369
1370 void hammer_start_transaction(struct hammer_transaction *trans,
1371                               struct hammer_mount *hmp);
1372 void hammer_simple_transaction(struct hammer_transaction *trans,
1373                               struct hammer_mount *hmp);
1374 void hammer_start_transaction_fls(struct hammer_transaction *trans,
1375                                   struct hammer_mount *hmp);
1376 void hammer_done_transaction(struct hammer_transaction *trans);
1377 hammer_tid_t hammer_alloc_tid(hammer_mount_t hmp, int count);
1378
1379 void hammer_modify_inode(hammer_transaction_t trans, hammer_inode_t ip, int flags);
1380 void hammer_flush_inode(hammer_inode_t ip, int flags);
1381 void hammer_flush_inode_done(hammer_inode_t ip, int error);
1382 void hammer_wait_inode(hammer_inode_t ip);
1383
1384 int  hammer_create_inode(struct hammer_transaction *trans, struct vattr *vap,
1385                         struct ucred *cred, struct hammer_inode *dip,
1386                         const char *name, int namelen,
1387                         hammer_pseudofs_inmem_t pfsm,
1388                         struct hammer_inode **ipp);
1389 void hammer_rel_inode(hammer_inode_t ip, int flush);
1390 int hammer_reload_inode(hammer_inode_t ip, void *arg __unused);
1391 int hammer_ino_rb_compare(hammer_inode_t ip1, hammer_inode_t ip2);
1392 int hammer_redo_rb_compare(hammer_inode_t ip1, hammer_inode_t ip2);
1393 int hammer_destroy_inode_callback(hammer_inode_t ip, void *data __unused);
1394
1395 int hammer_sync_inode(hammer_transaction_t trans, hammer_inode_t ip);
1396 void hammer_test_inode(hammer_inode_t dip);
1397 void hammer_inode_unloadable_check(hammer_inode_t ip, int getvp);
1398 int hammer_update_atime_quick(hammer_inode_t ip);
1399
1400 int  hammer_ip_add_directory(struct hammer_transaction *trans,
1401                         hammer_inode_t dip, const char *name, int bytes,
1402                         hammer_inode_t nip);
1403 int  hammer_ip_del_directory(struct hammer_transaction *trans,
1404                         hammer_cursor_t cursor, hammer_inode_t dip,
1405                         hammer_inode_t ip);
1406 void hammer_ip_replace_bulk(hammer_mount_t hmp, hammer_record_t record);
1407 hammer_record_t hammer_ip_add_bulk(hammer_inode_t ip, off_t file_offset,
1408                         void *data, int bytes, int *errorp);
1409 int  hammer_ip_frontend_trunc(struct hammer_inode *ip, off_t file_size);
1410 int  hammer_ip_add_record(struct hammer_transaction *trans,
1411                         hammer_record_t record);
1412 int  hammer_ip_delete_range(hammer_cursor_t cursor, hammer_inode_t ip,
1413                         int64_t ran_beg, int64_t ran_end, int truncating);
1414 int  hammer_ip_delete_clean(hammer_cursor_t cursor, hammer_inode_t ip,
1415                         int *countp);
1416 int  hammer_ip_sync_data(hammer_cursor_t cursor, hammer_inode_t ip,
1417                         int64_t offset, void *data, int bytes);
1418 int  hammer_ip_sync_record_cursor(hammer_cursor_t cursor, hammer_record_t rec);
1419 hammer_pseudofs_inmem_t  hammer_load_pseudofs(hammer_transaction_t trans,
1420                         uint32_t localization, int *errorp);
1421 int  hammer_mkroot_pseudofs(hammer_transaction_t trans, struct ucred *cred,
1422                         hammer_pseudofs_inmem_t pfsm);
1423 int  hammer_save_pseudofs(hammer_transaction_t trans,
1424                         hammer_pseudofs_inmem_t pfsm);
1425 int  hammer_unload_pseudofs(hammer_transaction_t trans, uint32_t localization);
1426 void hammer_rel_pseudofs(hammer_mount_t hmp, hammer_pseudofs_inmem_t pfsm);
1427 int hammer_ioctl(hammer_inode_t ip, u_long com, caddr_t data, int fflag,
1428                         struct ucred *cred);
1429
1430 void hammer_io_init(hammer_io_t io, hammer_volume_t volume,
1431                         enum hammer_io_type type);
1432 int hammer_io_read(struct vnode *devvp, struct hammer_io *io, int limit);
1433 void hammer_io_advance(struct hammer_io *io);
1434 int hammer_io_new(struct vnode *devvp, struct hammer_io *io);
1435 int hammer_io_inval(hammer_volume_t volume, hammer_off_t zone2_offset);
1436 struct buf *hammer_io_release(struct hammer_io *io, int flush);
1437 void hammer_io_flush(struct hammer_io *io, int reclaim);
