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/mountctl.h>
  53 #include <sys/vnode.h>
  54 #include <sys/proc.h>
  55 #include <sys/priv.h>
  56 #include <sys/stat.h>
  57 #include <sys/globaldata.h>
  58 #include <sys/lockf.h>
  59 #include <sys/buf.h>
  60 #include <sys/queue.h>
  61 #include <sys/ktr.h>
  62 #include <sys/limits.h>
  63 #include <vm/vm_extern.h>
  64
  65 #ifdef _KERNEL
  66 #include <sys/buf2.h>
  67 #include <sys/signal2.h>
  68 #include <vm/vm_page2.h>
  69 #endif
  70
  71 #include "hammer_disk.h"
  72 #include "hammer_mount.h"
  73 #include "hammer_ioctl.h"
  74
  75 #if defined(_KERNEL) || defined(_KERNEL_STRUCTURES)
  76
  77 MALLOC_DECLARE(M_HAMMER);
  78
  79 /*
  80  * Kernel trace
  81  */
  82 #if !defined(KTR_HAMMER)
  83 #define KTR_HAMMER      KTR_ALL
  84 #endif
  85 /* KTR_INFO_MASTER_EXTERN(hammer); */
  86
  87 /*
  88  * Misc structures
  89  */
  90 struct hammer_mount;
  91
  92 /*
  93  * Key structure used for custom RB tree inode lookups.  This prototypes
  94  * the function hammer_ino_rb_tree_RB_LOOKUP_INFO(root, info).
  95  */
  96 typedef struct hammer_inode_info {
  97         int64_t         obj_id;         /* (key) object identifier */
  98         hammer_tid_t    obj_asof;       /* (key) snapshot transid or 0 */
  99         u_int32_t       obj_localization; /* (key) pseudo-fs */
 100         union {
 101                 struct hammer_btree_leaf_elm *leaf;
 102         } u;
 103 } *hammer_inode_info_t;
 104
 105 typedef enum hammer_transaction_type {
 106         HAMMER_TRANS_RO,
 107         HAMMER_TRANS_STD,
 108         HAMMER_TRANS_FLS
 109 } hammer_transaction_type_t;
 110
 111 /*
 112  * HAMMER Transaction tracking
 113  */
 114 struct hammer_transaction {
 115         hammer_transaction_type_t type;
 116         struct hammer_mount *hmp;
 117         hammer_tid_t    tid;
 118         u_int64_t       time;
 119         u_int32_t       time32;
 120         int             sync_lock_refs;
 121         int             flags;
 122         struct hammer_volume *rootvol;
 123 };
 124
 125 typedef struct hammer_transaction *hammer_transaction_t;
 126
 127 #define HAMMER_TRANSF_NEWINODE  0x0001
 128 #define HAMMER_TRANSF_DIDIO     0x0002
 129 #define HAMMER_TRANSF_CRCDOM    0x0004  /* EDOM on CRC error, less critical */
 130
 131 /*
 132  * HAMMER locks
 133  */
 134 struct hammer_lock {
 135         volatile u_int  refs;           /* active references */
 136         volatile u_int  lockval;        /* lock count and control bits */
 137         struct thread   *lowner;        /* owner if exclusively held */
 138         struct thread   *rowner;        /* owner if exclusively held */
 139 };
 140
 141 #define HAMMER_REFS_LOCKED      0x40000000      /* transition check */
 142 #define HAMMER_REFS_WANTED      0x20000000      /* transition check */
 143 #define HAMMER_REFS_CHECK       0x10000000      /* transition check */
 144
 145 #define HAMMER_REFS_FLAGS       (HAMMER_REFS_LOCKED | \
 146                                  HAMMER_REFS_WANTED | \
 147                                  HAMMER_REFS_CHECK)
 148
 149 #define HAMMER_LOCKF_EXCLUSIVE  0x40000000
 150 #define HAMMER_LOCKF_WANTED     0x20000000
 151
 152 static __inline int
 153 hammer_notlocked(struct hammer_lock *lock)
 154 {
 155         return(lock->lockval == 0);
 156 }
 157
 158 static __inline int
 159 hammer_islocked(struct hammer_lock *lock)
 160 {
 161         return(lock->lockval != 0);
 162 }
 163
 164 /*
 165  * Returns the number of refs on the object.
 166  */
 167 static __inline int
 168 hammer_isactive(struct hammer_lock *lock)
 169 {
 170         return(lock->refs & ~HAMMER_REFS_FLAGS);
 171 }
 172
 173 static __inline int
 174 hammer_oneref(struct hammer_lock *lock)
 175 {
 176         return((lock->refs & ~HAMMER_REFS_FLAGS) == 1);
 177 }
 178
 179 static __inline int
 180 hammer_norefs(struct hammer_lock *lock)
 181 {
 182         return((lock->refs & ~HAMMER_REFS_FLAGS) == 0);
 183 }
 184
 185 static __inline int
 186 hammer_norefsorlock(struct hammer_lock *lock)
 187 {
 188         return(lock->refs == 0);
 189 }
 190
 191 static __inline int
 192 hammer_refsorlock(struct hammer_lock *lock)
 193 {
 194         return(lock->refs != 0);
 195 }
 196
 197 /*
 198  * Return if we specifically own the lock exclusively.
 199  */
 200 static __inline int
 201 hammer_lock_excl_owned(struct hammer_lock *lock, thread_t td)
 202 {
 203         if ((lock->lockval & HAMMER_LOCKF_EXCLUSIVE) &&
 204             lock->lowner == td) {
 205                 return(1);
 206         }
 207         return(0);
 208 }
 209
 210 /*
 211  * Flush state, used by various structures
 212  */
 213 typedef enum hammer_inode_state {
 214         HAMMER_FST_IDLE,
 215         HAMMER_FST_SETUP,
 216         HAMMER_FST_FLUSH
 217 } hammer_inode_state_t;
 218
 219 TAILQ_HEAD(hammer_record_list, hammer_record);
 220
 221 /*
 222  * Pseudo-filesystem extended data tracking
 223  */
 224 struct hammer_pfs_rb_tree;
 225 struct hammer_pseudofs_inmem;
 226 RB_HEAD(hammer_pfs_rb_tree, hammer_pseudofs_inmem);
 227 RB_PROTOTYPE2(hammer_pfs_rb_tree, hammer_pseudofs_inmem, rb_node,
 228               hammer_pfs_rb_compare, u_int32_t);
 229
 230 struct hammer_pseudofs_inmem {
 231         RB_ENTRY(hammer_pseudofs_inmem) rb_node;
 232         struct hammer_lock      lock;
 233         u_int32_t               localization;
 234         hammer_tid_t            create_tid;
 235         int                     flags;
 236         udev_t                  fsid_udev;
 237         struct hammer_pseudofs_data pfsd;
 238 };
 239
 240 typedef struct hammer_pseudofs_inmem *hammer_pseudofs_inmem_t;
 241
 242 /*
 243  * Cache object ids.  A fixed number of objid cache structures are
 244  * created to reserve object id's for newly created files in multiples
 245  * of 100,000, localized to a particular directory, and recycled as
 246  * needed.  This allows parallel create operations in different
 247  * directories to retain fairly localized object ids which in turn
 248  * improves reblocking performance and layout.
 249  */
 250 #define OBJID_CACHE_SIZE        2048
 251 #define OBJID_CACHE_BULK_BITS   10              /* 10 bits (1024)       */
 252 #define OBJID_CACHE_BULK        (32 * 32)       /* two level (1024)     */
 253 #define OBJID_CACHE_BULK_MASK   (OBJID_CACHE_BULK - 1)
 254 #define OBJID_CACHE_BULK_MASK64 ((u_int64_t)(OBJID_CACHE_BULK - 1))
 255
 256 typedef struct hammer_objid_cache {
 257         TAILQ_ENTRY(hammer_objid_cache) entry;
 258         struct hammer_inode             *dip;
 259         hammer_tid_t                    base_tid;
 260         int                             count;
 261         u_int32_t                       bm0;
 262         u_int32_t                       bm1[32];
 263 } *hammer_objid_cache_t;
 264
 265 /*
 266  * Associate an inode with a B-Tree node to cache search start positions
 267  */
 268 typedef struct hammer_node_cache {
 269         TAILQ_ENTRY(hammer_node_cache)  entry;
 270         struct hammer_node              *node;
 271         struct hammer_inode             *ip;
 272 } *hammer_node_cache_t;
 273
 274 TAILQ_HEAD(hammer_node_cache_list, hammer_node_cache);
 275
 276 /*
 277  * Live dedup cache
 278  */
 279 struct hammer_dedup_crc_rb_tree;
 280 RB_HEAD(hammer_dedup_crc_rb_tree, hammer_dedup_cache);
 281 RB_PROTOTYPE2(hammer_dedup_crc_rb_tree, hammer_dedup_cache, crc_entry,
 282                 hammer_dedup_crc_rb_compare, hammer_crc_t);
 283
 284 struct hammer_dedup_off_rb_tree;
 285 RB_HEAD(hammer_dedup_off_rb_tree, hammer_dedup_cache);
 286 RB_PROTOTYPE2(hammer_dedup_off_rb_tree, hammer_dedup_cache, off_entry,
 287                 hammer_dedup_off_rb_compare, hammer_off_t);
 288
 289 #define DEDUP_CACHE_SIZE        4096 /* XXX make it a dynamic tunable */
 290
 291 typedef struct hammer_dedup_cache {
 292         RB_ENTRY(hammer_dedup_cache) crc_entry;
 293         RB_ENTRY(hammer_dedup_cache) off_entry;
 294         TAILQ_ENTRY(hammer_dedup_cache) lru_entry;
 295         struct hammer_mount *hmp;
 296         int64_t obj_id;
 297         u_int32_t localization;
 298         off_t file_offset;
 299         int bytes;
 300         hammer_off_t data_offset;
 301         hammer_crc_t crc;
 302 } *hammer_dedup_cache_t;
 303
 304 /*
 305  * Structure used to organize flush groups.  Flush groups must be
 306  * organized into chunks in order to avoid blowing out the UNDO FIFO.
 307  * Without this a 'sync' could end up flushing 50,000 inodes in a single
 308  * transaction.
