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  * This header file contains structures used internally by the HAMMERFS
  36  * implementation.  See hammer_disk.h for on-disk structures.
  37  */
  38
  39 #include <sys/param.h>
  40 #ifdef _KERNEL
  41 #include <sys/kernel.h>
  42 #include <sys/systm.h>
  43 #endif
  44 #include <sys/conf.h>
  45 #include <sys/tree.h>
  46 #include <sys/malloc.h>
  47 #include <sys/mount.h>
  48 #include <sys/mountctl.h>
  49 #include <sys/vnode.h>
  50 #include <sys/proc.h>
  51 #include <sys/priv.h>
  52 #include <sys/stat.h>
  53 #include <sys/globaldata.h>
  54 #include <sys/lockf.h>
  55 #include <sys/buf.h>
  56 #include <sys/queue.h>
  57 #include <sys/ktr.h>
  58 #include <sys/limits.h>
  59 #include <vm/vm_extern.h>
  60
  61 #ifdef _KERNEL
  62 #include <sys/buf2.h>
  63 #include <sys/signal2.h>
  64 #include <vm/vm_page2.h>
  65 #endif
  66
  67 #include "hammer_disk.h"
  68 #include "hammer_mount.h"
  69 #include "hammer_ioctl.h"
  70
  71 #if defined(_KERNEL) || defined(_KERNEL_STRUCTURES)
  72
  73 MALLOC_DECLARE(M_HAMMER);
  74
  75 /*
  76  * Kernel trace
  77  */
  78 #if !defined(KTR_HAMMER)
  79 #define KTR_HAMMER      KTR_ALL
  80 #endif
  81 /* KTR_INFO_MASTER_EXTERN(hammer); */
  82
  83 /*
  84  * Misc structures
  85  */
  86 struct hammer_mount;
  87
  88 /*
  89  * Key structure used for custom RB tree inode lookups.  This prototypes
  90  * the function hammer_ino_rb_tree_RB_LOOKUP_INFO(root, info).
  91  */
  92 typedef struct hammer_inode_info {
  93         int64_t         obj_id;         /* (key) object identifier */
  94         hammer_tid_t    obj_asof;       /* (key) snapshot transid or 0 */
  95         u_int32_t       obj_localization; /* (key) pseudo-fs */
  96         union {
  97                 struct hammer_btree_leaf_elm *leaf;
  98         } u;
  99 } *hammer_inode_info_t;
 100
 101 typedef enum hammer_transaction_type {
 102         HAMMER_TRANS_RO,
 103         HAMMER_TRANS_STD,
 104         HAMMER_TRANS_FLS
 105 } hammer_transaction_type_t;
 106
 107 /*
 108  * HAMMER Transaction tracking
 109  */
 110 struct hammer_transaction {
 111         hammer_transaction_type_t type;
 112         struct hammer_mount *hmp;
 113         hammer_tid_t    tid;
 114         u_int64_t       time;
 115         u_int32_t       time32;
 116         int             sync_lock_refs;
 117         int             flags;
 118         struct hammer_volume *rootvol;
 119 };
 120
 121 typedef struct hammer_transaction *hammer_transaction_t;
 122
 123 #define HAMMER_TRANSF_NEWINODE  0x0001
 124 #define HAMMER_TRANSF_DIDIO     0x0002
 125 #define HAMMER_TRANSF_CRCDOM    0x0004  /* EDOM on CRC error, less critical */
 126
 127 /*
 128  * HAMMER locks
 129  */
 130 struct hammer_lock {
 131         volatile u_int  refs;           /* active references */
 132         volatile u_int  lockval;        /* lock count and control bits */
 133         struct thread   *lowner;        /* owner if exclusively held */
 134         struct thread   *rowner;        /* owner if exclusively held */
 135 };
 136
 137 #define HAMMER_REFS_LOCKED      0x40000000      /* transition check */
 138 #define HAMMER_REFS_WANTED      0x20000000      /* transition check */
 139 #define HAMMER_REFS_CHECK       0x10000000      /* transition check */
 140
 141 #define HAMMER_REFS_FLAGS       (HAMMER_REFS_LOCKED | \
 142                                  HAMMER_REFS_WANTED | \
 143                                  HAMMER_REFS_CHECK)
 144
 145 #define HAMMER_LOCKF_EXCLUSIVE  0x40000000
 146 #define HAMMER_LOCKF_WANTED     0x20000000
 147
 148 static __inline int
 149 hammer_notlocked(struct hammer_lock *lock)
 150 {
 151         return(lock->lockval == 0);
 152 }
 153
 154 static __inline int
 155 hammer_islocked(struct hammer_lock *lock)
 156 {
 157         return(lock->lockval != 0);
 158 }
 159
 160 /*
 161  * Returns the number of refs on the object.
 162  */
 163 static __inline int
 164 hammer_isactive(struct hammer_lock *lock)
 165 {
 166         return(lock->refs & ~HAMMER_REFS_FLAGS);
 167 }
 168
 169 static __inline int
 170 hammer_oneref(struct hammer_lock *lock)
 171 {
 172         return((lock->refs & ~HAMMER_REFS_FLAGS) == 1);
 173 }
 174
 175 static __inline int
 176 hammer_norefs(struct hammer_lock *lock)
 177 {
 178         return((lock->refs & ~HAMMER_REFS_FLAGS) == 0);
 179 }
 180
 181 static __inline int
 182 hammer_norefsorlock(struct hammer_lock *lock)
 183 {
 184         return(lock->refs == 0);
 185 }
 186
 187 static __inline int
 188 hammer_refsorlock(struct hammer_lock *lock)
 189 {
 190         return(lock->refs != 0);
 191 }
 192
 193 /*
 194  * Return if we specifically own the lock exclusively.
 195  */
 196 static __inline int
 197 hammer_lock_excl_owned(struct hammer_lock *lock, thread_t td)
 198 {
 199         if ((lock->lockval & HAMMER_LOCKF_EXCLUSIVE) &&
 200             lock->lowner == td) {
 201                 return(1);
 202         }
 203         return(0);
 204 }
 205
 206 /*
 207  * Flush state, used by various structures
 208  */
 209 typedef enum hammer_inode_state {
 210         HAMMER_FST_IDLE,
 211         HAMMER_FST_SETUP,
 212         HAMMER_FST_FLUSH
 213 } hammer_inode_state_t;
 214
 215 TAILQ_HEAD(hammer_record_list, hammer_record);
 216
 217 /*
 218  * Pseudo-filesystem extended data tracking
 219  */
 220 struct hammer_pfs_rb_tree;
 221 struct hammer_pseudofs_inmem;
 222 RB_HEAD(hammer_pfs_rb_tree, hammer_pseudofs_inmem);
 223 RB_PROTOTYPE2(hammer_pfs_rb_tree, hammer_pseudofs_inmem, rb_node,
 224               hammer_pfs_rb_compare, u_int32_t);
 225
 226 struct hammer_pseudofs_inmem {
 227         RB_ENTRY(hammer_pseudofs_inmem) rb_node;
 228         struct hammer_lock      lock;
 229         u_int32_t               localization;
 230         hammer_tid_t            create_tid;
 231         int                     flags;
 232         udev_t                  fsid_udev;
 233         struct hammer_pseudofs_data pfsd;
 234 };
 235
 236 typedef struct hammer_pseudofs_inmem *hammer_pseudofs_inmem_t;
 237
 238 #define HAMMER_PFSM_DELETED     0x0001
 239
 240 /*
 241  * Cache object ids.  A fixed number of objid cache structures are
 242  * created to reserve object id's for newly created files in multiples
 243  * of 100,000, localized to a particular directory, and recycled as
 244  * needed.  This allows parallel create operations in different
 245  * directories to retain fairly localized object ids which in turn
 246  * improves reblocking performance and layout.
 247  */
 248 #define OBJID_CACHE_SIZE        2048
 249 #define OBJID_CACHE_BULK_BITS   10              /* 10 bits (1024)       */
 250 #define OBJID_CACHE_BULK        (32 * 32)       /* two level (1024)     */
 251 #define OBJID_CACHE_BULK_MASK   (OBJID_CACHE_BULK - 1)
 252 #define OBJID_CACHE_BULK_MASK64 ((u_int64_t)(OBJID_CACHE_BULK - 1))
 253
 254 typedef struct hammer_objid_cache {
 255         TAILQ_ENTRY(hammer_objid_cache) entry;
 256         struct hammer_inode             *dip;
 257         hammer_tid_t                    base_tid;
 258         int                             count;
 259         u_int32_t                       bm0;
 260         u_int32_t                       bm1[32];
 261 } *hammer_objid_cache_t;
 262
 263 /*
 264  * Associate an inode with a B-Tree node to cache search start positions
 265  */
 266 typedef struct hammer_node_cache {
 267         TAILQ_ENTRY(hammer_node_cache)  entry;
 268         struct hammer_node              *node;
 269         struct hammer_inode             *ip;
 270 } *hammer_node_cache_t;
 271
 272 TAILQ_HEAD(hammer_node_cache_list, hammer_node_cache);
 273
 274 /*
 275  * Live dedup cache
 276  */
 277 struct hammer_dedup_crc_rb_tree;
 278 RB_HEAD(hammer_dedup_crc_rb_tree, hammer_dedup_cache);
 279 RB_PROTOTYPE2(hammer_dedup_crc_rb_tree, hammer_dedup_cache, crc_entry,
 280                 hammer_dedup_crc_rb_compare, hammer_crc_t);
 281
 282 struct hammer_dedup_off_rb_tree;
 283 RB_HEAD(hammer_dedup_off_rb_tree, hammer_dedup_cache);
 284 RB_PROTOTYPE2(hammer_dedup_off_rb_tree, hammer_dedup_cache, off_entry,
 285                 hammer_dedup_off_rb_compare, hammer_off_t);
 286
 287 #define DEDUP_CACHE_SIZE        4096 /* XXX make it a dynamic tunable */
 288
 289 typedef struct hammer_dedup_cache {
 290         RB_ENTRY(hammer_dedup_cache) crc_entry;
 291         RB_ENTRY(hammer_dedup_cache) off_entry;
 292         TAILQ_ENTRY(hammer_dedup_cache) lru_entry;
 293         struct hammer_mount *hmp;
 294         int64_t obj_id;
 295         u_int32_t localization;
 296         off_t file_offset;
 297         int bytes;
 298         hammer_off_t data_offset;
 299         hammer_crc_t crc;
 300 } *hammer_dedup_cache_t;
 301
 302 /*
 303  * Structure used to organize flush groups.  Flush groups must be
 304  * organized into chunks in order to avoid blowing out the UNDO FIFO.