1438 void hammer_io_wait(struct hammer_io *io);
1439 void hammer_io_waitdep(struct hammer_io *io);
1440 void hammer_io_wait_all(hammer_mount_t hmp, const char *ident, int doflush);
1441 int hammer_io_direct_read(hammer_mount_t hmp, struct bio *bio,
1442                         hammer_btree_leaf_elm_t leaf);
1443 int hammer_io_indirect_read(hammer_mount_t hmp, struct bio *bio,
1444                         hammer_btree_leaf_elm_t leaf);
1445 int hammer_io_direct_write(hammer_mount_t hmp, struct bio *bio,
1446                         hammer_record_t record);
1447 void hammer_io_direct_wait(hammer_record_t record);
1448 void hammer_io_direct_uncache(hammer_mount_t hmp, hammer_btree_leaf_elm_t leaf);
1449 void hammer_io_write_interlock(hammer_io_t io);
1450 void hammer_io_done_interlock(hammer_io_t io);
1451 void hammer_io_clear_modify(struct hammer_io *io, int inval);
1452 void hammer_io_clear_modlist(struct hammer_io *io);
1453 void hammer_io_flush_sync(hammer_mount_t hmp);
1454 void hammer_io_clear_error(struct hammer_io *io);
1455 void hammer_io_clear_error_noassert(struct hammer_io *io);
1456 void hammer_io_notmeta(hammer_buffer_t buffer);
1457 void hammer_io_limit_backlog(hammer_mount_t hmp);
1458
1459 void hammer_modify_volume(hammer_transaction_t trans, hammer_volume_t volume,
1460                         void *base, int len);
1461 void hammer_modify_buffer(hammer_transaction_t trans, hammer_buffer_t buffer,
1462                         void *base, int len);
1463 void hammer_modify_volume_done(hammer_volume_t volume);
1464 void hammer_modify_buffer_done(hammer_buffer_t buffer);
1465
1466 int hammer_ioc_reblock(hammer_transaction_t trans, hammer_inode_t ip,
1467                         struct hammer_ioc_reblock *reblock);
1468 int hammer_ioc_rebalance(hammer_transaction_t trans, hammer_inode_t ip,
1469                         struct hammer_ioc_rebalance *rebal);
1470 int hammer_ioc_prune(hammer_transaction_t trans, hammer_inode_t ip,
1471                         struct hammer_ioc_prune *prune);
1472 int hammer_ioc_mirror_read(hammer_transaction_t trans, hammer_inode_t ip,
1473                         struct hammer_ioc_mirror_rw *mirror);
1474 int hammer_ioc_mirror_write(hammer_transaction_t trans, hammer_inode_t ip,
1475                         struct hammer_ioc_mirror_rw *mirror);
1476 int hammer_ioc_set_pseudofs(hammer_transaction_t trans, hammer_inode_t ip,
1477                         struct ucred *cred, struct hammer_ioc_pseudofs_rw *pfs);
1478 int hammer_ioc_get_pseudofs(hammer_transaction_t trans, hammer_inode_t ip,
1479                         struct hammer_ioc_pseudofs_rw *pfs);
1480 int hammer_ioc_destroy_pseudofs(hammer_transaction_t trans, hammer_inode_t ip,
1481                         struct hammer_ioc_pseudofs_rw *pfs);
1482 int hammer_ioc_downgrade_pseudofs(hammer_transaction_t trans, hammer_inode_t ip,
1483                         struct hammer_ioc_pseudofs_rw *pfs);
1484 int hammer_ioc_upgrade_pseudofs(hammer_transaction_t trans, hammer_inode_t ip,
1485                         struct hammer_ioc_pseudofs_rw *pfs);
1486 int hammer_ioc_wait_pseudofs(hammer_transaction_t trans, hammer_inode_t ip,
1487                         struct hammer_ioc_pseudofs_rw *pfs);
1488 int hammer_ioc_iterate_pseudofs(hammer_transaction_t trans, hammer_inode_t ip,
1489                         struct hammer_ioc_pfs_iterate *pi);
1490 int hammer_ioc_volume_add(hammer_transaction_t trans, hammer_inode_t ip,
1491                         struct hammer_ioc_volume *ioc);
1492 int hammer_ioc_volume_del(hammer_transaction_t trans, hammer_inode_t ip,
1493                         struct hammer_ioc_volume *ioc);
1494 int hammer_ioc_volume_list(hammer_transaction_t trans, hammer_inode_t ip,
1495                         struct hammer_ioc_volume_list *ioc);
1496 int hammer_ioc_dedup(hammer_transaction_t trans, hammer_inode_t ip,
1497                         struct hammer_ioc_dedup *dedup);
1498
1499 int hammer_signal_check(hammer_mount_t hmp);
1500
1501 void hammer_flusher_create(hammer_mount_t hmp);
1502 void hammer_flusher_destroy(hammer_mount_t hmp);
1503 void hammer_flusher_sync(hammer_mount_t hmp);
1504 int  hammer_flusher_async(hammer_mount_t hmp, hammer_flush_group_t flg);
1505 int  hammer_flusher_async_one(hammer_mount_t hmp);