 309  */
 310 struct hammer_fls_rb_tree;
 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 struct hammer_ino_rb_tree;
 346 struct hammer_inode;
 347 RB_HEAD(hammer_ino_rb_tree, hammer_inode);
 348 RB_PROTOTYPEX(hammer_ino_rb_tree, INFO, hammer_inode, rb_node,
 349               hammer_ino_rb_compare, hammer_inode_info_t);
 350
 351 struct hammer_redo_rb_tree;
 352 RB_HEAD(hammer_redo_rb_tree, hammer_inode);
 353 RB_PROTOTYPE2(hammer_redo_rb_tree, hammer_inode, rb_redonode,
 354               hammer_redo_rb_compare, hammer_off_t);
 355
 356 struct hammer_rec_rb_tree;
 357 struct hammer_record;
 358 RB_HEAD(hammer_rec_rb_tree, hammer_record);
 359 RB_PROTOTYPEX(hammer_rec_rb_tree, INFO, hammer_record, rb_node,
 360               hammer_rec_rb_compare, hammer_btree_leaf_elm_t);
 361
 362 TAILQ_HEAD(hammer_node_list, hammer_node);
 363
 364 struct hammer_inode {
 365         RB_ENTRY(hammer_inode)  rb_node;
 366         hammer_inode_state_t    flush_state;
 367         hammer_flush_group_t    flush_group;
 368         RB_ENTRY(hammer_inode)  rb_flsnode;     /* when on flush list */
 369         RB_ENTRY(hammer_inode)  rb_redonode;    /* when INODE_RDIRTY is set */
 370         struct hammer_record_list target_list;  /* target of dependant recs */
 371         int64_t                 obj_id;         /* (key) object identifier */
 372         hammer_tid_t            obj_asof;       /* (key) snapshot or 0 */
 373         u_int32_t               obj_localization; /* (key) pseudo-fs */
 374         struct hammer_mount     *hmp;
 375         hammer_objid_cache_t    objid_cache;
 376         int                     flags;
 377         int                     error;          /* flush error */
 378         int                     cursor_ip_refs; /* sanity */
 379 #if 0
 380         int                     cursor_exclreq_count;
 381 #endif
 382         int                     rsv_recs;
 383         struct vnode            *vp;
 384         hammer_pseudofs_inmem_t pfsm;
 385         struct lockf            advlock;
 386         struct hammer_lock      lock;           /* sync copy interlock */
 387         off_t                   trunc_off;
 388         struct hammer_btree_leaf_elm ino_leaf;  /* in-memory cache */
 389         struct hammer_inode_data ino_data;      /* in-memory cache */
 390         struct hammer_rec_rb_tree rec_tree;     /* in-memory cache */
 391         int                     rec_generation;
 392
 393         /*
 394          * search initiate cache
 395          * cache[0] - this inode
 396          * cache[1] - related data, the content depends on situations
 397          * cache[2] - for dip to cache ip to shortcut B-Tree search
 398          * cache[3] - related data copied from dip to a new ip's cache[1]
 399          */
 400         struct hammer_node_cache cache[4];
 401
 402         /*
 403          * When a demark is created to synchronize an inode to
 404          * disk, certain fields are copied so the front-end VOPs
 405          * can continue to run in parallel with the synchronization
 406          * occuring in the background.
 407          */
 408         int             sync_flags;             /* to-sync flags cache */
 409         off_t           sync_trunc_off;         /* to-sync truncation */
 410         off_t           save_trunc_off;         /* write optimization */
 411         struct hammer_btree_leaf_elm sync_ino_leaf; /* to-sync cache */
 412         struct hammer_inode_data sync_ino_data; /* to-sync cache */
 413         size_t          redo_count;
 414
 415         /*
 416          * Track the earliest offset in the UNDO/REDO FIFO containing
 417          * REDO records.  This is staged to the backend during flush
 418          * sequences.  While the inode is staged redo_fifo_next is used
 419          * to track the earliest offset for rotation into redo_fifo_start
 420          * on completion of the flush.
 421          */
 422         hammer_off_t    redo_fifo_start;
 423         hammer_off_t    redo_fifo_next;
 424 };
 425
 426 typedef struct hammer_inode *hammer_inode_t;
 427
 428 #define VTOI(vp)        ((struct hammer_inode *)(vp)->v_data)
 429
 430 /*
 431  * NOTE: DDIRTY does not include atime or mtime and does not include
 432  *       write-append size changes.  SDIRTY handles write-append size
 433  *       changes.
 434  *
 435  *       REDO indicates that REDO logging is active, creating a definitive
 436  *       stream of REDO records in the UNDO/REDO log for writes and
 437  *       truncations, including boundary records when/if REDO is turned off.
 438  *       REDO is typically enabled by fsync() and turned off if excessive
 439  *       writes without an fsync() occurs.
 440  *
 441  *       RDIRTY indicates that REDO records were laid down in the UNDO/REDO
 442  *       FIFO (even if REDO is turned off some might still be active) and
 443  *       still being tracked for this inode.  See hammer_redo.c
 444  */
 445                                         /* (not including atime/mtime) */
 446 #define HAMMER_INODE_DDIRTY     0x0001  /* in-memory ino_data is dirty */
 447 #define HAMMER_INODE_RSV_INODES 0x0002  /* hmp->rsv_inodes bumped */
 448 #define HAMMER_INODE_CONN_DOWN  0x0004  /* include in downward recursion */
 449 #define HAMMER_INODE_XDIRTY     0x0008  /* in-memory records */
 450 #define HAMMER_INODE_ONDISK     0x0010  /* inode is on-disk (else not yet) */
 451 #define HAMMER_INODE_FLUSH      0x0020  /* flush on last ref */
 452 #define HAMMER_INODE_DELETED    0x0080  /* inode delete (backend) */
 453 #define HAMMER_INODE_DELONDISK  0x0100  /* delete synchronized to disk */
 454 #define HAMMER_INODE_RO         0x0200  /* read-only (because of as-of) */
 455 #define HAMMER_INODE_RECSW      0x0400  /* waiting on data record flush */
 456 #define HAMMER_INODE_DONDISK    0x0800  /* data records may be on disk */
 457 #define HAMMER_INODE_BUFS       0x1000  /* dirty high level bps present */
 458 #define HAMMER_INODE_REFLUSH    0x2000  /* flush on dependancy / reflush */
 459 #define HAMMER_INODE_RECLAIM    0x4000  /* trying to reclaim */
 460 #define HAMMER_INODE_FLUSHW     0x8000  /* Someone waiting for flush */
 461
 462 #define HAMMER_INODE_TRUNCATED  0x00010000
 463 #define HAMMER_INODE_DELETING   0x00020000 /* inode delete request (frontend)*/
 464 #define HAMMER_INODE_RESIGNAL   0x00040000 /* re-signal on re-flush */
 465 #define HAMMER_INODE_ATIME      0x00100000 /* in-memory atime modified */
 466 #define HAMMER_INODE_MTIME      0x00200000 /* in-memory mtime modified */
 467 #define HAMMER_INODE_WOULDBLOCK 0x00400000 /* re-issue to new flush group */
 468 #define HAMMER_INODE_DUMMY      0x00800000 /* dummy inode covering bad file */
 469 #define HAMMER_INODE_SDIRTY     0x01000000 /* in-memory ino_data.size is dirty*/
 470 #define HAMMER_INODE_REDO       0x02000000 /* REDO logging active */
 471 #define HAMMER_INODE_RDIRTY     0x04000000 /* REDO records active in fifo */
 472 #define HAMMER_INODE_SLAVEFLUSH 0x08000000 /* being flushed by slave */
 473
 474 #define HAMMER_INODE_MODMASK    (HAMMER_INODE_DDIRTY|HAMMER_INODE_SDIRTY|   \
 475                                  HAMMER_INODE_XDIRTY|HAMMER_INODE_BUFS|     \
 476                                  HAMMER_INODE_ATIME|HAMMER_INODE_MTIME|     \
 477                                  HAMMER_INODE_TRUNCATED|HAMMER_INODE_DELETING)
 478
 479 #define HAMMER_INODE_MODMASK_NOXDIRTY   \
 480                                 (HAMMER_INODE_MODMASK & ~HAMMER_INODE_XDIRTY)
 481
 482 #define HAMMER_INODE_MODMASK_NOREDO     \
 483                                 (HAMMER_INODE_DDIRTY|                       \
 484                                  HAMMER_INODE_XDIRTY|                       \
 485                                  HAMMER_INODE_TRUNCATED|HAMMER_INODE_DELETING)
 486
 487 #define HAMMER_FLUSH_SIGNAL     0x0001
 488 #define HAMMER_FLUSH_RECURSION  0x0002
 489
 490 /*
 491  * Used by the inode reclaim code to pipeline reclaims and avoid
 492  * blowing out kernel memory or letting the flusher get too far
 493  * behind.  The reclaim wakes up when count reaches 0 or the
 494  * timer expires.
 495  */
 496 struct hammer_reclaim {
 497         TAILQ_ENTRY(hammer_reclaim) entry;
 498         int     count;
 499 };
 500
 501 /*
 502  * Track who is creating the greatest burden on the
 503  * inode cache.
 504  */
 505 struct hammer_inostats {
 506         pid_t           pid;    /* track user process */
 507         int             ltick;  /* last tick */
 508         int             count;  /* count (degenerates) */
 509 };
 510
 511 #define HAMMER_INOSTATS_HSIZE   32
 512 #define HAMMER_INOSTATS_HMASK   (HAMMER_INOSTATS_HSIZE - 1)
 513
 514 /*
 515  * Structure used to represent an unsynchronized record in-memory.  These
 516  * records typically represent directory entries.  Only non-historical
 517  * records are kept in-memory.
 518  *
 519  * Records are organized as a per-inode RB-Tree.  If the inode is not
 520  * on disk then neither are any records and the in-memory record tree
 521  * represents the entire contents of the inode.  If the inode is on disk
 522  * then the on-disk B-Tree is scanned in parallel with the in-memory
 523  * RB-Tree to synthesize the current state of the file.
 524  *
 525  * Records are also used to enforce the ordering of directory create/delete
 526  * operations.  A new inode will not be flushed to disk unless its related
 527  * directory entry is also being flushed at the same time.  A directory entry
 528  * will not be removed unless its related inode is also being removed at the
 529  * same time.
 530  */
 531 typedef enum hammer_record_type {
 532         HAMMER_MEM_RECORD_GENERAL,      /* misc record */
 533         HAMMER_MEM_RECORD_INODE,        /* inode record */
 534         HAMMER_MEM_RECORD_ADD,          /* positive memory cache record */
 535         HAMMER_MEM_RECORD_DEL,          /* negative delete-on-disk record */
 536         HAMMER_MEM_RECORD_DATA          /* bulk-data record w/on-disk ref */
 537 } hammer_record_type_t;
 538
 539 struct hammer_record {
 540         RB_ENTRY(hammer_record)         rb_node;
 541         TAILQ_ENTRY(hammer_record)      target_entry;
 542         hammer_inode_state_t            flush_state;
 543         hammer_flush_group_t            flush_group;
 544         hammer_record_type_t            type;
 545         struct hammer_lock              lock;
 546         struct hammer_reserve           *resv;
 547         struct hammer_inode             *ip;
 548         struct hammer_inode             *target_ip;
 549         struct hammer_btree_leaf_elm    leaf;
 550         union hammer_data_ondisk        *data;
 551         int                             flags;
 552         int                             gflags;
 553         hammer_off_t                    zone2_offset;   /* direct-write only */
 554 };
 555
 556 typedef struct hammer_record *hammer_record_t;
 557
 558 /*
 559  * Record flags.  Note that FE can only be set by the frontend if the
 560  * record has not been interlocked by the backend w/ BE.