 305  * Without this a 'sync' could end up flushing 50,000 inodes in a single
 306  * transaction.
 307  */
 308 struct hammer_fls_rb_tree;
 309 RB_HEAD(hammer_fls_rb_tree, hammer_inode);
 310 RB_PROTOTYPE(hammer_fls_rb_tree, hammer_inode, rb_flsnode,
 311               hammer_ino_rb_compare);
 312
 313 struct hammer_flush_group {
 314         TAILQ_ENTRY(hammer_flush_group) flush_entry;
 315         struct hammer_fls_rb_tree       flush_tree;
 316         int                             seq;            /* our seq no */
 317         int                             total_count;    /* record load */
 318         int                             running;        /* group is running */
 319         int                             closed;
 320         int                             refs;
 321 };
 322
 323 typedef struct hammer_flush_group *hammer_flush_group_t;
 324
 325 TAILQ_HEAD(hammer_flush_group_list, hammer_flush_group);
 326
 327 /*
 328  * Structure used to represent an inode in-memory.
 329  *
 330  * The record and data associated with an inode may be out of sync with
 331  * the disk (xDIRTY flags), or not even on the disk at all (ONDISK flag
 332  * clear).
 333  *
 334  * An inode may also hold a cache of unsynchronized records, used for
 335  * database and directories only.  Unsynchronized regular file data is
 336  * stored in the buffer cache.
 337  *
 338  * NOTE: A file which is created and destroyed within the initial
 339  * synchronization period can wind up not doing any disk I/O at all.
 340  *
 341  * Finally, an inode may cache numerous disk-referencing B-Tree cursors.
 342  */
 343 struct hammer_ino_rb_tree;
 344 struct hammer_inode;
 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 struct hammer_redo_rb_tree;
 350 RB_HEAD(hammer_redo_rb_tree, hammer_inode);
 351 RB_PROTOTYPE2(hammer_redo_rb_tree, hammer_inode, rb_redonode,
 352               hammer_redo_rb_compare, hammer_off_t);
 353
 354 struct hammer_rec_rb_tree;
 355 struct hammer_record;
 356 RB_HEAD(hammer_rec_rb_tree, hammer_record);
 357 RB_PROTOTYPEX(hammer_rec_rb_tree, INFO, hammer_record, rb_node,
 358               hammer_rec_rb_compare, hammer_btree_leaf_elm_t);
 359
 360 TAILQ_HEAD(hammer_node_list, hammer_node);
 361
 362 struct hammer_inode {
 363         RB_ENTRY(hammer_inode)  rb_node;
 364         hammer_inode_state_t    flush_state;
 365         hammer_flush_group_t    flush_group;
 366         RB_ENTRY(hammer_inode)  rb_flsnode;     /* when on flush list */
 367         RB_ENTRY(hammer_inode)  rb_redonode;    /* when INODE_RDIRTY is set */
 368         struct hammer_record_list target_list;  /* target of dependant recs */
 369         int64_t                 obj_id;         /* (key) object identifier */
 370         hammer_tid_t            obj_asof;       /* (key) snapshot or 0 */
 371         u_int32_t               obj_localization; /* (key) pseudo-fs */
 372         struct hammer_mount     *hmp;
 373         hammer_objid_cache_t    objid_cache;
 374         int                     flags;
 375         int                     error;          /* flush error */
 376         int                     cursor_ip_refs; /* sanity */
 377         int                     cursor_exclreq_count;
 378         int                     rsv_recs;
 379         struct vnode            *vp;
 380         hammer_pseudofs_inmem_t pfsm;
 381         struct lockf            advlock;
 382         struct hammer_lock      lock;           /* sync copy interlock */
 383         off_t                   trunc_off;
 384         struct hammer_btree_leaf_elm ino_leaf;  /* in-memory cache */
 385         struct hammer_inode_data ino_data;      /* in-memory cache */
 386         struct hammer_rec_rb_tree rec_tree;     /* in-memory cache */
 387         int                     rec_generation;
 388         struct hammer_node_cache cache[4];      /* search initiate cache */
 389
 390         /*
 391          * When a demark is created to synchronize an inode to
 392          * disk, certain fields are copied so the front-end VOPs
 393          * can continue to run in parallel with the synchronization
 394          * occuring in the background.
 395          */
 396         int             sync_flags;             /* to-sync flags cache */
 397         off_t           sync_trunc_off;         /* to-sync truncation */
 398         off_t           save_trunc_off;         /* write optimization */
 399         struct hammer_btree_leaf_elm sync_ino_leaf; /* to-sync cache */
 400         struct hammer_inode_data sync_ino_data; /* to-sync cache */
 401         size_t          redo_count;
 402
 403         /*
 404          * Track the earliest offset in the UNDO/REDO FIFO containing
 405          * REDO records.  This is staged to the backend during flush
 406          * sequences.  While the inode is staged redo_fifo_next is used
 407          * to track the earliest offset for rotation into redo_fifo_start
 408          * on completion of the flush.
 409          */
 410         hammer_off_t    redo_fifo_start;
 411         hammer_off_t    redo_fifo_next;
 412 };
 413
 414 typedef struct hammer_inode *hammer_inode_t;
 415
 416 #define VTOI(vp)        ((struct hammer_inode *)(vp)->v_data)
 417
 418 /*
 419  * NOTE: DDIRTY does not include atime or mtime and does not include
 420  *       write-append size changes.  SDIRTY handles write-append size
 421  *       changes.
 422  *
 423  *       REDO indicates that REDO logging is active, creating a definitive
 424  *       stream of REDO records in the UNDO/REDO log for writes and
 425  *       truncations, including boundary records when/if REDO is turned off.
 426  *       REDO is typically enabled by fsync() and turned off if excessive
 427  *       writes without an fsync() occurs.
 428  *
 429  *       RDIRTY indicates that REDO records were laid down in the UNDO/REDO
 430  *       FIFO (even if REDO is turned off some might still be active) and
 431  *       still being tracked for this inode.  See hammer_redo.c
 432  */
 433                                         /* (not including atime/mtime) */
 434 #define HAMMER_INODE_DDIRTY     0x0001  /* in-memory ino_data is dirty */
 435 #define HAMMER_INODE_RSV_INODES 0x0002  /* hmp->rsv_inodes bumped */
 436 #define HAMMER_INODE_CONN_DOWN  0x0004  /* include in downward recursion */
 437 #define HAMMER_INODE_XDIRTY     0x0008  /* in-memory records */
 438 #define HAMMER_INODE_ONDISK     0x0010  /* inode is on-disk (else not yet) */
 439 #define HAMMER_INODE_FLUSH      0x0020  /* flush on last ref */
 440 #define HAMMER_INODE_DELETED    0x0080  /* inode delete (backend) */
 441 #define HAMMER_INODE_DELONDISK  0x0100  /* delete synchronized to disk */
 442 #define HAMMER_INODE_RO         0x0200  /* read-only (because of as-of) */
 443 #define HAMMER_INODE_RECSW      0x0400  /* waiting on data record flush */
 444 #define HAMMER_INODE_DONDISK    0x0800  /* data records may be on disk */
 445 #define HAMMER_INODE_BUFS       0x1000  /* dirty high level bps present */
 446 #define HAMMER_INODE_REFLUSH    0x2000  /* flush on dependancy / reflush */
 447 #define HAMMER_INODE_RECLAIM    0x4000  /* trying to reclaim */
 448 #define HAMMER_INODE_FLUSHW     0x8000  /* Someone waiting for flush */
 449
 450 #define HAMMER_INODE_TRUNCATED  0x00010000
 451 #define HAMMER_INODE_DELETING   0x00020000 /* inode delete request (frontend)*/
 452 #define HAMMER_INODE_RESIGNAL   0x00040000 /* re-signal on re-flush */
 453 #define HAMMER_INODE_ATIME      0x00100000 /* in-memory atime modified */
 454 #define HAMMER_INODE_MTIME      0x00200000 /* in-memory mtime modified */
 455 #define HAMMER_INODE_WOULDBLOCK 0x00400000 /* re-issue to new flush group */
 456 #define HAMMER_INODE_DUMMY      0x00800000 /* dummy inode covering bad file */
 457 #define HAMMER_INODE_SDIRTY     0x01000000 /* in-memory ino_data.size is dirty*/
 458 #define HAMMER_INODE_REDO       0x02000000 /* REDO logging active */
 459 #define HAMMER_INODE_RDIRTY     0x04000000 /* REDO records active in fifo */
 460 #define HAMMER_INODE_SLAVEFLUSH 0x08000000 /* being flushed by slave */
 461
 462 #define HAMMER_INODE_MODMASK    (HAMMER_INODE_DDIRTY|HAMMER_INODE_SDIRTY|   \
 463                                  HAMMER_INODE_XDIRTY|HAMMER_INODE_BUFS|     \
 464                                  HAMMER_INODE_ATIME|HAMMER_INODE_MTIME|     \
 465                                  HAMMER_INODE_TRUNCATED|HAMMER_INODE_DELETING)
 466
 467 #define HAMMER_INODE_MODMASK_NOXDIRTY   \
 468                                 (HAMMER_INODE_MODMASK & ~HAMMER_INODE_XDIRTY)
 469
 470 #define HAMMER_INODE_MODMASK_NOREDO     \
 471                                 (HAMMER_INODE_DDIRTY|                       \
 472                                  HAMMER_INODE_XDIRTY|                       \
 473                                  HAMMER_INODE_TRUNCATED|HAMMER_INODE_DELETING)
 474
 475 #define HAMMER_FLUSH_SIGNAL     0x0001
 476 #define HAMMER_FLUSH_RECURSION  0x0002
 477
 478 /*
 479  * Used by the inode reclaim code to pipeline reclaims and avoid
 480  * blowing out kernel memory or letting the flusher get too far
 481  * behind.  The reclaim wakes up when count reaches 0 or the
 482  * timer expires.