1506 int hammer_flusher_running(hammer_mount_t hmp);
1507 void hammer_flusher_wait(hammer_mount_t hmp, int seq);
1508 void hammer_flusher_wait_next(hammer_mount_t hmp);
1509 int  hammer_flusher_meta_limit(hammer_mount_t hmp);
1510 int  hammer_flusher_meta_halflimit(hammer_mount_t hmp);
1511 int  hammer_flusher_undo_exhausted(hammer_transaction_t trans, int quarter);
1512 void hammer_flusher_clean_loose_ios(hammer_mount_t hmp);
1513 void hammer_flusher_finalize(hammer_transaction_t trans, int final);
1514 int  hammer_flusher_haswork(hammer_mount_t hmp);
1515 int  hammer_flush_dirty(hammer_mount_t hmp, int max_count);
1516 void hammer_flusher_flush_undos(hammer_mount_t hmp, int already_flushed);
1517
1518 int hammer_recover_stage1(hammer_mount_t hmp, hammer_volume_t rootvol);
1519 int hammer_recover_stage2(hammer_mount_t hmp, hammer_volume_t rootvol);
1520 void hammer_recover_flush_buffers(hammer_mount_t hmp,
1521                         hammer_volume_t root_volume, int final);
1522
1523 void hammer_crc_set_blockmap(hammer_blockmap_t blockmap);
1524 void hammer_crc_set_volume(hammer_volume_ondisk_t ondisk);
1525 void hammer_crc_set_leaf(void *data, hammer_btree_leaf_elm_t leaf);
1526
1527 int hammer_crc_test_blockmap(hammer_blockmap_t blockmap);
1528 int hammer_crc_test_volume(hammer_volume_ondisk_t ondisk);
1529 int hammer_crc_test_btree(hammer_node_ondisk_t ondisk);
1530 int hammer_crc_test_leaf(void *data, hammer_btree_leaf_elm_t leaf);
1531 udev_t hammer_fsid_to_udev(uuid_t *uuid);
1532
1533
1534 int hammer_blocksize(int64_t file_offset);
1535 int hammer_blockoff(int64_t file_offset);
1536 int64_t hammer_blockdemarc(int64_t file_offset1, int64_t file_offset2);
1537
1538 /*
1539  * Shortcut for _hammer_checkspace(), used all over the code.
1540  */
1541 static __inline int
1542 hammer_checkspace(hammer_mount_t hmp, int slop)
1543 {
1544         return(_hammer_checkspace(hmp, slop, NULL));
1545 }
1546
1547 static __inline void
1548 hammer_wait_mem_record(hammer_record_t record)
1549 {
1550         hammer_wait_mem_record_ident(record, "hmmwai");
1551 }
1552
1553 static __inline void
1554 hammer_lock_ex(struct hammer_lock *lock)
1555 {
1556         hammer_lock_ex_ident(lock, "hmrlck");
1557 }
1558
1559 static __inline void
1560 hammer_modify_volume_noundo(hammer_transaction_t trans, hammer_volume_t volume)
1561 {
1562         hammer_modify_volume(trans, volume, NULL, 0);
1563 }
1564
1565 static __inline void
1566 hammer_modify_buffer_noundo(hammer_transaction_t trans, hammer_buffer_t buffer)
1567 {
1568         hammer_modify_buffer(trans, buffer, NULL, 0);
1569 }
1570
1571 /*
1572  * Indicate that a B-Tree node is being modified.
1573  */
1574 static __inline void
1575 hammer_modify_node_noundo(hammer_transaction_t trans, hammer_node_t node)
1576 {
1577         KKASSERT((node->flags & HAMMER_NODE_CRCBAD) == 0);
1578         hammer_modify_buffer(trans, node->buffer, NULL, 0);
1579 }
1580
1581 static __inline void
1582 hammer_modify_node_all(hammer_transaction_t trans, struct hammer_node *node)
1583 {
1584         KKASSERT((node->flags & HAMMER_NODE_CRCBAD) == 0);
1585         hammer_modify_buffer(trans, node->buffer,
1586                              node->ondisk, sizeof(*node->ondisk));
1587 }
1588
1589 static __inline void
1590 hammer_modify_node(hammer_transaction_t trans, hammer_node_t node,
1591                    void *base, int len)
1592 {
1593         hammer_crc_t *crcptr;
1594
1595         KKASSERT((char *)base >= (char *)node->ondisk &&
1596                  (char *)base + len <=
1597                     (char *)node->ondisk + sizeof(*node->ondisk));
1598         KKASSERT((node->flags & HAMMER_NODE_CRCBAD) == 0);
1599
1600         if (hammer_btree_full_undo) {
1601                 hammer_modify_node_all(trans, node);
1602         } else {
1603                 hammer_modify_buffer(trans, node->buffer, base, len);
1604                 crcptr = &node->ondisk->crc;
1605                 hammer_modify_buffer(trans, node->buffer,
1606                                      crcptr, sizeof(hammer_crc_t));
1607                 --node->buffer->io.modify_refs; /* only want one ref */
1608         }
1609 }
1610
1611 /*
1612  * Indicate that the specified modifications have been completed.