 561  */
 562 #define HAMMER_RECF_ALLOCDATA           0x0001
 563 #define HAMMER_RECF_ONRBTREE            0x0002
 564 #define HAMMER_RECF_DELETED_FE          0x0004  /* deleted (frontend) */
 565 #define HAMMER_RECF_DELETED_BE          0x0008  /* deleted (backend) */
 566 #define HAMMER_RECF_COMMITTED           0x0010  /* committed to the B-Tree */
 567 #define HAMMER_RECF_INTERLOCK_BE        0x0020  /* backend interlock */
 568 #define HAMMER_RECF_WANTED              0x0040  /* wanted by the frontend */
 569 #define HAMMER_RECF_DEDUPED             0x0080  /* will be live-dedup'ed */
 570 #define HAMMER_RECF_CONVERT_DELETE      0x0100  /* special case */
 571 #define HAMMER_RECF_REDO                0x1000  /* REDO was laid down */
 572
 573 /*
 574  * These flags must be separate to deal with SMP races
 575  */
 576 #define HAMMER_RECG_DIRECT_IO           0x0001  /* related direct I/O running*/
 577 #define HAMMER_RECG_DIRECT_WAIT         0x0002  /* related direct I/O running*/
 578 #define HAMMER_RECG_DIRECT_INVAL        0x0004  /* buffer alias invalidation */
 579 /*
 580  * hammer_create_at_cursor() and hammer_delete_at_cursor() flags.
 581  */
 582 #define HAMMER_CREATE_MODE_UMIRROR      0x0001
 583 #define HAMMER_CREATE_MODE_SYS          0x0002
 584
 585 #define HAMMER_DELETE_ADJUST            0x0001
 586 #define HAMMER_DELETE_DESTROY           0x0002
 587
 588 /*
 589  * In-memory structures representing on-disk structures.
 590  */
 591 struct hammer_volume;
 592 struct hammer_buffer;
 593 struct hammer_node;
 594 struct hammer_undo;
 595 struct hammer_reserve;
 596
 597 RB_HEAD(hammer_vol_rb_tree, hammer_volume);
 598 RB_HEAD(hammer_buf_rb_tree, hammer_buffer);
 599 RB_HEAD(hammer_nod_rb_tree, hammer_node);
 600 RB_HEAD(hammer_und_rb_tree, hammer_undo);
 601 RB_HEAD(hammer_res_rb_tree, hammer_reserve);
 602 RB_HEAD(hammer_mod_rb_tree, hammer_io);
 603
 604 RB_PROTOTYPE2(hammer_vol_rb_tree, hammer_volume, rb_node,
 605               hammer_vol_rb_compare, int32_t);
 606 RB_PROTOTYPE2(hammer_buf_rb_tree, hammer_buffer, rb_node,
 607               hammer_buf_rb_compare, hammer_off_t);
 608 RB_PROTOTYPE2(hammer_nod_rb_tree, hammer_node, rb_node,
 609               hammer_nod_rb_compare, hammer_off_t);
 610 RB_PROTOTYPE2(hammer_und_rb_tree, hammer_undo, rb_node,
 611               hammer_und_rb_compare, hammer_off_t);
 612 RB_PROTOTYPE2(hammer_res_rb_tree, hammer_reserve, rb_node,
 613               hammer_res_rb_compare, hammer_off_t);
 614 RB_PROTOTYPE2(hammer_mod_rb_tree, hammer_io, rb_node,
 615               hammer_mod_rb_compare, hammer_off_t);
 616
 617 /*
 618  * IO management - embedded at the head of various in-memory structures
 619  *
 620  * VOLUME       - hammer_volume containing meta-data
 621  * META_BUFFER  - hammer_buffer containing meta-data
 622  * UNDO_BUFFER  - hammer_buffer containing undo-data
 623  * DATA_BUFFER  - hammer_buffer containing pure-data
 624  * DUMMY        - hammer_buffer not containing valid data
 625  *
 626  * Dirty volume headers and dirty meta-data buffers are locked until the
 627  * flusher can sequence them out.  Dirty pure-data buffers can be written.
 628  * Clean buffers can be passively released.
 629  */
 630 typedef enum hammer_io_type {
 631         HAMMER_STRUCTURE_VOLUME,
 632         HAMMER_STRUCTURE_META_BUFFER,
 633         HAMMER_STRUCTURE_UNDO_BUFFER,
 634         HAMMER_STRUCTURE_DATA_BUFFER,
 635         HAMMER_STRUCTURE_DUMMY
 636 } hammer_io_type_t;
 637
 638 union hammer_io_structure;
 639 struct hammer_io;
 640
 641 struct worklist {
 642         LIST_ENTRY(worklist) node;
 643 };
 644
 645 TAILQ_HEAD(hammer_io_list, hammer_io);
 646 typedef struct hammer_io_list *hammer_io_list_t;
 647
 648 struct hammer_io {
 649         struct worklist         worklist;
 650         struct hammer_lock      lock;
 651         enum hammer_io_type     type;
 652         struct hammer_mount     *hmp;
 653         struct hammer_volume    *volume;
 654         RB_ENTRY(hammer_io)     rb_node;     /* if modified */
 655         TAILQ_ENTRY(hammer_io)  iorun_entry; /* iorun_list */
 656         struct hammer_mod_rb_tree *mod_root;
 657         struct buf              *bp;
 658         int64_t                 offset;    /* zone-2 offset */
 659         int                     bytes;     /* buffer cache buffer size */
 660         int                     modify_refs;
 661
 662         /*
 663          * These can be modified at any time by the backend while holding
 664          * io_token, due to bio_done and hammer_io_complete() callbacks.
 665          */
 666         u_int           running : 1;    /* bp write IO in progress */
 667         u_int           waiting : 1;    /* someone is waiting on us */
 668         u_int           ioerror : 1;    /* abort on io-error */
 669         u_int           unusedA : 29;
 670
 671         /*
 672          * These can only be modified by the frontend while holding
 673          * fs_token, or by the backend while holding the io interlocked
 674          * with no references (which will block the frontend when it
 675          * tries to reference it).
 676          *
 677          * WARNING! SMP RACES will create havoc if the callbacks ever tried
 678          *          to modify any of these outside the above restrictions.
 679          */
 680         u_int           modified : 1;   /* bp's data was modified */
 681         u_int           released : 1;   /* bp released (w/ B_LOCKED set) */
 682         u_int           validated : 1;  /* ondisk has been validated */
 683         u_int           waitdep : 1;    /* flush waits for dependancies */
 684         u_int           recovered : 1;  /* has recovery ref */
 685         u_int           waitmod : 1;    /* waiting for modify_refs */
 686         u_int           reclaim : 1;    /* reclaim requested */
 687         u_int           gencrc : 1;     /* crc needs to be generated */
 688         u_int           unusedB : 24;
 689 };
 690
 691 typedef struct hammer_io *hammer_io_t;
 692
 693 #define HAMMER_CLUSTER_SIZE     (64 * 1024)
 694 #if HAMMER_CLUSTER_SIZE > MAXBSIZE
 695 #undef  HAMMER_CLUSTER_SIZE
 696 #define HAMMER_CLUSTER_SIZE     MAXBSIZE
 697 #endif
 698
 699 /*
 700  * In-memory volume representing on-disk buffer
 701  */
 702 struct hammer_volume {
 703         struct hammer_io io;
 704         RB_ENTRY(hammer_volume) rb_node;
 705         struct hammer_volume_ondisk *ondisk;
 706         int32_t vol_no;
 707         int64_t nblocks;        /* note: special calculation for statfs */
 708         int64_t buffer_base;    /* base offset of buffer 0 */
 709         hammer_off_t maxbuf_off; /* Maximum buffer offset (zone-2) */
 710         hammer_off_t maxraw_off; /* Maximum raw offset for device */
 711         char    *vol_name;
 712         struct vnode *devvp;
 713         int     vol_flags;
 714 };
 715
 716 typedef struct hammer_volume *hammer_volume_t;
 717
 718 /*
 719  * In-memory buffer representing an on-disk buffer.
 720  */
 721 struct hammer_buffer {
 722         struct hammer_io io;
 723         RB_ENTRY(hammer_buffer) rb_node;
 724         void *ondisk;
 725         hammer_off_t zoneX_offset;
 726         hammer_off_t zone2_offset;
 727         struct hammer_reserve *resv;
 728         struct hammer_node_list clist;
 729 };
 730
 731 typedef struct hammer_buffer *hammer_buffer_t;
 732
 733 /*
 734  * In-memory B-Tree node, representing an on-disk B-Tree node.
 735  *
 736  * This is a hang-on structure which is backed by a hammer_buffer,
 737  * and used for fine-grained locking of B-Tree nodes in order to
 738  * properly control lock ordering.
 739  */
 740 struct hammer_node {
 741         struct hammer_lock      lock;           /* node-by-node lock */
 742         TAILQ_ENTRY(hammer_node) entry;         /* per-buffer linkage */
 743         RB_ENTRY(hammer_node)   rb_node;        /* per-mount linkage */
 744         hammer_off_t            node_offset;    /* full offset spec */
 745         struct hammer_mount     *hmp;
 746         struct hammer_buffer    *buffer;        /* backing buffer */
 747         hammer_node_ondisk_t    ondisk;         /* ptr to on-disk structure */
 748         TAILQ_HEAD(, hammer_cursor) cursor_list;  /* deadlock recovery */
 749         struct hammer_node_cache_list cache_list; /* passive caches */
 750         int                     flags;
 751 #if 0
 752         int                     cursor_exclreq_count;
 753 #endif
 754 };
 755
 756 #define HAMMER_NODE_DELETED     0x0001
 757 #define HAMMER_NODE_FLUSH       0x0002
 758 #define HAMMER_NODE_CRCGOOD     0x0004
 759 #define HAMMER_NODE_NEEDSCRC    0x0008
 760 #define HAMMER_NODE_NEEDSMIRROR 0x0010
 761 #define HAMMER_NODE_CRCBAD      0x0020
 762 #define HAMMER_NODE_NONLINEAR   0x0040          /* linear heuristic */
 763
 764 #define HAMMER_NODE_CRCANY      (HAMMER_NODE_CRCGOOD | HAMMER_NODE_CRCBAD)
 765
 766 typedef struct hammer_node      *hammer_node_t;
 767
 768 /*
 769  * List of locked nodes.  This structure is used to lock potentially large
 770  * numbers of nodes as an aid for complex B-Tree operations.