 483  */
 484 struct hammer_reclaim {
 485         TAILQ_ENTRY(hammer_reclaim) entry;
 486         int     count;
 487 };
 488
 489 /*
 490  * Track who is creating the greatest burden on the
 491  * inode cache.
 492  */
 493 struct hammer_inostats {
 494         pid_t           pid;    /* track user process */
 495         int             ltick;  /* last tick */
 496         int             count;  /* count (degenerates) */
 497 };
 498
 499 #define HAMMER_INOSTATS_HSIZE   32
 500 #define HAMMER_INOSTATS_HMASK   (HAMMER_INOSTATS_HSIZE - 1)
 501
 502 /*
 503  * Structure used to represent an unsynchronized record in-memory.  These
 504  * records typically represent directory entries.  Only non-historical
 505  * records are kept in-memory.
 506  *
 507  * Records are organized as a per-inode RB-Tree.  If the inode is not
 508  * on disk then neither are any records and the in-memory record tree
 509  * represents the entire contents of the inode.  If the inode is on disk
 510  * then the on-disk B-Tree is scanned in parallel with the in-memory
 511  * RB-Tree to synthesize the current state of the file.
 512  *
 513  * Records are also used to enforce the ordering of directory create/delete
 514  * operations.  A new inode will not be flushed to disk unless its related
 515  * directory entry is also being flushed at the same time.  A directory entry
 516  * will not be removed unless its related inode is also being removed at the
 517  * same time.
 518  */
 519 typedef enum hammer_record_type {
 520         HAMMER_MEM_RECORD_GENERAL,      /* misc record */
 521         HAMMER_MEM_RECORD_INODE,        /* inode record */
 522         HAMMER_MEM_RECORD_ADD,          /* positive memory cache record */
 523         HAMMER_MEM_RECORD_DEL,          /* negative delete-on-disk record */
 524         HAMMER_MEM_RECORD_DATA          /* bulk-data record w/on-disk ref */
 525 } hammer_record_type_t;
 526
 527 struct hammer_record {
 528         RB_ENTRY(hammer_record)         rb_node;
 529         TAILQ_ENTRY(hammer_record)      target_entry;
 530         hammer_inode_state_t            flush_state;
 531         hammer_flush_group_t            flush_group;
 532         hammer_record_type_t            type;
 533         struct hammer_lock              lock;
 534         struct hammer_reserve           *resv;
 535         struct hammer_inode             *ip;
 536         struct hammer_inode             *target_ip;
 537         struct hammer_btree_leaf_elm    leaf;
 538         union hammer_data_ondisk        *data;
 539         int                             flags;
 540         int                             gflags;
 541         hammer_off_t                    zone2_offset;   /* direct-write only */
 542 };
 543
 544 typedef struct hammer_record *hammer_record_t;
 545
 546 /*
 547  * Record flags.  Note that FE can only be set by the frontend if the
 548  * record has not been interlocked by the backend w/ BE.
 549  */
 550 #define HAMMER_RECF_ALLOCDATA           0x0001
 551 #define HAMMER_RECF_ONRBTREE            0x0002
 552 #define HAMMER_RECF_DELETED_FE          0x0004  /* deleted (frontend) */
 553 #define HAMMER_RECF_DELETED_BE          0x0008  /* deleted (backend) */
 554 #define HAMMER_RECF_COMMITTED           0x0010  /* committed to the B-Tree */
 555 #define HAMMER_RECF_INTERLOCK_BE        0x0020  /* backend interlock */
 556 #define HAMMER_RECF_WANTED              0x0040  /* wanted by the frontend */
 557 #define HAMMER_RECF_DEDUPED             0x0080  /* will be live-dedup'ed */
 558 #define HAMMER_RECF_CONVERT_DELETE      0x0100  /* special case */
 559 #define HAMMER_RECF_REDO                0x1000  /* REDO was laid down */
 560
 561 /*
 562  * These flags must be separate to deal with SMP races
 563  */
 564 #define HAMMER_RECG_DIRECT_IO           0x0001  /* related direct I/O running*/
 565 #define HAMMER_RECG_DIRECT_WAIT         0x0002  /* related direct I/O running*/
 566 #define HAMMER_RECG_DIRECT_INVAL        0x0004  /* buffer alias invalidation */
 567 /*
 568  * hammer_create_at_cursor() and hammer_delete_at_cursor() flags.
 569  */
 570 #define HAMMER_CREATE_MODE_UMIRROR      0x0001
 571 #define HAMMER_CREATE_MODE_SYS          0x0002
 572
 573 #define HAMMER_DELETE_ADJUST            0x0001
 574 #define HAMMER_DELETE_DESTROY           0x0002
 575
 576 /*
 577  * In-memory structures representing on-disk structures.
 578  */
 579 struct hammer_volume;
 580 struct hammer_buffer;
 581 struct hammer_node;
 582 struct hammer_undo;
 583 struct hammer_reserve;
 584
 585 RB_HEAD(hammer_vol_rb_tree, hammer_volume);
 586 RB_HEAD(hammer_buf_rb_tree, hammer_buffer);
 587 RB_HEAD(hammer_nod_rb_tree, hammer_node);
 588 RB_HEAD(hammer_und_rb_tree, hammer_undo);
 589 RB_HEAD(hammer_res_rb_tree, hammer_reserve);
 590 RB_HEAD(hammer_mod_rb_tree, hammer_io);
 591
 592 RB_PROTOTYPE2(hammer_vol_rb_tree, hammer_volume, rb_node,
 593               hammer_vol_rb_compare, int32_t);
 594 RB_PROTOTYPE2(hammer_buf_rb_tree, hammer_buffer, rb_node,
 595               hammer_buf_rb_compare, hammer_off_t);
 596 RB_PROTOTYPE2(hammer_nod_rb_tree, hammer_node, rb_node,
 597               hammer_nod_rb_compare, hammer_off_t);
 598 RB_PROTOTYPE2(hammer_und_rb_tree, hammer_undo, rb_node,
 599               hammer_und_rb_compare, hammer_off_t);
 600 RB_PROTOTYPE2(hammer_res_rb_tree, hammer_reserve, rb_node,
 601               hammer_res_rb_compare, hammer_off_t);
 602 RB_PROTOTYPE2(hammer_mod_rb_tree, hammer_io, rb_node,
 603               hammer_mod_rb_compare, hammer_off_t);
 604
 605 /*
 606  * IO management - embedded at the head of various in-memory structures
 607  *
 608  * VOLUME       - hammer_volume containing meta-data
 609  * META_BUFFER  - hammer_buffer containing meta-data
 610  * UNDO_BUFFER  - hammer_buffer containing undo-data
 611  * DATA_BUFFER  - hammer_buffer containing pure-data
 612  * DUMMY        - hammer_buffer not containing valid data
 613  *
 614  * Dirty volume headers and dirty meta-data buffers are locked until the
 615  * flusher can sequence them out.  Dirty pure-data buffers can be written.
 616  * Clean buffers can be passively released.
 617  */
 618 typedef enum hammer_io_type {
 619         HAMMER_STRUCTURE_VOLUME,
 620         HAMMER_STRUCTURE_META_BUFFER,
 621         HAMMER_STRUCTURE_UNDO_BUFFER,
 622         HAMMER_STRUCTURE_DATA_BUFFER,
 623         HAMMER_STRUCTURE_DUMMY
 624 } hammer_io_type_t;
 625
 626 union hammer_io_structure;
 627 struct hammer_io;
 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           validated : 1;  /* ondisk has been validated */
 671         u_int           waitdep : 1;    /* flush waits for dependancies */
 672         u_int           recovered : 1;  /* has recovery ref */
 673         u_int           waitmod : 1;    /* waiting for modify_refs */
 674         u_int           reclaim : 1;    /* reclaim requested */
 675         u_int           gencrc : 1;     /* crc needs to be generated */
 676         u_int           unusedB : 24;
 677 };
 678
 679 typedef struct hammer_io *hammer_io_t;
 680
 681 #define HAMMER_CLUSTER_SIZE     (64 * 1024)
 682 #if HAMMER_CLUSTER_SIZE > MAXBSIZE
 683 #undef  HAMMER_CLUSTER_SIZE
 684 #define HAMMER_CLUSTER_SIZE     MAXBSIZE
 685 #endif
 686 #define HAMMER_CLUSTER_BUFS     (HAMMER_CLUSTER_SIZE / HAMMER_BUFSIZE)
 687
 688 /*
 689  * In-memory volume representing on-disk buffer
 690  */
 691 struct hammer_volume {
 692         struct hammer_io io;
 693         RB_ENTRY(hammer_volume) rb_node;
 694         struct hammer_volume_ondisk *ondisk;
 695         int32_t vol_no;
 696         int64_t nblocks;        /* note: special calculation for statfs */
 697         int64_t buffer_base;    /* base offset of buffer 0 */
 698         hammer_off_t maxbuf_off; /* Maximum buffer offset (zone-2) */
 699         hammer_off_t maxraw_off; /* Maximum raw offset for device */
 700         char    *vol_name;
 701         struct vnode *devvp;
 702         int     vol_flags;
 703 };
 704
 705 typedef struct hammer_volume *hammer_volume_t;
 706
 707 /*
 708  * In-memory buffer (other then volume, super-cluster, or cluster),
 709  * representing an on-disk buffer.