1613  *
1614  * Do not try to generate the crc here, it's very expensive to do and a
1615  * sequence of insertions or deletions can result in many calls to this
1616  * function on the same node.
1617  */
1618 static __inline void
1619 hammer_modify_node_done(hammer_node_t node)
1620 {
1621         node->flags |= HAMMER_NODE_CRCGOOD;
1622         if ((node->flags & HAMMER_NODE_NEEDSCRC) == 0) {
1623                 node->flags |= HAMMER_NODE_NEEDSCRC;
1624                 node->buffer->io.gencrc = 1;
1625                 hammer_ref_node(node);
1626         }
1627         hammer_modify_buffer_done(node->buffer);
1628 }
1629
1630 /*
1631  * Lookup a blockmap offset.
1632  */
1633 static __inline hammer_off_t
1634 hammer_blockmap_lookup(hammer_mount_t hmp, hammer_off_t zone_offset,
1635                         int *errorp)
1636 {
1637 #if defined INVARIANTS
1638         int zone = HAMMER_ZONE_DECODE(zone_offset);
1639         KKASSERT(hammer_is_zone2_mapped_index(zone));
1640 #endif
1641
1642         /*
1643          * We can actually skip blockmap verify by default,
1644          * as normal blockmaps are now direct-mapped onto the freemap
1645          * and so represent zone-2 addresses.
1646          */
1647         if (hammer_verify_zone == 0) {
1648                 *errorp = 0;
1649                 return hammer_xlate_to_zone2(zone_offset);
1650         }
1651
1652         return hammer_blockmap_lookup_verify(hmp, zone_offset, errorp);
1653 }
1654
1655 #define hammer_modify_volume_field(trans, vol, field)           \
1656         hammer_modify_volume(trans, vol, &(vol)->ondisk->field, \
1657                              sizeof((vol)->ondisk->field))
1658
1659 #define hammer_modify_node_field(trans, node, field)            \
1660         hammer_modify_node(trans, node, &(node)->ondisk->field, \
1661                              sizeof((node)->ondisk->field))
1662
1663 /*
1664  * The HAMMER_INODE_CAP_DIR_LOCAL_INO capability is set on newly
1665  * created directories for HAMMER version 2 or greater and causes
1666  * directory entries to be placed the inode localization zone in
1667  * the B-Tree instead of the misc zone.
1668  *
1669  * This greatly improves localization between directory entries and
1670  * inodes
1671  */
1672 static __inline uint32_t
1673 hammer_dir_localization(hammer_inode_t dip)
1674 {
1675         return(HAMMER_DIR_INODE_LOCALIZATION(&dip->ino_data));
1676 }
1677
1678 #define hkprintf(format, args...)                       \
1679         kprintf("HAMMER: "format,## args)
1680 #define hvkprintf(vol, format, args...)                 \
1681         kprintf("HAMMER(%s) "format, vol->ondisk->vol_name,## args)
1682 #define hmkprintf(hmp, format, args...)                 \
1683         kprintf("HAMMER(%s) "format, hmp->mp->mnt_stat.f_mntfromname,## args)
1684 #define hdkprintf(format, args...)                      \
1685         kprintf("%s: "format, __func__,## args)
1686
1687 #define hkrateprintf(rate , format, args...)            \
1688         krateprintf(rate, "HAMMER: "format,## args)
1689 #define hvkrateprintf(rate, vol, format, args...)       \
1690         krateprintf(rate, "HAMMER(%s) "format, vol->ondisk->vol_name,## args)
1691 #define hmkrateprintf(rate, hmp, format, args...)       \
1692         krateprintf(rate, "HAMMER(%s) "format, hmp->mp->mnt_stat.f_mntfromname,## args)
1693 #define hdkrateprintf(rate, format, args...)            \
1694         krateprintf(rate, "%s: "format, __func__,## args)
1695
1696 #define hpanic(format, args...)                         \
1697         panic("%s: "format, __func__,## args)
1698 #endif  /* _KERNEL */
1699
1700 #endif /* !VFS_HAMMER_HAMMER_H_ */