 771  */
 772 struct hammer_node_lock;
 773 TAILQ_HEAD(hammer_node_lock_list, hammer_node_lock);
 774
 775 struct hammer_node_lock {
 776         TAILQ_ENTRY(hammer_node_lock) entry;
 777         struct hammer_node_lock_list  list;
 778         struct hammer_node_lock       *parent;
 779         hammer_node_t   node;
 780         hammer_node_ondisk_t copy;      /* copy of on-disk data */
 781         int             index;          /* index of this node in parent */
 782         int             count;          /* count children */
 783         int             flags;
 784 };
 785
 786 typedef struct hammer_node_lock *hammer_node_lock_t;
 787
 788 #define HAMMER_NODE_LOCK_UPDATED        0x0001
 789 #define HAMMER_NODE_LOCK_LCACHE         0x0002
 790
 791 /*
 792  * Common I/O management structure - embedded in in-memory structures
 793  * which are backed by filesystem buffers.
 794  */
 795 union hammer_io_structure {
 796         struct hammer_io        io;
 797         struct hammer_volume    volume;
 798         struct hammer_buffer    buffer;
 799 };
 800
 801 typedef union hammer_io_structure *hammer_io_structure_t;
 802
 803 /*
 804  * The reserve structure prevents the blockmap from allocating
 805  * out of a reserved big-block.  Such reservations are used by
 806  * the direct-write mechanism.
 807  *
 808  * The structure is also used to hold off on reallocations of
 809  * big-blocks from the freemap until flush dependancies have
 810  * been dealt with.
 811  */
 812 struct hammer_reserve {
 813         RB_ENTRY(hammer_reserve) rb_node;
 814         TAILQ_ENTRY(hammer_reserve) delay_entry;
 815         int             flush_group;
 816         int             flags;
 817         int             refs;
 818         int             zone;
 819         int             append_off;
 820         int32_t         bytes_free;
 821         hammer_off_t    zone_offset;
 822 };
 823
 824 typedef struct hammer_reserve *hammer_reserve_t;
 825
 826 #define HAMMER_RESF_ONDELAY     0x0001
 827 #define HAMMER_RESF_LAYER2FREE  0x0002
 828
 829 #include "hammer_cursor.h"
 830
 831 /*
 832  * The undo structure tracks recent undos to avoid laying down duplicate
 833  * undos within a flush group, saving us a significant amount of overhead.
 834  *
 835  * This is strictly a heuristic.
 836  */
 837 #define HAMMER_MAX_UNDOS                1024
 838 #define HAMMER_MAX_FLUSHERS             4
 839
 840 struct hammer_undo {
 841         RB_ENTRY(hammer_undo)   rb_node;
 842         TAILQ_ENTRY(hammer_undo) lru_entry;
 843         hammer_off_t            offset;
 844         int                     bytes;
 845 };
 846
 847 typedef struct hammer_undo *hammer_undo_t;
 848
 849 struct hammer_flusher_info;
 850 TAILQ_HEAD(hammer_flusher_info_list, hammer_flusher_info);
 851
 852 struct hammer_flusher {
 853         int             signal;         /* flusher thread sequencer */
 854         int             done;           /* last completed flush group */
 855         int             next;           /* next unallocated flg seqno */
 856         int             group_lock;     /* lock sequencing of the next flush */
 857         int             exiting;        /* request master exit */
 858         thread_t        td;             /* master flusher thread */
 859         hammer_tid_t    tid;            /* last flushed transaction id */
 860         int             finalize_want;          /* serialize finalization */
 861         struct hammer_lock finalize_lock;       /* serialize finalization */
 862         struct hammer_transaction trans;        /* shared transaction */
 863         struct hammer_flusher_info_list run_list;
 864         struct hammer_flusher_info_list ready_list;
 865 };
 866
 867 #define HAMMER_FLUSH_UNDOS_RELAXED      0
 868 #define HAMMER_FLUSH_UNDOS_FORCED       1
 869 #define HAMMER_FLUSH_UNDOS_AUTO         2
 870 /*
 871  * Internal hammer mount data structure
 872  */
 873 struct hammer_mount {
 874         struct mount *mp;
 875         struct hammer_ino_rb_tree rb_inos_root;
 876         struct hammer_redo_rb_tree rb_redo_root;
 877         struct hammer_vol_rb_tree rb_vols_root;
 878         struct hammer_nod_rb_tree rb_nods_root;
 879         struct hammer_und_rb_tree rb_undo_root;
 880         struct hammer_res_rb_tree rb_resv_root;
 881         struct hammer_buf_rb_tree rb_bufs_root;
 882         struct hammer_pfs_rb_tree rb_pfsm_root;
 883
 884         struct hammer_dedup_crc_rb_tree rb_dedup_crc_root;
 885         struct hammer_dedup_off_rb_tree rb_dedup_off_root;
 886
 887         struct hammer_volume *rootvol;
 888         struct hammer_base_elm root_btree_beg;
 889         struct hammer_base_elm root_btree_end;
 890
 891         struct malloc_type      *m_misc;
 892         struct malloc_type      *m_inodes;
 893
 894         int     flags;          /* HAMMER_MOUNT_xxx flags */
 895         int     hflags;
 896         int     ronly;
 897         int     nvolumes;
 898         int     volume_iterator;
 899         int     master_id;      /* -1 or 0-15 for mirroring */
 900         int     version;        /* hammer filesystem version to use */
 901         int     rsv_inodes;     /* reserved space due to dirty inodes */
 902         int64_t rsv_databytes;  /* reserved space due to record data */
 903         int     rsv_recs;       /* reserved space due to dirty records */
 904         int     rsv_fromdelay;  /* big-blocks reserved due to flush delay */
 905         int     undo_rec_limit; /* based on size of undo area */
 906         int     last_newrecords;
 907         int     count_newrecords;
 908
 909         int     volume_to_remove; /* volume that is currently being removed */
 910
 911         int     count_inodes;   /* total number of inodes */
 912         int     count_iqueued;  /* inodes queued to flusher */
 913         int     count_reclaims; /* inodes pending reclaim by flusher */
 914
 915         struct hammer_flusher flusher;
 916
 917         u_int   check_interrupt;
 918         u_int   check_yield;
 919         uuid_t  fsid;
 920         struct hammer_mod_rb_tree volu_root;    /* dirty undo buffers */
 921         struct hammer_mod_rb_tree undo_root;    /* dirty undo buffers */
 922         struct hammer_mod_rb_tree data_root;    /* dirty data buffers */
 923         struct hammer_mod_rb_tree meta_root;    /* dirty meta bufs    */
 924         struct hammer_mod_rb_tree lose_root;    /* loose buffers      */
 925         long    locked_dirty_space;             /* meta/volu count    */
 926         long    io_running_space;               /* io_token */
 927         int     unused01;
 928         int     objid_cache_count;
 929         int     dedup_cache_count;
 930         int     error;                          /* critical I/O error */
 931         struct krate    krate;                  /* rate limited kprintf */
 932         struct krate    kdiag;                  /* rate limited kprintf */
 933         hammer_tid_t    asof;                   /* snapshot mount */
 934         hammer_tid_t    next_tid;
 935         hammer_tid_t    flush_tid1;             /* flusher tid sequencing */
 936         hammer_tid_t    flush_tid2;             /* flusher tid sequencing */
 937         int64_t copy_stat_freebigblocks;        /* number of free big-blocks */
 938         u_int32_t       undo_seqno;             /* UNDO/REDO FIFO seqno */
 939         u_int32_t       recover_stage2_seqno;   /* REDO recovery seqno */
 940         hammer_off_t    recover_stage2_offset;  /* REDO recovery offset */
 941
 942         struct netexport export;
 943         struct hammer_lock sync_lock;
 944         struct hammer_lock free_lock;
 945         struct hammer_lock undo_lock;
 946         struct hammer_lock blkmap_lock;
 947         struct hammer_lock snapshot_lock;
 948         struct hammer_lock volume_lock;
 949         struct hammer_blockmap  blockmap[HAMMER_MAX_ZONES];
 950         struct hammer_undo      undos[HAMMER_MAX_UNDOS];
 951         int                     undo_alloc;
 952         TAILQ_HEAD(, hammer_undo)  undo_lru_list;
 953         TAILQ_HEAD(, hammer_reserve) delay_list;
 954         struct hammer_flush_group_list  flush_group_list;
 955         hammer_flush_group_t    fill_flush_group;
 956         hammer_flush_group_t    next_flush_group;
 957         TAILQ_HEAD(, hammer_objid_cache) objid_cache_list;
 958         TAILQ_HEAD(, hammer_dedup_cache) dedup_lru_list;
 959         hammer_dedup_cache_t    dedup_free_cache;
 960         TAILQ_HEAD(, hammer_reclaim) reclaim_list;
 961         TAILQ_HEAD(, hammer_io) iorun_list;
 962
 963         struct lwkt_token       fs_token;       /* high level */
 964         struct lwkt_token       io_token;       /* low level (IO callback) */
 965
 966         struct hammer_inostats  inostats[HAMMER_INOSTATS_HSIZE];
 967         uint64_t volume_map[4];  /* 256 bits bitfield */
 968 };
 969
 970 typedef struct hammer_mount     *hammer_mount_t;
 971
 972 #define HAMMER_MOUNT_CRITICAL_ERROR     0x0001
 973 #define HAMMER_MOUNT_FLUSH_RECOVERY     0x0002
 974 #define HAMMER_MOUNT_REDO_SYNC          0x0004
 975 #define HAMMER_MOUNT_REDO_RECOVERY_REQ  0x0008
 976 #define HAMMER_MOUNT_REDO_RECOVERY_RUN  0x0010
 977
 978 #define HAMMER_VOLUME_NUMBER_ADD(hmp, vol)                      \
 979         (hmp)->volume_map[(vol)->vol_no >> 6] |=                \
 980         ((uint64_t)1 << ((vol)->vol_no & ((1 << 6) - 1)))
 981
 982 #define HAMMER_VOLUME_NUMBER_DEL(hmp, vol)                      \
 983         (hmp)->volume_map[(vol)->vol_no >> 6] &=                \
 984         ~((uint64_t)1 << ((vol)->vol_no & ((1 << 6) - 1)))
 985
 986 #define HAMMER_VOLUME_NUMBER_IS_SET(hmp, n)                     \
 987         (((hmp)->volume_map[(n) >> 6] &                         \
 988         ((uint64_t)1 << ((n) & ((1 << 6) - 1)))) != 0)
 989
 990 #define HAMMER_VOLUME_NUMBER_FOREACH(hmp, n)                    \
 991         for (n = 0; n < HAMMER_MAX_VOLUMES; n++)                \
 992                 if (HAMMER_VOLUME_NUMBER_IS_SET(hmp, n))
 993
 994 struct hammer_sync_info {
 995         int error;
 996         int waitfor;
 997 };
 998
 999 /*
1000  * Minium buffer cache bufs required to rebalance the B-Tree.