 710  */
 711 struct hammer_buffer {
 712         struct hammer_io io;
 713         RB_ENTRY(hammer_buffer) rb_node;
 714         void *ondisk;
 715         hammer_off_t zoneX_offset;
 716         hammer_off_t zone2_offset;
 717         struct hammer_reserve *resv;
 718         struct hammer_node_list clist;
 719 };
 720
 721 typedef struct hammer_buffer *hammer_buffer_t;
 722
 723 /*
 724  * In-memory B-Tree node, representing an on-disk B-Tree node.
 725  *
 726  * This is a hang-on structure which is backed by a hammer_buffer,
 727  * indexed by a hammer_cluster, and used for fine-grained locking of
 728  * B-Tree nodes in order to properly control lock ordering.  A hammer_buffer
 729  * can contain multiple nodes representing wildly disassociated portions
 730  * of the B-Tree so locking cannot be done on a buffer-by-buffer basis.
 731  *
 732  * This structure uses a cluster-relative index to reduce the number
 733  * of layers required to access it, and also because all on-disk B-Tree
 734  * references are cluster-relative offsets.
 735  */
 736 struct hammer_node {
 737         struct hammer_lock      lock;           /* node-by-node lock */
 738         TAILQ_ENTRY(hammer_node) entry;         /* per-buffer linkage */
 739         RB_ENTRY(hammer_node)   rb_node;        /* per-cluster linkage */
 740         hammer_off_t            node_offset;    /* full offset spec */
 741         struct hammer_mount     *hmp;
 742         struct hammer_buffer    *buffer;        /* backing buffer */
 743         hammer_node_ondisk_t    ondisk;         /* ptr to on-disk structure */
 744         TAILQ_HEAD(, hammer_cursor) cursor_list;  /* deadlock recovery */
 745         struct hammer_node_cache_list cache_list; /* passive caches */
 746         int                     flags;
 747         int                     cursor_exclreq_count;
 748 };
 749
 750 #define HAMMER_NODE_DELETED     0x0001
 751 #define HAMMER_NODE_FLUSH       0x0002
 752 #define HAMMER_NODE_CRCGOOD     0x0004
 753 #define HAMMER_NODE_NEEDSCRC    0x0008
 754 #define HAMMER_NODE_NEEDSMIRROR 0x0010
 755 #define HAMMER_NODE_CRCBAD      0x0020
 756 #define HAMMER_NODE_NONLINEAR   0x0040          /* linear heuristic */
 757
 758 #define HAMMER_NODE_CRCANY      (HAMMER_NODE_CRCGOOD | HAMMER_NODE_CRCBAD)
 759
 760 typedef struct hammer_node      *hammer_node_t;
 761
 762 /*
 763  * List of locked nodes.  This structure is used to lock potentially large
 764  * numbers of nodes as an aid for complex B-Tree operations.
 765  */
 766 struct hammer_node_lock;
 767 TAILQ_HEAD(hammer_node_lock_list, hammer_node_lock);
 768
 769 struct hammer_node_lock {
 770         TAILQ_ENTRY(hammer_node_lock) entry;
 771         struct hammer_node_lock_list  list;
 772         struct hammer_node_lock       *parent;
 773         hammer_node_t   node;
 774         hammer_node_ondisk_t copy;      /* copy of on-disk data */
 775         int             index;          /* index of this node in parent */
 776         int             count;          /* count children */
 777         int             flags;
 778 };
 779
 780 typedef struct hammer_node_lock *hammer_node_lock_t;
 781
 782 #define HAMMER_NODE_LOCK_UPDATED        0x0001
 783 #define HAMMER_NODE_LOCK_LCACHE         0x0002
 784
 785 /*
 786  * Common I/O management structure - embedded in in-memory structures
 787  * which are backed by filesystem buffers.
 788  */
 789 union hammer_io_structure {
 790         struct hammer_io        io;
 791         struct hammer_volume    volume;
 792         struct hammer_buffer    buffer;
 793 };
 794
 795 typedef union hammer_io_structure *hammer_io_structure_t;
 796
 797 /*
 798  * The reserve structure prevents the blockmap from allocating
 799  * out of a reserved big-block.  Such reservations are used by
 800  * the direct-write mechanism.
 801  *
 802  * The structure is also used to hold off on reallocations of
 803  * big-blocks from the freemap until flush dependancies have
 804  * been dealt with.
 805  */
 806 struct hammer_reserve {
 807         RB_ENTRY(hammer_reserve) rb_node;
 808         TAILQ_ENTRY(hammer_reserve) delay_entry;
 809         int             flush_group;
 810         int             flags;
 811         int             refs;
 812         int             zone;
 813         int             append_off;
 814         int32_t         bytes_free;
 815         hammer_off_t    zone_offset;
 816 };
 817
 818 typedef struct hammer_reserve *hammer_reserve_t;
 819
 820 #define HAMMER_RESF_ONDELAY     0x0001
 821 #define HAMMER_RESF_LAYER2FREE  0x0002
 822
 823 #include "hammer_cursor.h"
 824
 825 /*
 826  * The undo structure tracks recent undos to avoid laying down duplicate
 827  * undos within a flush group, saving us a significant amount of overhead.
 828  *
 829  * This is strictly a heuristic.
 830  */
 831 #define HAMMER_MAX_UNDOS                1024
 832 #define HAMMER_MAX_FLUSHERS             4
 833
 834 struct hammer_undo {
 835         RB_ENTRY(hammer_undo)   rb_node;
 836         TAILQ_ENTRY(hammer_undo) lru_entry;
 837         hammer_off_t            offset;
 838         int                     bytes;
 839 };
 840
 841 typedef struct hammer_undo *hammer_undo_t;
 842
 843 struct hammer_flusher_info;
 844 TAILQ_HEAD(hammer_flusher_info_list, hammer_flusher_info);
 845
 846 struct hammer_flusher {
 847         int             signal;         /* flusher thread sequencer */
 848         int             done;           /* last completed flush group */
 849         int             next;           /* next unallocated flg seqno */
 850         int             group_lock;     /* lock sequencing of the next flush */
 851         int             exiting;        /* request master exit */
 852         thread_t        td;             /* master flusher thread */
 853         hammer_tid_t    tid;            /* last flushed transaction id */
 854         int             finalize_want;          /* serialize finalization */
 855         struct hammer_lock finalize_lock;       /* serialize finalization */
 856         struct hammer_transaction trans;        /* shared transaction */
 857         struct hammer_flusher_info_list run_list;
 858         struct hammer_flusher_info_list ready_list;
 859 };
 860
 861 #define HAMMER_FLUSH_UNDOS_RELAXED      0
 862 #define HAMMER_FLUSH_UNDOS_FORCED       1
 863 #define HAMMER_FLUSH_UNDOS_AUTO         2
 864 /*
 865  * Internal hammer mount data structure
 866  */
 867 struct hammer_mount {
 868         struct mount *mp;
 869         /*struct vnode *rootvp;*/
 870         struct hammer_ino_rb_tree rb_inos_root;
 871         struct hammer_redo_rb_tree rb_redo_root;
 872         struct hammer_vol_rb_tree rb_vols_root;
 873         struct hammer_nod_rb_tree rb_nods_root;
 874         struct hammer_und_rb_tree rb_undo_root;
 875         struct hammer_res_rb_tree rb_resv_root;
 876         struct hammer_buf_rb_tree rb_bufs_root;
 877         struct hammer_pfs_rb_tree rb_pfsm_root;
 878
 879         struct hammer_dedup_crc_rb_tree rb_dedup_crc_root;
 880         struct hammer_dedup_off_rb_tree rb_dedup_off_root;
 881
 882         struct hammer_volume *rootvol;
 883         struct hammer_base_elm root_btree_beg;
 884         struct hammer_base_elm root_btree_end;
 885
 886         struct malloc_type      *m_misc;
 887         struct malloc_type      *m_inodes;
 888
 889         int     flags;          /* HAMMER_MOUNT_xxx flags */
 890         int     hflags;
 891         int     ronly;
 892         int     nvolumes;
 893         int     volume_iterator;
 894         int     master_id;      /* -1 or 0-15 - clustering and mirroring */
 895         int     version;        /* hammer filesystem version to use */
 896         int     rsv_inodes;     /* reserved space due to dirty inodes */
 897         int64_t rsv_databytes;  /* reserved space due to record data */
 898         int     rsv_recs;       /* reserved space due to dirty records */
 899         int     rsv_fromdelay;  /* big-blocks reserved due to flush delay */
 900         int     undo_rec_limit; /* based on size of undo area */
 901         int     last_newrecords;
 902         int     count_newrecords;
 903
 904         int     volume_to_remove; /* volume that is currently being removed */
 905
 906         int     count_inodes;   /* total number of inodes */
 907         int     count_iqueued;  /* inodes queued to flusher */
 908         int     count_reclaims; /* inodes pending reclaim by flusher */
 909
 910         struct hammer_flusher flusher;
 911
 912         u_int   check_interrupt;
 913         u_int   check_yield;
 914         uuid_t  fsid;
 915         struct hammer_mod_rb_tree volu_root;    /* dirty undo buffers */
 916         struct hammer_mod_rb_tree undo_root;    /* dirty undo buffers */