1001  * This is because we must hold the children and the children's children
1002  * locked.  Even this might not be enough if things are horribly out
1003  * of balance.
1004  */
1005 #define HAMMER_REBALANCE_MIN_BUFS       \
1006         (HAMMER_BTREE_LEAF_ELMS * HAMMER_BTREE_LEAF_ELMS)
1007
1008 #endif  /* _KERNEL || _KERNEL_STRUCTURES */
1009
1010 #if defined(_KERNEL)
1011 /*
1012  * checkspace slop (8MB chunks), higher numbers are more conservative.
1013  */
1014 #define HAMMER_CHKSPC_REBLOCK   25
1015 #define HAMMER_CHKSPC_MIRROR    20
1016 #define HAMMER_CHKSPC_WRITE     20
1017 #define HAMMER_CHKSPC_CREATE    20
1018 #define HAMMER_CHKSPC_REMOVE    10
1019 #define HAMMER_CHKSPC_EMERGENCY 0
1020
1021 extern struct vop_ops hammer_vnode_vops;
1022 extern struct vop_ops hammer_spec_vops;
1023 extern struct vop_ops hammer_fifo_vops;
1024 extern struct bio_ops hammer_bioops;
1025
1026 extern int hammer_debug_io;
1027 extern int hammer_debug_general;
1028 extern int hammer_debug_debug;
1029 extern int hammer_debug_inode;
1030 extern int hammer_debug_locks;
1031 extern int hammer_debug_btree;
1032 extern int hammer_debug_tid;
1033 extern int hammer_debug_recover;
1034 extern int hammer_debug_recover_faults;
1035 extern int hammer_debug_critical;
1036 extern int hammer_cluster_enable;
1037 extern int hammer_live_dedup;
1038 extern int hammer_tdmux_ticks;
1039 extern int hammer_count_fsyncs;
1040 extern int hammer_count_inodes;
1041 extern int hammer_count_iqueued;
1042 extern int hammer_count_reclaims;
1043 extern int hammer_count_records;
1044 extern int hammer_count_record_datas;
1045 extern int hammer_count_volumes;
1046 extern int hammer_count_buffers;
1047 extern int hammer_count_nodes;
1048 extern int64_t hammer_count_extra_space_used;
1049 extern int64_t hammer_stats_btree_lookups;
1050 extern int64_t hammer_stats_btree_searches;
1051 extern int64_t hammer_stats_btree_inserts;
1052 extern int64_t hammer_stats_btree_deletes;
1053 extern int64_t hammer_stats_btree_elements;
1054 extern int64_t hammer_stats_btree_splits;
1055 extern int64_t hammer_stats_btree_iterations;
1056 extern int64_t hammer_stats_btree_root_iterations;
1057 extern int64_t hammer_stats_record_iterations;
1058 extern int64_t hammer_stats_file_read;
1059 extern int64_t hammer_stats_file_write;
1060 extern int64_t hammer_stats_file_iopsr;
1061 extern int64_t hammer_stats_file_iopsw;
1062 extern int64_t hammer_stats_disk_read;
1063 extern int64_t hammer_stats_disk_write;
1064 extern int64_t hammer_stats_inode_flushes;
1065 extern int64_t hammer_stats_commits;
1066 extern int64_t hammer_stats_undo;
1067 extern int64_t hammer_stats_redo;
1068 extern long hammer_count_dirtybufspace;
1069 extern int hammer_count_refedbufs;
1070 extern int hammer_count_reservations;
1071 extern long hammer_count_io_running_read;
1072 extern long hammer_count_io_running_write;
1073 extern int hammer_count_io_locked;
1074 extern long hammer_limit_dirtybufspace;
1075 extern int hammer_limit_recs;
1076 extern int hammer_limit_inode_recs;
1077 extern int hammer_limit_reclaims;
1078 extern int hammer_live_dedup_cache_size;
1079 extern int hammer_limit_redo;
1080 extern int hammer_bio_count;
1081 extern int hammer_verify_zone;
1082 extern int hammer_verify_data;
1083 extern int hammer_write_mode;
1084 extern int hammer_double_buffer;
1085 extern int hammer_btree_full_undo;
1086 extern int hammer_yield_check;
1087 extern int hammer_fsync_mode;
1088 extern int hammer_autoflush;
1089 extern int64_t hammer_contention_count;
1090
1091 extern int64_t hammer_live_dedup_vnode_bcmps;
1092 extern int64_t hammer_live_dedup_device_bcmps;
1093 extern int64_t hammer_live_dedup_findblk_failures;
1094 extern int64_t hammer_live_dedup_bmap_saves;
1095
1096 void    hammer_critical_error(hammer_mount_t hmp, hammer_inode_t ip,
1097                         int error, const char *msg);
1098 int     hammer_vop_inactive(struct vop_inactive_args *);
1099 int     hammer_vop_reclaim(struct vop_reclaim_args *);
1100 int     hammer_get_vnode(struct hammer_inode *ip, struct vnode **vpp);
1101 struct hammer_inode *hammer_get_inode(hammer_transaction_t trans,
1102                         hammer_inode_t dip, int64_t obj_id,
1103                         hammer_tid_t asof, u_int32_t localization,
1104                         int flags, int *errorp);
1105 struct hammer_inode *hammer_get_dummy_inode(hammer_transaction_t trans,
1106                         hammer_inode_t dip, int64_t obj_id,
1107                         hammer_tid_t asof, u_int32_t localization,
1108                         int flags, int *errorp);
1109 struct hammer_inode *hammer_find_inode(hammer_transaction_t trans,
1110                         int64_t obj_id, hammer_tid_t asof,
1111                         u_int32_t localization);
1112 void    hammer_scan_inode_snapshots(hammer_mount_t hmp,
1113                         hammer_inode_info_t iinfo,
1114                         int (*callback)(hammer_inode_t ip, void *data),
1115                         void *data);
1116 void    hammer_put_inode(struct hammer_inode *ip);
1117 void    hammer_put_inode_ref(struct hammer_inode *ip);
1118 void    hammer_inode_waitreclaims(hammer_transaction_t trans);
1119 void    hammer_inode_dirty(struct hammer_inode *ip);
1120
1121 int     hammer_unload_volume(hammer_volume_t volume, void *data);
1122 int     hammer_adjust_volume_mode(hammer_volume_t volume, void *data __unused);
1123
1124 int     hammer_unload_buffer(hammer_buffer_t buffer, void *data);
1125 int     hammer_install_volume(hammer_mount_t hmp, const char *volname,
1126                         struct vnode *devvp, void *data);
1127 int     hammer_mountcheck_volumes(hammer_mount_t hmp);
1128 int     hammer_get_installed_volumes(hammer_mount_t hmp);
1129
1130 int     hammer_mem_add(hammer_record_t record);
1131 int     hammer_ip_lookup(hammer_cursor_t cursor);
1132 int     hammer_ip_first(hammer_cursor_t cursor);
1133 int     hammer_ip_next(hammer_cursor_t cursor);
1134 int     hammer_ip_resolve_data(hammer_cursor_t cursor);
1135 int     hammer_ip_delete_record(hammer_cursor_t cursor, hammer_inode_t ip,
1136                         hammer_tid_t tid);
1137 int     hammer_create_at_cursor(hammer_cursor_t cursor,
1138                         hammer_btree_leaf_elm_t leaf, void *udata, int mode);
1139 int     hammer_delete_at_cursor(hammer_cursor_t cursor, int delete_flags,
1140                         hammer_tid_t delete_tid, u_int32_t delete_ts,
1141                         int track, int64_t *stat_bytes);
1142 int     hammer_ip_check_directory_empty(hammer_transaction_t trans,
1143                         hammer_inode_t ip);
1144 int     hammer_sync_hmp(hammer_mount_t hmp, int waitfor);
1145 int     hammer_queue_inodes_flusher(hammer_mount_t hmp, int waitfor);
1146
1147 hammer_record_t
1148         hammer_alloc_mem_record(hammer_inode_t ip, int data_len);
1149 void    hammer_flush_record_done(hammer_record_t record, int error);
1150 void    hammer_wait_mem_record_ident(hammer_record_t record, const char *ident);
1151 void    hammer_rel_mem_record(hammer_record_t record);
1152
1153 int     hammer_cursor_up(hammer_cursor_t cursor);
1154 int     hammer_cursor_up_locked(hammer_cursor_t cursor);
1155 int     hammer_cursor_down(hammer_cursor_t cursor);
1156 int     hammer_cursor_upgrade(hammer_cursor_t cursor);
1157 int     hammer_cursor_upgrade_node(hammer_cursor_t cursor);
1158 void    hammer_cursor_downgrade(hammer_cursor_t cursor);
1159 int     hammer_cursor_upgrade2(hammer_cursor_t c1, hammer_cursor_t c2);
1160 void    hammer_cursor_downgrade2(hammer_cursor_t c1, hammer_cursor_t c2);
1161 int     hammer_cursor_seek(hammer_cursor_t cursor, hammer_node_t node,
1162                         int index);
1163 void    hammer_lock_ex_ident(struct hammer_lock *lock, const char *ident);
1164 int     hammer_lock_ex_try(struct hammer_lock *lock);
1165 void    hammer_lock_sh(struct hammer_lock *lock);
1166 int     hammer_lock_sh_try(struct hammer_lock *lock);
1167 int     hammer_lock_upgrade(struct hammer_lock *lock, int shcount);
1168 void    hammer_lock_downgrade(struct hammer_lock *lock, int shcount);
1169 int     hammer_lock_status(struct hammer_lock *lock);
1170 void    hammer_unlock(struct hammer_lock *lock);
1171 void    hammer_ref(struct hammer_lock *lock);
1172 int     hammer_ref_interlock(struct hammer_lock *lock);
1173 int     hammer_ref_interlock_true(struct hammer_lock *lock);
1174 void    hammer_ref_interlock_done(struct hammer_lock *lock);
1175 void    hammer_rel(struct hammer_lock *lock);
1176 int     hammer_rel_interlock(struct hammer_lock *lock, int locked);
1177 void    hammer_rel_interlock_done(struct hammer_lock *lock, int orig_locked);
1178 int     hammer_get_interlock(struct hammer_lock *lock);
1179 int     hammer_try_interlock_norefs(struct hammer_lock *lock);
1180 void    hammer_put_interlock(struct hammer_lock *lock, int error);
1181
1182 void    hammer_sync_lock_ex(hammer_transaction_t trans);
1183 void    hammer_sync_lock_sh(hammer_transaction_t trans);
1184 int     hammer_sync_lock_sh_try(hammer_transaction_t trans);
1185 void    hammer_sync_unlock(hammer_transaction_t trans);
1186
1187 u_int32_t hammer_to_unix_xid(uuid_t *uuid);
1188 void hammer_guid_to_uuid(uuid_t *uuid, u_int32_t guid);
1189 void    hammer_time_to_timespec(u_int64_t xtime, struct timespec *ts);
1190 u_int64_t hammer_timespec_to_time(struct timespec *ts);
1191 int     hammer_str_to_tid(const char *str, int *ispfsp,
1192                         hammer_tid_t *tidp, u_int32_t *localizationp);
1193 int     hammer_is_atatext(const char *name, int len);
1194 hammer_tid_t hammer_alloc_objid(hammer_mount_t hmp, hammer_inode_t dip,
1195                         int64_t namekey);
1196 void hammer_clear_objid(hammer_inode_t dip);
1197 void hammer_destroy_objid_cache(hammer_mount_t hmp);
1198
1199 int hammer_dedup_crc_rb_compare(hammer_dedup_cache_t dc1,
1200                         hammer_dedup_cache_t dc2);
1201 int hammer_dedup_off_rb_compare(hammer_dedup_cache_t dc1,
1202                         hammer_dedup_cache_t dc2);
1203 hammer_dedup_cache_t hammer_dedup_cache_add(hammer_inode_t ip,
1204                         hammer_btree_leaf_elm_t leaf);
1205 hammer_dedup_cache_t hammer_dedup_cache_lookup(hammer_mount_t hmp,
1206                         hammer_crc_t crc);
1207 void hammer_dedup_cache_inval(hammer_mount_t hmp, hammer_off_t base_offset);
1208 void hammer_destroy_dedup_cache(hammer_mount_t hmp);
1209 void hammer_dump_dedup_cache(hammer_mount_t hmp);
1210 int hammer_dedup_validate(hammer_dedup_cache_t dcp, int zone, int bytes,
1211                         void *data);
1212
1213 int hammer_enter_undo_history(hammer_mount_t hmp, hammer_off_t offset,
1214                         int bytes);
1215 void hammer_clear_undo_history(hammer_mount_t hmp);
1216 enum vtype hammer_get_vnode_type(u_int8_t obj_type);
1217 int hammer_get_dtype(u_int8_t obj_type);
1218 u_int8_t hammer_get_obj_type(enum vtype vtype);
1219 int64_t hammer_directory_namekey(hammer_inode_t dip, const void *name, int len,
1220                         u_int32_t *max_iterationsp);
1221 int     hammer_nohistory(hammer_inode_t ip);
1222
1223 int     hammer_init_cursor(hammer_transaction_t trans, hammer_cursor_t cursor,
1224                         hammer_node_cache_t cache, hammer_inode_t ip);
1225 void    hammer_normalize_cursor(hammer_cursor_t cursor);
1226 void    hammer_done_cursor(hammer_cursor_t cursor);
1227 int     hammer_recover_cursor(hammer_cursor_t cursor);
1228 void    hammer_unlock_cursor(hammer_cursor_t cursor);
1229 int     hammer_lock_cursor(hammer_cursor_t cursor);
1230 hammer_cursor_t hammer_push_cursor(hammer_cursor_t ocursor);
1231 void    hammer_pop_cursor(hammer_cursor_t ocursor, hammer_cursor_t ncursor);
1232
1233 void    hammer_cursor_replaced_node(hammer_node_t onode, hammer_node_t nnode);
1234 void    hammer_cursor_removed_node(hammer_node_t onode, hammer_node_t parent,
1235                         int index);
1236 void    hammer_cursor_split_node(hammer_node_t onode, hammer_node_t nnode,
1237                         int index);
1238 void    hammer_cursor_moved_element(hammer_node_t oparent, int pindex,
1239                         hammer_node_t onode, int oindex,
1240                         hammer_node_t nnode, int nindex);
1241 void    hammer_cursor_parent_changed(hammer_node_t node, hammer_node_t oparent,
1242                         hammer_node_t nparent, int nindex);
1243 void    hammer_cursor_inserted_element(hammer_node_t node, int index);
1244 void    hammer_cursor_deleted_element(hammer_node_t node, int index);
1245 void    hammer_cursor_invalidate_cache(hammer_cursor_t cursor);
1246
1247 int     hammer_btree_lookup(hammer_cursor_t cursor);
1248 int     hammer_btree_first(hammer_cursor_t cursor);
1249 int     hammer_btree_last(hammer_cursor_t cursor);
1250 int     hammer_btree_extract(hammer_cursor_t cursor, int flags);
1251 int     hammer_btree_iterate(hammer_cursor_t cursor);
1252 int     hammer_btree_iterate_reverse(hammer_cursor_t cursor);
1253 int     hammer_btree_insert(hammer_cursor_t cursor,
1254                             hammer_btree_leaf_elm_t elm, int *doprop);
1255 int     hammer_btree_delete(hammer_cursor_t cursor, int *ndelete);
1256 void    hammer_btree_do_propagation(hammer_cursor_t cursor,
1257                             hammer_pseudofs_inmem_t pfsm,
1258                             hammer_btree_leaf_elm_t leaf);
1259 int     hammer_btree_cmp(hammer_base_elm_t key1, hammer_base_elm_t key2);
1260 int     hammer_btree_chkts(hammer_tid_t ts, hammer_base_elm_t key);
1261 int     hammer_btree_correct_rhb(hammer_cursor_t cursor, hammer_tid_t tid);
1262 int     hammer_btree_correct_lhb(hammer_cursor_t cursor, hammer_tid_t tid);
1263
1264 int     btree_set_parent_of_child(hammer_transaction_t trans,
1265                         hammer_node_t node,
1266                         hammer_btree_elm_t elm);
1267 void    hammer_node_lock_init(hammer_node_lock_t parent, hammer_node_t node);
1268 void    hammer_btree_lcache_init(hammer_mount_t hmp, hammer_node_lock_t lcache,
1269                         int depth);
1270 void    hammer_btree_lcache_free(hammer_mount_t hmp, hammer_node_lock_t lcache);
1271 int     hammer_btree_lock_children(hammer_cursor_t cursor, int depth,
1272                         hammer_node_lock_t parent,
1273                         hammer_node_lock_t lcache);
1274 void    hammer_btree_lock_copy(hammer_cursor_t cursor,
1275                         hammer_node_lock_t parent);
1276 int     hammer_btree_sync_copy(hammer_cursor_t cursor,
1277                         hammer_node_lock_t parent);
1278 void    hammer_btree_unlock_children(hammer_mount_t hmp,
1279                         hammer_node_lock_t parent,
1280                         hammer_node_lock_t lcache);
1281 int     hammer_btree_search_node(hammer_base_elm_t elm, hammer_node_ondisk_t node);
1282 hammer_node_t hammer_btree_get_parent(hammer_transaction_t trans,
1283                         hammer_node_t node, int *parent_indexp,
1284                         int *errorp, int try_exclusive);
1285
1286 void    hammer_print_btree_node(hammer_node_ondisk_t ondisk);
1287 void    hammer_print_btree_elm(hammer_btree_elm_t elm);
1288
1289 void    *hammer_bread(hammer_mount_t hmp, hammer_off_t off,
1290                         int *errorp, struct hammer_buffer **bufferp);
1291 void    *hammer_bnew(hammer_mount_t hmp, hammer_off_t off,
1292                         int *errorp, struct hammer_buffer **bufferp);
1293 void    *hammer_bread_ext(hammer_mount_t hmp, hammer_off_t off, int bytes,
1294                         int *errorp, struct hammer_buffer **bufferp);
1295 void    *hammer_bnew_ext(hammer_mount_t hmp, hammer_off_t off, int bytes,
1296                         int *errorp, struct hammer_buffer **bufferp);
1297
1298 hammer_volume_t hammer_get_root_volume(hammer_mount_t hmp, int *errorp);
1299
1300 hammer_volume_t hammer_get_volume(hammer_mount_t hmp,
1301                         int32_t vol_no, int *errorp);
1302 hammer_buffer_t hammer_get_buffer(hammer_mount_t hmp, hammer_off_t buf_offset,
1303                         int bytes, int isnew, int *errorp);
1304 void            hammer_sync_buffers(hammer_mount_t hmp,
1305                         hammer_off_t base_offset, int bytes);
1306 int             hammer_del_buffers(hammer_mount_t hmp,
1307                         hammer_off_t base_offset,
1308                         hammer_off_t zone2_offset, int bytes,
1309                         int report_conflicts);
1310
1311 int             hammer_ref_volume(hammer_volume_t volume);
1312 int             hammer_ref_buffer(hammer_buffer_t buffer);
1313 void            hammer_flush_buffer_nodes(hammer_buffer_t buffer);
1314
1315 void            hammer_rel_volume(hammer_volume_t volume, int locked);
1316 void            hammer_rel_buffer(hammer_buffer_t buffer, int locked);
1317
1318 int             hammer_vfs_export(struct mount *mp, int op,
1319                         const struct export_args *export);
1320 hammer_node_t   hammer_get_node(hammer_transaction_t trans,
1321                         hammer_off_t node_offset, int isnew, int *errorp);
1322 void            hammer_ref_node(hammer_node_t node);
1323 hammer_node_t   hammer_ref_node_safe(hammer_transaction_t trans,
1324                         hammer_node_cache_t cache, int *errorp);
1325 void            hammer_rel_node(hammer_node_t node);
1326 void            hammer_delete_node(hammer_transaction_t trans,
1327                         hammer_node_t node);
1328 void            hammer_cache_node(hammer_node_cache_t cache,
1329                         hammer_node_t node);
1330 void            hammer_uncache_node(hammer_node_cache_t cache);
1331 void            hammer_flush_node(hammer_node_t node, int locked);
1332
1333 void hammer_dup_buffer(struct hammer_buffer **bufferp,
1334                         struct hammer_buffer *buffer);
1335 hammer_node_t hammer_alloc_btree(hammer_transaction_t trans,
1336                         hammer_off_t hint, int *errorp);
1337 void *hammer_alloc_data(hammer_transaction_t trans, int32_t data_len,
1338                         u_int16_t rec_type, hammer_off_t *data_offsetp,
1339                         struct hammer_buffer **data_bufferp,
1340                         hammer_off_t hint, int *errorp);
1341