 917         struct hammer_mod_rb_tree data_root;    /* dirty data buffers */
 918         struct hammer_mod_rb_tree meta_root;    /* dirty meta bufs    */
 919         struct hammer_mod_rb_tree lose_root;    /* loose buffers      */
 920         long    locked_dirty_space;             /* meta/volu count    */
 921         long    io_running_space;               /* io_token */
 922         int     unused01;
 923         int     objid_cache_count;
 924         int     dedup_cache_count;
 925         int     error;                          /* critical I/O error */
 926         struct krate    krate;                  /* rate limited kprintf */
 927         struct krate    kdiag;                  /* rate limited kprintf */
 928         hammer_tid_t    asof;                   /* snapshot mount */
 929         hammer_tid_t    next_tid;
 930         hammer_tid_t    flush_tid1;             /* flusher tid sequencing */
 931         hammer_tid_t    flush_tid2;             /* flusher tid sequencing */
 932         int64_t copy_stat_freebigblocks;        /* number of free big-blocks */
 933         u_int32_t       undo_seqno;             /* UNDO/REDO FIFO seqno */
 934         u_int32_t       recover_stage2_seqno;   /* REDO recovery seqno */
 935         hammer_off_t    recover_stage2_offset;  /* REDO recovery offset */
 936
 937         struct netexport export;
 938         struct hammer_lock sync_lock;
 939         struct hammer_lock free_lock;
 940         struct hammer_lock undo_lock;
 941         struct hammer_lock blkmap_lock;
 942         struct hammer_lock snapshot_lock;
 943         struct hammer_lock volume_lock;
 944         struct hammer_blockmap  blockmap[HAMMER_MAX_ZONES];
 945         struct hammer_undo      undos[HAMMER_MAX_UNDOS];
 946         int                     undo_alloc;
 947         TAILQ_HEAD(, hammer_undo)  undo_lru_list;
 948         TAILQ_HEAD(, hammer_reserve) delay_list;
 949         struct hammer_flush_group_list  flush_group_list;
 950         hammer_flush_group_t    fill_flush_group;
 951         hammer_flush_group_t    next_flush_group;
 952         TAILQ_HEAD(, hammer_objid_cache) objid_cache_list;
 953         TAILQ_HEAD(, hammer_dedup_cache) dedup_lru_list;
 954         hammer_dedup_cache_t    dedup_free_cache;
 955         TAILQ_HEAD(, hammer_reclaim) reclaim_list;
 956         TAILQ_HEAD(, hammer_io) iorun_list;
 957
 958         struct lwkt_token       fs_token;       /* high level */
 959         struct lwkt_token       io_token;       /* low level (IO callback) */
 960
 961         struct hammer_inostats  inostats[HAMMER_INOSTATS_HSIZE];
 962         uint64_t volume_map[4];  /* 256 bits bitfield */
 963 };
 964
 965 typedef struct hammer_mount     *hammer_mount_t;
 966
 967 #define HAMMER_MOUNT_CRITICAL_ERROR     0x0001
 968 #define HAMMER_MOUNT_FLUSH_RECOVERY     0x0002
 969 #define HAMMER_MOUNT_REDO_SYNC          0x0004
 970 #define HAMMER_MOUNT_REDO_RECOVERY_REQ  0x0008
 971 #define HAMMER_MOUNT_REDO_RECOVERY_RUN  0x0010
 972
 973 #define HAMMER_VOLUME_NUMBER_ADD(hmp, vol)                      \
 974         (hmp)->volume_map[(vol)->vol_no >> 6] |=                \
 975         ((uint64_t)1 << ((vol)->vol_no & ((1 << 6) - 1)))
 976
 977 #define HAMMER_VOLUME_NUMBER_DEL(hmp, vol)                      \
 978         (hmp)->volume_map[(vol)->vol_no >> 6] &=                \
 979         ~((uint64_t)1 << ((vol)->vol_no & ((1 << 6) - 1)))
 980
 981 #define HAMMER_VOLUME_NUMBER_IS_SET(hmp, n)                     \
 982         (((hmp)->volume_map[(n) >> 6] &                         \
 983         ((uint64_t)1 << ((n) & ((1 << 6) - 1)))) != 0)
 984
 985 #define HAMMER_VOLUME_NUMBER_FOREACH(hmp, n)                    \
 986         for (n = 0; n < HAMMER_MAX_VOLUMES; n++)                \
 987                 if (HAMMER_VOLUME_NUMBER_IS_SET(hmp, n))
 988
 989 struct hammer_sync_info {
 990         int error;
 991         int waitfor;
 992 };
 993
 994 /*
 995  * Minium buffer cache bufs required to rebalance the B-Tree.
 996  * This is because we must hold the children and the children's children
 997  * locked.  Even this might not be enough if things are horribly out
 998  * of balance.
 999  */
1000 #define HAMMER_REBALANCE_MIN_BUFS       \
1001         (HAMMER_BTREE_LEAF_ELMS * HAMMER_BTREE_LEAF_ELMS)
1002
1003 #endif  /* _KERNEL || _KERNEL_STRUCTURES */
1004
1005 #if defined(_KERNEL)
1006 /*
1007  * checkspace slop (8MB chunks), higher numbers are more conservative.
1008  */
1009 #define HAMMER_CHKSPC_REBLOCK   25
1010 #define HAMMER_CHKSPC_MIRROR    20
1011 #define HAMMER_CHKSPC_WRITE     20
1012 #define HAMMER_CHKSPC_CREATE    20
1013 #define HAMMER_CHKSPC_REMOVE    10
1014 #define HAMMER_CHKSPC_EMERGENCY 0
1015
1016 extern struct vop_ops hammer_vnode_vops;
1017 extern struct vop_ops hammer_spec_vops;
1018 extern struct vop_ops hammer_fifo_vops;
1019 extern struct bio_ops hammer_bioops;
1020
1021 extern int hammer_debug_io;
1022 extern int hammer_debug_general;
1023 extern int hammer_debug_debug;
1024 extern int hammer_debug_inode;
1025 extern int hammer_debug_locks;
1026 extern int hammer_debug_btree;
1027 extern int hammer_debug_tid;
1028 extern int hammer_debug_recover;
1029 extern int hammer_debug_recover_faults;
1030 extern int hammer_debug_critical;
1031 extern int hammer_cluster_enable;
1032 extern int hammer_live_dedup;
1033 extern int hammer_tdmux_ticks;
1034 extern int hammer_count_fsyncs;
1035 extern int hammer_count_inodes;
1036 extern int hammer_count_iqueued;
1037 extern int hammer_count_reclaims;
1038 extern int hammer_count_records;
1039 extern int hammer_count_record_datas;
1040 extern int hammer_count_volumes;
1041 extern int hammer_count_buffers;
1042 extern int hammer_count_nodes;
1043 extern int64_t hammer_count_extra_space_used;
1044 extern int64_t hammer_stats_btree_lookups;
1045 extern int64_t hammer_stats_btree_searches;
1046 extern int64_t hammer_stats_btree_inserts;
1047 extern int64_t hammer_stats_btree_deletes;
1048 extern int64_t hammer_stats_btree_elements;
1049 extern int64_t hammer_stats_btree_splits;
1050 extern int64_t hammer_stats_btree_iterations;
1051 extern int64_t hammer_stats_btree_root_iterations;
1052 extern int64_t hammer_stats_record_iterations;
1053 extern int64_t hammer_stats_file_read;
1054 extern int64_t hammer_stats_file_write;
1055 extern int64_t hammer_stats_file_iopsr;
1056 extern int64_t hammer_stats_file_iopsw;
1057 extern int64_t hammer_stats_disk_read;
1058 extern int64_t hammer_stats_disk_write;
1059 extern int64_t hammer_stats_inode_flushes;
1060 extern int64_t hammer_stats_commits;
1061 extern int64_t hammer_stats_undo;
1062 extern int64_t hammer_stats_redo;
1063 extern long hammer_count_dirtybufspace;
1064 extern int hammer_count_refedbufs;
1065 extern int hammer_count_reservations;
1066 extern long hammer_count_io_running_read;
1067 extern long hammer_count_io_running_write;
1068 extern int hammer_count_io_locked;
1069 extern long hammer_limit_dirtybufspace;
1070 extern int hammer_limit_recs;
1071 extern int hammer_limit_inode_recs;
1072 extern int hammer_limit_reclaims;
1073 extern int hammer_live_dedup_cache_size;
1074 extern int hammer_limit_redo;
1075 extern int hammer_bio_count;
1076 extern int hammer_verify_zone;
1077 extern int hammer_verify_data;
1078 extern int hammer_write_mode;
1079 extern int hammer_double_buffer;
1080 extern int hammer_btree_full_undo;
1081 extern int hammer_yield_check;
1082 extern int hammer_fsync_mode;
1083 extern int hammer_autoflush;
1084 extern int64_t hammer_contention_count;
1085
1086 extern int64_t hammer_live_dedup_vnode_bcmps;
1087 extern int64_t hammer_live_dedup_device_bcmps;
1088 extern int64_t hammer_live_dedup_findblk_failures;
1089 extern int64_t hammer_live_dedup_bmap_saves;
1090
1091 void    hammer_critical_error(hammer_mount_t hmp, hammer_inode_t ip,
1092                         int error, const char *msg);
1093 int     hammer_vop_inactive(struct vop_inactive_args *);
1094 int     hammer_vop_reclaim(struct vop_reclaim_args *);
1095 int     hammer_get_vnode(struct hammer_inode *ip, struct vnode **vpp);
1096 struct hammer_inode *hammer_get_inode(hammer_transaction_t trans,
1097                         hammer_inode_t dip, int64_t obj_id,
1098                         hammer_tid_t asof, u_int32_t localization,
1099                         int flags, int *errorp);
1100 struct hammer_inode *hammer_get_dummy_inode(hammer_transaction_t trans,
1101                         hammer_inode_t dip, int64_t obj_id,
1102                         hammer_tid_t asof, u_int32_t localization,
1103                         int flags, int *errorp);
1104 struct hammer_inode *hammer_find_inode(hammer_transaction_t trans,
1105                         int64_t obj_id, hammer_tid_t asof,