1342 int hammer_generate_undo(hammer_transaction_t trans,
1343                         hammer_off_t zone_offset, void *base, int len);
1344 int hammer_generate_redo(hammer_transaction_t trans, hammer_inode_t ip,
1345                         hammer_off_t file_offset, u_int32_t flags,
1346                         void *base, int len);
1347 void hammer_generate_redo_sync(hammer_transaction_t trans);
1348 void hammer_redo_fifo_start_flush(hammer_inode_t ip);
1349 void hammer_redo_fifo_end_flush(hammer_inode_t ip);
1350
1351 void hammer_format_undo(void *base, u_int32_t seqno);
1352 int hammer_upgrade_undo_4(hammer_transaction_t trans);
1353
1354 void hammer_put_volume(struct hammer_volume *volume, int flush);
1355 void hammer_put_buffer(struct hammer_buffer *buffer, int flush);
1356
1357 hammer_off_t hammer_freemap_alloc(hammer_transaction_t trans,
1358                         hammer_off_t owner, int *errorp);
1359 void hammer_freemap_free(hammer_transaction_t trans, hammer_off_t phys_offset,
1360                         hammer_off_t owner, int *errorp);
1361 int _hammer_checkspace(hammer_mount_t hmp, int slop, int64_t *resp);
1362 hammer_off_t hammer_blockmap_alloc(hammer_transaction_t trans, int zone,
1363                         int bytes, hammer_off_t hint, int *errorp);
1364 hammer_reserve_t hammer_blockmap_reserve(hammer_mount_t hmp, int zone,
1365                         int bytes, hammer_off_t *zone_offp, int *errorp);
1366 hammer_reserve_t hammer_blockmap_reserve_dedup(hammer_mount_t hmp, int zone,
1367                         int bytes, hammer_off_t zone_offset, int *errorp);
1368 void hammer_blockmap_reserve_complete(hammer_mount_t hmp,
1369                         hammer_reserve_t resv);
1370 void hammer_reserve_clrdelay(hammer_mount_t hmp, hammer_reserve_t resv);
1371 void hammer_blockmap_free(hammer_transaction_t trans,
1372                         hammer_off_t zone_offset, int bytes);
1373 int hammer_blockmap_dedup(hammer_transaction_t trans,
1374                         hammer_off_t zone_offset, int bytes);
1375 int hammer_blockmap_finalize(hammer_transaction_t trans,
1376                         hammer_reserve_t resv,
1377                         hammer_off_t zone_offset, int bytes);
1378 int hammer_blockmap_getfree(hammer_mount_t hmp, hammer_off_t zone_offset,
1379                         int *curp, int *errorp);
1380 hammer_off_t hammer_blockmap_lookup_verify(hammer_mount_t hmp,
1381                         hammer_off_t zone_offset, int *errorp);
1382
1383 hammer_off_t hammer_undo_lookup(hammer_mount_t hmp, hammer_off_t zone_offset,
1384                         int *errorp);
1385 int64_t hammer_undo_used(hammer_transaction_t trans);
1386 int64_t hammer_undo_space(hammer_transaction_t trans);
1387 int64_t hammer_undo_max(hammer_mount_t hmp);
1388 int hammer_undo_reclaim(hammer_io_t io);
1389
1390 void hammer_start_transaction(struct hammer_transaction *trans,
1391                               struct hammer_mount *hmp);
1392 void hammer_simple_transaction(struct hammer_transaction *trans,
1393                               struct hammer_mount *hmp);
1394 void hammer_start_transaction_fls(struct hammer_transaction *trans,
1395                                   struct hammer_mount *hmp);
1396 void hammer_done_transaction(struct hammer_transaction *trans);
1397 hammer_tid_t hammer_alloc_tid(hammer_mount_t hmp, int count);
1398
1399 void hammer_modify_inode(hammer_transaction_t trans, hammer_inode_t ip, int flags);
1400 void hammer_flush_inode(hammer_inode_t ip, int flags);
1401 void hammer_flush_inode_done(hammer_inode_t ip, int error);
1402 void hammer_wait_inode(hammer_inode_t ip);
1403
1404 int  hammer_create_inode(struct hammer_transaction *trans, struct vattr *vap,
1405                         struct ucred *cred, struct hammer_inode *dip,
1406                         const char *name, int namelen,
1407                         hammer_pseudofs_inmem_t pfsm,
1408                         struct hammer_inode **ipp);
1409 void hammer_rel_inode(hammer_inode_t ip, int flush);
1410 int hammer_reload_inode(hammer_inode_t ip, void *arg __unused);
1411 int hammer_ino_rb_compare(hammer_inode_t ip1, hammer_inode_t ip2);
1412 int hammer_redo_rb_compare(hammer_inode_t ip1, hammer_inode_t ip2);
1413 int hammer_destroy_inode_callback(hammer_inode_t ip, void *data __unused);
1414
1415 int hammer_sync_inode(hammer_transaction_t trans, hammer_inode_t ip);
1416 void hammer_test_inode(hammer_inode_t dip);
1417 void hammer_inode_unloadable_check(hammer_inode_t ip, int getvp);
1418 int hammer_update_atime_quick(hammer_inode_t ip);
1419
1420 int  hammer_ip_add_directory(struct hammer_transaction *trans,
1421                         hammer_inode_t dip, const char *name, int bytes,
1422                         hammer_inode_t nip);
1423 int  hammer_ip_del_directory(struct hammer_transaction *trans,
1424                         hammer_cursor_t cursor, hammer_inode_t dip,
1425                         hammer_inode_t ip);
1426 void hammer_ip_replace_bulk(hammer_mount_t hmp, hammer_record_t record);
1427 hammer_record_t hammer_ip_add_bulk(hammer_inode_t ip, off_t file_offset,
1428                         void *data, int bytes, int *errorp);
1429 int  hammer_ip_frontend_trunc(struct hammer_inode *ip, off_t file_size);
1430 int  hammer_ip_add_record(struct hammer_transaction *trans,
1431                         hammer_record_t record);
1432 int  hammer_ip_delete_range(hammer_cursor_t cursor, hammer_inode_t ip,
1433                         int64_t ran_beg, int64_t ran_end, int truncating);
1434 int  hammer_ip_delete_clean(hammer_cursor_t cursor, hammer_inode_t ip,
1435                         int *countp);
1436 int  hammer_ip_sync_data(hammer_cursor_t cursor, hammer_inode_t ip,
1437                         int64_t offset, void *data, int bytes);
1438 int  hammer_ip_sync_record_cursor(hammer_cursor_t cursor, hammer_record_t rec);
1439 hammer_pseudofs_inmem_t  hammer_load_pseudofs(hammer_transaction_t trans,
1440                         u_int32_t localization, int *errorp);
1441 int  hammer_mkroot_pseudofs(hammer_transaction_t trans, struct ucred *cred,
1442                         hammer_pseudofs_inmem_t pfsm);
1443 int  hammer_save_pseudofs(hammer_transaction_t trans,
1444                         hammer_pseudofs_inmem_t pfsm);
1445 int  hammer_unload_pseudofs(hammer_transaction_t trans, u_int32_t localization);
1446 void hammer_rel_pseudofs(hammer_mount_t hmp, hammer_pseudofs_inmem_t pfsm);
1447 int hammer_ioctl(hammer_inode_t ip, u_long com, caddr_t data, int fflag,
1448                         struct ucred *cred);
1449
1450 void hammer_io_init(hammer_io_t io, hammer_volume_t volume,
1451                         enum hammer_io_type type);
1452 int hammer_io_read(struct vnode *devvp, struct hammer_io *io, int limit);
1453 void hammer_io_advance(struct hammer_io *io);
1454 int hammer_io_new(struct vnode *devvp, struct hammer_io *io);
1455 int hammer_io_inval(hammer_volume_t volume, hammer_off_t zone2_offset);
1456 struct buf *hammer_io_release(struct hammer_io *io, int flush);
1457 void hammer_io_flush(struct hammer_io *io, int reclaim);
1458 void hammer_io_wait(struct hammer_io *io);
1459 void hammer_io_waitdep(struct hammer_io *io);
1460 void hammer_io_wait_all(hammer_mount_t hmp, const char *ident, int doflush);
1461 int hammer_io_direct_read(hammer_mount_t hmp, struct bio *bio,
1462                         hammer_btree_leaf_elm_t leaf);
1463 int hammer_io_indirect_read(hammer_mount_t hmp, struct bio *bio,
1464                         hammer_btree_leaf_elm_t leaf);
1465 int hammer_io_direct_write(hammer_mount_t hmp, struct bio *bio,
1466                         hammer_record_t record);
1467 void hammer_io_direct_wait(hammer_record_t record);
1468 void hammer_io_direct_uncache(hammer_mount_t hmp, hammer_btree_leaf_elm_t leaf);
1469 void hammer_io_write_interlock(hammer_io_t io);
1470 void hammer_io_done_interlock(hammer_io_t io);
1471 void hammer_io_clear_modify(struct hammer_io *io, int inval);
1472 void hammer_io_clear_modlist(struct hammer_io *io);
1473 void hammer_io_flush_sync(hammer_mount_t hmp);
1474 void hammer_io_clear_error(struct hammer_io *io);
1475 void hammer_io_clear_error_noassert(struct hammer_io *io);
1476 void hammer_io_notmeta(hammer_buffer_t buffer);
1477 void hammer_io_limit_backlog(hammer_mount_t hmp);
1478
1479 void hammer_modify_volume(hammer_transaction_t trans, hammer_volume_t volume,
1480                         void *base, int len);
1481 void hammer_modify_buffer(hammer_transaction_t trans, hammer_buffer_t buffer,
1482                         void *base, int len);
1483 void hammer_modify_volume_done(hammer_volume_t volume);
1484 void hammer_modify_buffer_done(hammer_buffer_t buffer);
1485
1486 int hammer_ioc_reblock(hammer_transaction_t trans, hammer_inode_t ip,
1487                         struct hammer_ioc_reblock *reblock);
1488 int hammer_ioc_rebalance(hammer_transaction_t trans, hammer_inode_t ip,
1489                         struct hammer_ioc_rebalance *rebal);
1490 int hammer_ioc_prune(hammer_transaction_t trans, hammer_inode_t ip,
1491                         struct hammer_ioc_prune *prune);