1106                         u_int32_t localization);
1107 void    hammer_scan_inode_snapshots(hammer_mount_t hmp,
1108                         hammer_inode_info_t iinfo,
1109                         int (*callback)(hammer_inode_t ip, void *data),
1110                         void *data);
1111 void    hammer_put_inode(struct hammer_inode *ip);
1112 void    hammer_put_inode_ref(struct hammer_inode *ip);
1113 void    hammer_inode_waitreclaims(hammer_transaction_t trans);
1114 void    hammer_inode_dirty(struct hammer_inode *ip);
1115
1116 int     hammer_unload_volume(hammer_volume_t volume, void *data __unused);
1117 int     hammer_adjust_volume_mode(hammer_volume_t volume, void *data __unused);
1118
1119 int     hammer_unload_buffer(hammer_buffer_t buffer, void *data);
1120 int     hammer_install_volume(hammer_mount_t hmp, const char *volname,
1121                         struct vnode *devvp);
1122 int     hammer_mountcheck_volumes(hammer_mount_t hmp);
1123
1124 int     hammer_mem_add(hammer_record_t record);
1125 int     hammer_ip_lookup(hammer_cursor_t cursor);
1126 int     hammer_ip_first(hammer_cursor_t cursor);
1127 int     hammer_ip_next(hammer_cursor_t cursor);
1128 int     hammer_ip_resolve_data(hammer_cursor_t cursor);
1129 int     hammer_ip_delete_record(hammer_cursor_t cursor, hammer_inode_t ip,
1130                         hammer_tid_t tid);
1131 int     hammer_create_at_cursor(hammer_cursor_t cursor,
1132                         hammer_btree_leaf_elm_t leaf, void *udata, int mode);
1133 int     hammer_delete_at_cursor(hammer_cursor_t cursor, int delete_flags,
1134                         hammer_tid_t delete_tid, u_int32_t delete_ts,
1135                         int track, int64_t *stat_bytes);
1136 int     hammer_ip_check_directory_empty(hammer_transaction_t trans,
1137                         hammer_inode_t ip);
1138 int     hammer_sync_hmp(hammer_mount_t hmp, int waitfor);
1139 int     hammer_queue_inodes_flusher(hammer_mount_t hmp, int waitfor);
1140
1141 hammer_record_t
1142         hammer_alloc_mem_record(hammer_inode_t ip, int data_len);
1143 void    hammer_flush_record_done(hammer_record_t record, int error);
1144 void    hammer_wait_mem_record_ident(hammer_record_t record, const char *ident);
1145 void    hammer_rel_mem_record(hammer_record_t record);
1146
1147 int     hammer_cursor_up(hammer_cursor_t cursor);
1148 int     hammer_cursor_up_locked(hammer_cursor_t cursor);
1149 int     hammer_cursor_down(hammer_cursor_t cursor);
1150 int     hammer_cursor_upgrade(hammer_cursor_t cursor);
1151 int     hammer_cursor_upgrade_node(hammer_cursor_t cursor);
1152 void    hammer_cursor_downgrade(hammer_cursor_t cursor);
1153 int     hammer_cursor_upgrade2(hammer_cursor_t c1, hammer_cursor_t c2);
1154 void    hammer_cursor_downgrade2(hammer_cursor_t c1, hammer_cursor_t c2);
1155 int     hammer_cursor_seek(hammer_cursor_t cursor, hammer_node_t node,
1156                         int index);
1157 void    hammer_lock_ex_ident(struct hammer_lock *lock, const char *ident);
1158 int     hammer_lock_ex_try(struct hammer_lock *lock);
1159 void    hammer_lock_sh(struct hammer_lock *lock);
1160 int     hammer_lock_sh_try(struct hammer_lock *lock);
1161 int     hammer_lock_upgrade(struct hammer_lock *lock, int shcount);
1162 void    hammer_lock_downgrade(struct hammer_lock *lock, int shcount);
1163 int     hammer_lock_status(struct hammer_lock *lock);
1164 void    hammer_unlock(struct hammer_lock *lock);
1165 void    hammer_ref(struct hammer_lock *lock);
1166 int     hammer_ref_interlock(struct hammer_lock *lock);
1167 int     hammer_ref_interlock_true(struct hammer_lock *lock);
1168 void    hammer_ref_interlock_done(struct hammer_lock *lock);
1169 void    hammer_rel(struct hammer_lock *lock);
1170 int     hammer_rel_interlock(struct hammer_lock *lock, int locked);
1171 void    hammer_rel_interlock_done(struct hammer_lock *lock, int orig_locked);
1172 int     hammer_get_interlock(struct hammer_lock *lock);
1173 int     hammer_try_interlock_norefs(struct hammer_lock *lock);
1174 void    hammer_put_interlock(struct hammer_lock *lock, int error);
1175
1176 void    hammer_sync_lock_ex(hammer_transaction_t trans);
1177 void    hammer_sync_lock_sh(hammer_transaction_t trans);
1178 int     hammer_sync_lock_sh_try(hammer_transaction_t trans);
1179 void    hammer_sync_unlock(hammer_transaction_t trans);
1180
1181 u_int32_t hammer_to_unix_xid(uuid_t *uuid);
1182 void hammer_guid_to_uuid(uuid_t *uuid, u_int32_t guid);
1183 void    hammer_time_to_timespec(u_int64_t xtime, struct timespec *ts);
1184 u_int64_t hammer_timespec_to_time(struct timespec *ts);
1185 int     hammer_str_to_tid(const char *str, int *ispfsp,
1186                         hammer_tid_t *tidp, u_int32_t *localizationp);
1187 int     hammer_is_atatext(const char *name, int len);
1188 hammer_tid_t hammer_alloc_objid(hammer_mount_t hmp, hammer_inode_t dip,
1189                         int64_t namekey);
1190 void hammer_clear_objid(hammer_inode_t dip);
1191 void hammer_destroy_objid_cache(hammer_mount_t hmp);
1192
1193 int hammer_dedup_crc_rb_compare(hammer_dedup_cache_t dc1,
1194                         hammer_dedup_cache_t dc2);
1195 int hammer_dedup_off_rb_compare(hammer_dedup_cache_t dc1,
1196                         hammer_dedup_cache_t dc2);
1197 hammer_dedup_cache_t hammer_dedup_cache_add(hammer_inode_t ip,
1198                         hammer_btree_leaf_elm_t leaf);
1199 hammer_dedup_cache_t hammer_dedup_cache_lookup(hammer_mount_t hmp,
1200                         hammer_crc_t crc);
1201 void hammer_dedup_cache_inval(hammer_mount_t hmp, hammer_off_t base_offset);
1202 void hammer_destroy_dedup_cache(hammer_mount_t hmp);
1203 void hammer_dump_dedup_cache(hammer_mount_t hmp);
1204 int hammer_dedup_validate(hammer_dedup_cache_t dcp, int zone, int bytes,
1205                         void *data);
1206
1207 int hammer_enter_undo_history(hammer_mount_t hmp, hammer_off_t offset,
1208                         int bytes);
1209 void hammer_clear_undo_history(hammer_mount_t hmp);
1210 enum vtype hammer_get_vnode_type(u_int8_t obj_type);
1211 int hammer_get_dtype(u_int8_t obj_type);
1212 u_int8_t hammer_get_obj_type(enum vtype vtype);
1213 int64_t hammer_directory_namekey(hammer_inode_t dip, const void *name, int len,
1214                         u_int32_t *max_iterationsp);
1215 int     hammer_nohistory(hammer_inode_t ip);
1216
1217 int     hammer_init_cursor(hammer_transaction_t trans, hammer_cursor_t cursor,
1218                         hammer_node_cache_t cache, hammer_inode_t ip);
1219 void    hammer_normalize_cursor(hammer_cursor_t cursor);
1220 void    hammer_done_cursor(hammer_cursor_t cursor);
1221 int     hammer_recover_cursor(hammer_cursor_t cursor);
1222 void    hammer_unlock_cursor(hammer_cursor_t cursor);
1223 int     hammer_lock_cursor(hammer_cursor_t cursor);
1224 hammer_cursor_t hammer_push_cursor(hammer_cursor_t ocursor);
1225 void    hammer_pop_cursor(hammer_cursor_t ocursor, hammer_cursor_t ncursor);
1226
1227 void    hammer_cursor_replaced_node(hammer_node_t onode, hammer_node_t nnode);
1228 void    hammer_cursor_removed_node(hammer_node_t onode, hammer_node_t parent,
1229                         int index);
1230 void    hammer_cursor_split_node(hammer_node_t onode, hammer_node_t nnode,
1231                         int index);
1232 void    hammer_cursor_moved_element(hammer_node_t oparent, int pindex,
1233                         hammer_node_t onode, int oindex,
1234                         hammer_node_t nnode, int nindex);
1235 void    hammer_cursor_parent_changed(hammer_node_t node, hammer_node_t oparent,
1236                         hammer_node_t nparent, int nindex);
1237 void    hammer_cursor_inserted_element(hammer_node_t node, int index);
1238 void    hammer_cursor_deleted_element(hammer_node_t node, int index);
1239 void    hammer_cursor_invalidate_cache(hammer_cursor_t cursor);
1240
1241 int     hammer_btree_lookup(hammer_cursor_t cursor);
1242 int     hammer_btree_first(hammer_cursor_t cursor);
1243 int     hammer_btree_last(hammer_cursor_t cursor);
1244 int     hammer_btree_extract(hammer_cursor_t cursor, int flags);
1245 int     hammer_btree_iterate(hammer_cursor_t cursor);
1246 int     hammer_btree_iterate_reverse(hammer_cursor_t cursor);
1247 int     hammer_btree_insert(hammer_cursor_t cursor,
1248                             hammer_btree_leaf_elm_t elm, int *doprop);
1249 int     hammer_btree_delete(hammer_cursor_t cursor, int *ndelete);
1250 void    hammer_btree_do_propagation(hammer_cursor_t cursor,
1251                             hammer_pseudofs_inmem_t pfsm,
1252                             hammer_btree_leaf_elm_t leaf);
1253 int     hammer_btree_cmp(hammer_base_elm_t key1, hammer_base_elm_t key2);
1254 int     hammer_btree_chkts(hammer_tid_t ts, hammer_base_elm_t key);
1255 int     hammer_btree_correct_rhb(hammer_cursor_t cursor, hammer_tid_t tid);
1256 int     hammer_btree_correct_lhb(hammer_cursor_t cursor, hammer_tid_t tid);
1257
1258 int     btree_set_parent(hammer_transaction_t trans, hammer_node_t node,
1259                         hammer_btree_elm_t elm);
1260 void    hammer_node_lock_init(hammer_node_lock_t parent, hammer_node_t node);
1261 void    hammer_btree_lcache_init(hammer_mount_t hmp, hammer_node_lock_t lcache,
1262                         int depth);
1263 void    hammer_btree_lcache_free(hammer_mount_t hmp, hammer_node_lock_t lcache);
1264 int     hammer_btree_lock_children(hammer_cursor_t cursor, int depth,
1265                         hammer_node_lock_t parent,
1266                         hammer_node_lock_t lcache);
1267 void    hammer_btree_lock_copy(hammer_cursor_t cursor,
1268                         hammer_node_lock_t parent);
1269 int     hammer_btree_sync_copy(hammer_cursor_t cursor,
1270                         hammer_node_lock_t parent);
1271 void    hammer_btree_unlock_children(hammer_mount_t hmp,
1272                         hammer_node_lock_t parent,
1273                         hammer_node_lock_t lcache);
1274 int     hammer_btree_search_node(hammer_base_elm_t elm, hammer_node_ondisk_t node);
1275 hammer_node_t hammer_btree_get_parent(hammer_transaction_t trans,
1276                         hammer_node_t node, int *parent_indexp,
1277                         int *errorp, int try_exclusive);
1278
1279 void    hammer_print_btree_node(hammer_node_ondisk_t ondisk);
1280 void    hammer_print_btree_elm(hammer_btree_elm_t elm, u_int8_t type, int i);
1281
1282 void    *hammer_bread(struct hammer_mount *hmp, hammer_off_t off,
1283                         int *errorp, struct hammer_buffer **bufferp);
1284 void    *hammer_bnew(struct hammer_mount *hmp, hammer_off_t off,
1285                         int *errorp, struct hammer_buffer **bufferp);
1286 void    *hammer_bread_ext(struct hammer_mount *hmp, hammer_off_t off, int bytes,
1287                         int *errorp, struct hammer_buffer **bufferp);
1288 void    *hammer_bnew_ext(struct hammer_mount *hmp, hammer_off_t off, int bytes,
1289                         int *errorp, struct hammer_buffer **bufferp);
1290
1291 hammer_volume_t hammer_get_root_volume(hammer_mount_t hmp, int *errorp);
1292
1293 hammer_volume_t hammer_get_volume(hammer_mount_t hmp,
1294                         int32_t vol_no, int *errorp);
1295 hammer_buffer_t hammer_get_buffer(hammer_mount_t hmp, hammer_off_t buf_offset,
1296                         int bytes, int isnew, int *errorp);
1297 void            hammer_sync_buffers(hammer_mount_t hmp,
1298                         hammer_off_t base_offset, int bytes);
1299 int             hammer_del_buffers(hammer_mount_t hmp,
1300                         hammer_off_t base_offset,
1301                         hammer_off_t zone2_offset, int bytes,
1302                         int report_conflicts);
1303
1304 int             hammer_ref_volume(hammer_volume_t volume);
1305 int             hammer_ref_buffer(hammer_buffer_t buffer);
1306 void            hammer_flush_buffer_nodes(hammer_buffer_t buffer);
1307
1308 void            hammer_rel_volume(hammer_volume_t volume, int locked);
1309 void            hammer_rel_buffer(hammer_buffer_t buffer, int locked);
1310
1311 int             hammer_vfs_export(struct mount *mp, int op,
1312                         const struct export_args *export);
1313 hammer_node_t   hammer_get_node(hammer_transaction_t trans,
1314                         hammer_off_t node_offset, int isnew, int *errorp);
1315 void            hammer_ref_node(hammer_node_t node);
1316 hammer_node_t   hammer_ref_node_safe(hammer_transaction_t trans,
1317                         hammer_node_cache_t cache, int *errorp);
1318 void            hammer_rel_node(hammer_node_t node);
1319 void            hammer_delete_node(hammer_transaction_t trans,
1320                         hammer_node_t node);
1321 void            hammer_cache_node(hammer_node_cache_t cache,
1322                         hammer_node_t node);
1323 void            hammer_uncache_node(hammer_node_cache_t cache);
1324 void            hammer_flush_node(hammer_node_t node, int locked);
1325
1326 void hammer_dup_buffer(struct hammer_buffer **bufferp,
1327                         struct hammer_buffer *buffer);
1328 hammer_node_t hammer_alloc_btree(hammer_transaction_t trans,
1329                         hammer_off_t hint, int *errorp);
1330 void *hammer_alloc_data(hammer_transaction_t trans, int32_t data_len,
1331                         u_int16_t rec_type, hammer_off_t *data_offsetp,
1332                         struct hammer_buffer **data_bufferp,
1333                         hammer_off_t hint, int *errorp);
1334
1335 int hammer_generate_undo(hammer_transaction_t trans,
1336                         hammer_off_t zone_offset, void *base, int len);
1337 int hammer_generate_redo(hammer_transaction_t trans, hammer_inode_t ip,
1338                         hammer_off_t file_offset, u_int32_t flags,
1339                         void *base, int len);
1340 void hammer_generate_redo_sync(hammer_transaction_t trans);
1341 void hammer_redo_fifo_start_flush(hammer_inode_t ip);
1342 void hammer_redo_fifo_end_flush(hammer_inode_t ip);
1343
1344 void hammer_format_undo(void *base, u_int32_t seqno);
1345 int hammer_upgrade_undo_4(hammer_transaction_t trans);
1346
1347 void hammer_put_volume(struct hammer_volume *volume, int flush);
1348 void hammer_put_buffer(struct hammer_buffer *buffer, int flush);
1349
1350 hammer_off_t hammer_freemap_alloc(hammer_transaction_t trans,
1351                         hammer_off_t owner, int *errorp);
1352 void hammer_freemap_free(hammer_transaction_t trans, hammer_off_t phys_offset,
1353                         hammer_off_t owner, int *errorp);
1354 int _hammer_checkspace(hammer_mount_t hmp, int slop, int64_t *resp);
1355 hammer_off_t hammer_blockmap_alloc(hammer_transaction_t trans, int zone,
1356                         int bytes, hammer_off_t hint, int *errorp);
1357 hammer_reserve_t hammer_blockmap_reserve(hammer_mount_t hmp, int zone,
1358                         int bytes, hammer_off_t *zone_offp, int *errorp);
1359 hammer_reserve_t hammer_blockmap_reserve_dedup(hammer_mount_t hmp, int zone,
1360                         int bytes, hammer_off_t zone_offset, int *errorp);
1361 void hammer_blockmap_reserve_complete(hammer_mount_t hmp,
1362                         hammer_reserve_t resv);
1363 void hammer_reserve_clrdelay(hammer_mount_t hmp, hammer_reserve_t resv);
1364 void hammer_blockmap_free(hammer_transaction_t trans,
1365                         hammer_off_t zone_offset, int bytes);
1366 int hammer_blockmap_dedup(hammer_transaction_t trans,
1367                         hammer_off_t zone_offset, int bytes);
1368 int hammer_blockmap_finalize(hammer_transaction_t trans,
1369                         hammer_reserve_t resv,
1370                         hammer_off_t zone_offset, int bytes);
1371 int hammer_blockmap_getfree(hammer_mount_t hmp, hammer_off_t zone_offset,
1372                         int *curp, int *errorp);
1373 hammer_off_t hammer_blockmap_lookup_verify(hammer_mount_t hmp,
1374                         hammer_off_t zone_offset, int *errorp);
1375
1376 hammer_off_t hammer_undo_lookup(hammer_mount_t hmp, hammer_off_t zone_offset,
1377                         int *errorp);
1378 int64_t hammer_undo_used(hammer_transaction_t trans);
1379 int64_t hammer_undo_space(hammer_transaction_t trans);
1380 int64_t hammer_undo_max(hammer_mount_t hmp);
1381 int hammer_undo_reclaim(hammer_io_t io);
1382
1383 void hammer_start_transaction(struct hammer_transaction *trans,
1384                               struct hammer_mount *hmp);
1385 void hammer_simple_transaction(struct hammer_transaction *trans,
1386                               struct hammer_mount *hmp);
1387 void hammer_start_transaction_fls(struct hammer_transaction *trans,
1388                                   struct hammer_mount *hmp);
1389 void hammer_done_transaction(struct hammer_transaction *trans);
1390 hammer_tid_t hammer_alloc_tid(hammer_mount_t hmp, int count);
1391
1392 void hammer_modify_inode(hammer_transaction_t trans, hammer_inode_t ip, int flags);
1393 void hammer_flush_inode(hammer_inode_t ip, int flags);
1394 void hammer_flush_inode_done(hammer_inode_t ip, int error);
1395 void hammer_wait_inode(hammer_inode_t ip);
1396
1397 int  hammer_create_inode(struct hammer_transaction *trans, struct vattr *vap,
1398                         struct ucred *cred, struct hammer_inode *dip,
1399                         const char *name, int namelen,
1400                         hammer_pseudofs_inmem_t pfsm,
1401                         struct hammer_inode **ipp);