1492 int hammer_ioc_mirror_read(hammer_transaction_t trans, hammer_inode_t ip,
1493                         struct hammer_ioc_mirror_rw *mirror);
1494 int hammer_ioc_mirror_write(hammer_transaction_t trans, hammer_inode_t ip,
1495                         struct hammer_ioc_mirror_rw *mirror);
1496 int hammer_ioc_set_pseudofs(hammer_transaction_t trans, hammer_inode_t ip,
1497                         struct ucred *cred, struct hammer_ioc_pseudofs_rw *pfs);
1498 int hammer_ioc_get_pseudofs(hammer_transaction_t trans, hammer_inode_t ip,
1499                         struct hammer_ioc_pseudofs_rw *pfs);
1500 int hammer_ioc_destroy_pseudofs(hammer_transaction_t trans, hammer_inode_t ip,
1501                         struct hammer_ioc_pseudofs_rw *pfs);
1502 int hammer_ioc_downgrade_pseudofs(hammer_transaction_t trans, hammer_inode_t ip,
1503                         struct hammer_ioc_pseudofs_rw *pfs);
1504 int hammer_ioc_upgrade_pseudofs(hammer_transaction_t trans, hammer_inode_t ip,
1505                         struct hammer_ioc_pseudofs_rw *pfs);
1506 int hammer_ioc_wait_pseudofs(hammer_transaction_t trans, hammer_inode_t ip,
1507                         struct hammer_ioc_pseudofs_rw *pfs);
1508 int hammer_ioc_iterate_pseudofs(hammer_transaction_t trans, hammer_inode_t ip,
1509                         struct hammer_ioc_pfs_iterate *pi);
1510 int hammer_ioc_volume_add(hammer_transaction_t trans, hammer_inode_t ip,
1511                         struct hammer_ioc_volume *ioc);
1512 int hammer_ioc_volume_del(hammer_transaction_t trans, hammer_inode_t ip,
1513                         struct hammer_ioc_volume *ioc);
1514 int hammer_ioc_volume_list(hammer_transaction_t trans, hammer_inode_t ip,
1515                         struct hammer_ioc_volume_list *ioc);
1516 int hammer_ioc_dedup(hammer_transaction_t trans, hammer_inode_t ip,
1517                         struct hammer_ioc_dedup *dedup);
1518
1519 int hammer_signal_check(hammer_mount_t hmp);
1520
1521 void hammer_flusher_create(hammer_mount_t hmp);
1522 void hammer_flusher_destroy(hammer_mount_t hmp);
1523 void hammer_flusher_sync(hammer_mount_t hmp);
1524 int  hammer_flusher_async(hammer_mount_t hmp, hammer_flush_group_t flg);
1525 int  hammer_flusher_async_one(hammer_mount_t hmp);
1526 int hammer_flusher_running(hammer_mount_t hmp);
1527 void hammer_flusher_wait(hammer_mount_t hmp, int seq);
1528 void hammer_flusher_wait_next(hammer_mount_t hmp);
1529 int  hammer_flusher_meta_limit(hammer_mount_t hmp);
1530 int  hammer_flusher_meta_halflimit(hammer_mount_t hmp);
1531 int  hammer_flusher_undo_exhausted(hammer_transaction_t trans, int quarter);
1532 void hammer_flusher_clean_loose_ios(hammer_mount_t hmp);
1533 void hammer_flusher_finalize(hammer_transaction_t trans, int final);
1534 int  hammer_flusher_haswork(hammer_mount_t hmp);
1535 int  hammer_flush_dirty(hammer_mount_t hmp, int max_count);
1536 void hammer_flusher_flush_undos(hammer_mount_t hmp, int already_flushed);
1537
1538 int hammer_recover_stage1(hammer_mount_t hmp, hammer_volume_t rootvol);
1539 int hammer_recover_stage2(hammer_mount_t hmp, hammer_volume_t rootvol);
1540 void hammer_recover_flush_buffers(hammer_mount_t hmp,
1541                         hammer_volume_t root_volume, int final);
1542
1543 void hammer_crc_set_blockmap(hammer_blockmap_t blockmap);
1544 void hammer_crc_set_volume(hammer_volume_ondisk_t ondisk);
1545 void hammer_crc_set_leaf(void *data, hammer_btree_leaf_elm_t leaf);
1546
1547 int hammer_crc_test_blockmap(hammer_blockmap_t blockmap);
1548 int hammer_crc_test_volume(hammer_volume_ondisk_t ondisk);
1549 int hammer_crc_test_btree(hammer_node_ondisk_t ondisk);
1550 int hammer_crc_test_leaf(void *data, hammer_btree_leaf_elm_t leaf);
1551 void hkprintf(const char *ctl, ...) __printflike(1, 2);
1552 udev_t hammer_fsid_to_udev(uuid_t *uuid);
1553
1554
1555 int hammer_blocksize(int64_t file_offset);
1556 int hammer_blockoff(int64_t file_offset);
1557 int64_t hammer_blockdemarc(int64_t file_offset1, int64_t file_offset2);
1558
1559 /*
1560  * Shortcut for _hammer_checkspace(), used all over the code.
1561  */
1562 static __inline int
1563 hammer_checkspace(hammer_mount_t hmp, int slop)
1564 {
1565         return(_hammer_checkspace(hmp, slop, NULL));
1566 }
1567
1568 static __inline void
1569 hammer_wait_mem_record(hammer_record_t record)
1570 {
1571         hammer_wait_mem_record_ident(record, "hmmwai");
1572 }
1573
1574 static __inline void
1575 hammer_lock_ex(struct hammer_lock *lock)
1576 {
1577         hammer_lock_ex_ident(lock, "hmrlck");
1578 }
1579
1580 static __inline void
1581 hammer_modify_volume_noundo(hammer_transaction_t trans, hammer_volume_t volume)
1582 {
1583         hammer_modify_volume(trans, volume, NULL, 0);
1584 }
1585
1586 static __inline void
1587 hammer_modify_buffer_noundo(hammer_transaction_t trans, hammer_buffer_t buffer)
1588 {
1589         hammer_modify_buffer(trans, buffer, NULL, 0);
1590 }
1591
1592 /*
1593  * Indicate that a B-Tree node is being modified.
1594  */
1595 static __inline void
1596 hammer_modify_node_noundo(hammer_transaction_t trans, hammer_node_t node)
1597 {
1598         KKASSERT((node->flags & HAMMER_NODE_CRCBAD) == 0);
1599         hammer_modify_buffer(trans, node->buffer, NULL, 0);
1600 }
1601
1602 static __inline void
1603 hammer_modify_node_all(hammer_transaction_t trans, struct hammer_node *node)
1604 {
1605         KKASSERT((node->flags & HAMMER_NODE_CRCBAD) == 0);
1606         hammer_modify_buffer(trans, node->buffer,
1607                              node->ondisk, sizeof(*node->ondisk));
1608 }
1609
1610 static __inline void
1611 hammer_modify_node(hammer_transaction_t trans, hammer_node_t node,
1612                    void *base, int len)
1613 {
1614         hammer_crc_t *crcptr;
1615
1616         KKASSERT((char *)base >= (char *)node->ondisk &&
1617                  (char *)base + len <=
1618                     (char *)node->ondisk + sizeof(*node->ondisk));
1619         KKASSERT((node->flags & HAMMER_NODE_CRCBAD) == 0);
1620
1621         if (hammer_btree_full_undo) {
1622                 hammer_modify_node_all(trans, node);
1623         } else {
1624                 hammer_modify_buffer(trans, node->buffer, base, len);
1625                 crcptr = &node->ondisk->crc;
1626                 hammer_modify_buffer(trans, node->buffer,
1627                                      crcptr, sizeof(hammer_crc_t));
1628                 --node->buffer->io.modify_refs; /* only want one ref */
1629         }
1630 }
1631
1632 /*
1633  * Indicate that the specified modifications have been completed.
1634  *
1635  * Do not try to generate the crc here, it's very expensive to do and a
1636  * sequence of insertions or deletions can result in many calls to this
1637  * function on the same node.
1638  */
1639 static __inline void
1640 hammer_modify_node_done(hammer_node_t node)
1641 {
1642         node->flags |= HAMMER_NODE_CRCGOOD;
1643         if ((node->flags & HAMMER_NODE_NEEDSCRC) == 0) {
1644                 node->flags |= HAMMER_NODE_NEEDSCRC;
1645                 node->buffer->io.gencrc = 1;
1646                 hammer_ref_node(node);
1647         }
1648         hammer_modify_buffer_done(node->buffer);
1649 }
1650
1651 /*
1652  * Lookup a blockmap offset.
1653  */
1654 static __inline hammer_off_t
1655 hammer_blockmap_lookup(hammer_mount_t hmp, hammer_off_t zone_offset,
1656                         int *errorp)
1657 {
1658 #if defined INVARIANTS
1659         int zone = HAMMER_ZONE_DECODE(zone_offset);
1660         KKASSERT(zone >= HAMMER_ZONE2_MAPPED_INDEX && zone < HAMMER_MAX_ZONES);
1661 #endif
1662
1663         /*
1664          * We can actually skip blockmap verify by default,
1665          * as normal blockmaps are now direct-mapped onto the freemap
1666          * and so represent zone-2 addresses.
1667          */
1668         if (hammer_verify_zone == 0) {
1669                 *errorp = 0;
1670                 return hammer_xlate_to_zone2(zone_offset);
1671         }
1672
1673         return hammer_blockmap_lookup_verify(hmp, zone_offset, errorp);
1674 }
1675
1676 #define hammer_modify_volume_field(trans, vol, field)           \
1677         hammer_modify_volume(trans, vol, &(vol)->ondisk->field, \
1678                              sizeof((vol)->ondisk->field))
1679
1680 #define hammer_modify_node_field(trans, node, field)            \
1681         hammer_modify_node(trans, node, &(node)->ondisk->field, \
1682                              sizeof((node)->ondisk->field))
1683
1684 /*
1685  * The HAMMER_INODE_CAP_DIR_LOCAL_INO capability is set on newly
1686  * created directories for HAMMER version 2 or greater and causes
1687  * directory entries to be placed the inode localization zone in
1688  * the B-Tree instead of the misc zone.
1689  *
1690  * This greatly improves localization between directory entries and
1691  * inodes
1692  */
1693 static __inline u_int32_t
1694 hammer_dir_localization(hammer_inode_t dip)
1695 {
1696         if (dip->ino_data.cap_flags & HAMMER_INODE_CAP_DIR_LOCAL_INO)
1697                 return(HAMMER_LOCALIZE_INODE);
1698         else
1699                 return(HAMMER_LOCALIZE_MISC);
1700 }
1701 #endif  /* _KERNEL */
1702
1703 #endif /* !VFS_HAMMER_HAMMER_H_ */