1402 void hammer_rel_inode(hammer_inode_t ip, int flush);
1403 int hammer_reload_inode(hammer_inode_t ip, void *arg __unused);
1404 int hammer_ino_rb_compare(hammer_inode_t ip1, hammer_inode_t ip2);
1405 int hammer_redo_rb_compare(hammer_inode_t ip1, hammer_inode_t ip2);
1406 int hammer_destroy_inode_callback(hammer_inode_t ip, void *data __unused);
1407
1408 int hammer_sync_inode(hammer_transaction_t trans, hammer_inode_t ip);
1409 void hammer_test_inode(hammer_inode_t dip);
1410 void hammer_inode_unloadable_check(hammer_inode_t ip, int getvp);
1411 int hammer_update_atime_quick(hammer_inode_t ip);
1412
1413 int  hammer_ip_add_directory(struct hammer_transaction *trans,
1414                         hammer_inode_t dip, const char *name, int bytes,
1415                         hammer_inode_t nip);
1416 int  hammer_ip_del_directory(struct hammer_transaction *trans,
1417                         hammer_cursor_t cursor, hammer_inode_t dip,
1418                         hammer_inode_t ip);
1419 void hammer_ip_replace_bulk(hammer_mount_t hmp, hammer_record_t record);
1420 hammer_record_t hammer_ip_add_bulk(hammer_inode_t ip, off_t file_offset,
1421                         void *data, int bytes, int *errorp);
1422 int  hammer_ip_frontend_trunc(struct hammer_inode *ip, off_t file_size);
1423 int  hammer_ip_add_record(struct hammer_transaction *trans,
1424                         hammer_record_t record);
1425 int  hammer_ip_delete_range(hammer_cursor_t cursor, hammer_inode_t ip,
1426                         int64_t ran_beg, int64_t ran_end, int truncating);
1427 int  hammer_ip_delete_clean(hammer_cursor_t cursor, hammer_inode_t ip,
1428                         int *countp);
1429 int  hammer_ip_sync_data(hammer_cursor_t cursor, hammer_inode_t ip,
1430                         int64_t offset, void *data, int bytes);
1431 int  hammer_ip_sync_record(hammer_transaction_t trans, hammer_record_t rec);
1432 int  hammer_ip_sync_record_cursor(hammer_cursor_t cursor, hammer_record_t rec);
1433 hammer_pseudofs_inmem_t  hammer_load_pseudofs(hammer_transaction_t trans,
1434                         u_int32_t localization, int *errorp);
1435 int  hammer_mkroot_pseudofs(hammer_transaction_t trans, struct ucred *cred,
1436                         hammer_pseudofs_inmem_t pfsm);
1437 int  hammer_save_pseudofs(hammer_transaction_t trans,
1438                         hammer_pseudofs_inmem_t pfsm);
1439 int  hammer_unload_pseudofs(hammer_transaction_t trans, u_int32_t localization);
1440 void hammer_rel_pseudofs(hammer_mount_t hmp, hammer_pseudofs_inmem_t pfsm);
1441 int hammer_ioctl(hammer_inode_t ip, u_long com, caddr_t data, int fflag,
1442                         struct ucred *cred);
1443
1444 void hammer_io_init(hammer_io_t io, hammer_volume_t volume,
1445                         enum hammer_io_type type);
1446 int hammer_io_read(struct vnode *devvp, struct hammer_io *io, int limit);
1447 void hammer_io_advance(struct hammer_io *io);
1448 int hammer_io_new(struct vnode *devvp, struct hammer_io *io);
1449 int hammer_io_inval(hammer_volume_t volume, hammer_off_t zone2_offset);
1450 struct buf *hammer_io_release(struct hammer_io *io, int flush);
1451 void hammer_io_flush(struct hammer_io *io, int reclaim);
1452 void hammer_io_wait(struct hammer_io *io);
1453 void hammer_io_waitdep(struct hammer_io *io);
1454 void hammer_io_wait_all(hammer_mount_t hmp, const char *ident, int doflush);
1455 int hammer_io_direct_read(hammer_mount_t hmp, struct bio *bio,
1456                         hammer_btree_leaf_elm_t leaf);
1457 int hammer_io_indirect_read(hammer_mount_t hmp, struct bio *bio,
1458                         hammer_btree_leaf_elm_t leaf);
1459 int hammer_io_direct_write(hammer_mount_t hmp, struct bio *bio,
1460                         hammer_record_t record);
1461 void hammer_io_direct_wait(hammer_record_t record);
1462 void hammer_io_direct_uncache(hammer_mount_t hmp, hammer_btree_leaf_elm_t leaf);
1463 void hammer_io_write_interlock(hammer_io_t io);
1464 void hammer_io_done_interlock(hammer_io_t io);
1465 void hammer_io_clear_modify(struct hammer_io *io, int inval);
1466 void hammer_io_clear_modlist(struct hammer_io *io);
1467 void hammer_io_flush_sync(hammer_mount_t hmp);
1468 void hammer_io_clear_error(struct hammer_io *io);
1469 void hammer_io_clear_error_noassert(struct hammer_io *io);
1470 void hammer_io_notmeta(hammer_buffer_t buffer);
1471 void hammer_io_limit_backlog(hammer_mount_t hmp);
1472
1473 void hammer_modify_volume(hammer_transaction_t trans, hammer_volume_t volume,
1474                         void *base, int len);
1475 void hammer_modify_buffer(hammer_transaction_t trans, hammer_buffer_t buffer,
1476                         void *base, int len);
1477 void hammer_modify_volume_done(hammer_volume_t volume);
1478 void hammer_modify_buffer_done(hammer_buffer_t buffer);
1479
1480 int hammer_ioc_reblock(hammer_transaction_t trans, hammer_inode_t ip,
1481                         struct hammer_ioc_reblock *reblock);
1482 int hammer_ioc_rebalance(hammer_transaction_t trans, hammer_inode_t ip,
1483                         struct hammer_ioc_rebalance *rebal);
1484 int hammer_ioc_prune(hammer_transaction_t trans, hammer_inode_t ip,
1485                         struct hammer_ioc_prune *prune);
1486 int hammer_ioc_mirror_read(hammer_transaction_t trans, hammer_inode_t ip,
1487                         struct hammer_ioc_mirror_rw *mirror);
1488 int hammer_ioc_mirror_write(hammer_transaction_t trans, hammer_inode_t ip,
1489                         struct hammer_ioc_mirror_rw *mirror);
1490 int hammer_ioc_set_pseudofs(hammer_transaction_t trans, hammer_inode_t ip,
1491                         struct ucred *cred, struct hammer_ioc_pseudofs_rw *pfs);
1492 int hammer_ioc_get_pseudofs(hammer_transaction_t trans, hammer_inode_t ip,
1493                         struct hammer_ioc_pseudofs_rw *pfs);
1494 int hammer_ioc_destroy_pseudofs(hammer_transaction_t trans, hammer_inode_t ip,
1495                         struct hammer_ioc_pseudofs_rw *pfs);
1496 int hammer_ioc_downgrade_pseudofs(hammer_transaction_t trans, hammer_inode_t ip,
1497                         struct hammer_ioc_pseudofs_rw *pfs);
1498 int hammer_ioc_upgrade_pseudofs(hammer_transaction_t trans, hammer_inode_t ip,
1499                         struct hammer_ioc_pseudofs_rw *pfs);
1500 int hammer_ioc_wait_pseudofs(hammer_transaction_t trans, hammer_inode_t ip,
1501                         struct hammer_ioc_pseudofs_rw *pfs);
1502 int hammer_ioc_iterate_pseudofs(hammer_transaction_t trans, hammer_inode_t ip,
1503                         struct hammer_ioc_pfs_iterate *pi);
1504 int hammer_ioc_volume_add(hammer_transaction_t trans, hammer_inode_t ip,
1505                         struct hammer_ioc_volume *ioc);
1506 int hammer_ioc_volume_del(hammer_transaction_t trans, hammer_inode_t ip,
1507                         struct hammer_ioc_volume *ioc);
1508 int hammer_ioc_volume_list(hammer_transaction_t trans, hammer_inode_t ip,
1509                         struct hammer_ioc_volume_list *ioc);
1510 int hammer_ioc_dedup(hammer_transaction_t trans, hammer_inode_t ip,
1511                         struct hammer_ioc_dedup *dedup);
1512
1513 int hammer_signal_check(hammer_mount_t hmp);
1514
1515 void hammer_flusher_create(hammer_mount_t hmp);
1516 void hammer_flusher_destroy(hammer_mount_t hmp);
1517 void hammer_flusher_sync(hammer_mount_t hmp);
1518 int  hammer_flusher_async(hammer_mount_t hmp, hammer_flush_group_t flg);
1519 int  hammer_flusher_async_one(hammer_mount_t hmp);
1520 int hammer_flusher_running(hammer_mount_t hmp);
1521 void hammer_flusher_wait(hammer_mount_t hmp, int seq);
1522 void hammer_flusher_wait_next(hammer_mount_t hmp);
1523 int  hammer_flusher_meta_limit(hammer_mount_t hmp);
1524 int  hammer_flusher_meta_halflimit(hammer_mount_t hmp);
1525 int  hammer_flusher_undo_exhausted(hammer_transaction_t trans, int quarter);
1526 void hammer_flusher_clean_loose_ios(hammer_mount_t hmp);
1527 void hammer_flusher_finalize(hammer_transaction_t trans, int final);
1528 int  hammer_flusher_haswork(hammer_mount_t hmp);
1529 void hammer_flusher_flush_undos(hammer_mount_t hmp, int already_flushed);
1530
1531 int hammer_recover_stage1(hammer_mount_t hmp, hammer_volume_t rootvol);
1532 int hammer_recover_stage2(hammer_mount_t hmp, hammer_volume_t rootvol);
1533 void hammer_recover_flush_buffers(hammer_mount_t hmp,
1534                         hammer_volume_t root_volume, int final);
1535
1536 void hammer_crc_set_blockmap(hammer_blockmap_t blockmap);
1537 void hammer_crc_set_volume(hammer_volume_ondisk_t ondisk);
1538 void hammer_crc_set_leaf(void *data, hammer_btree_leaf_elm_t leaf);
1539
1540 int hammer_crc_test_blockmap(hammer_blockmap_t blockmap);
1541 int hammer_crc_test_volume(hammer_volume_ondisk_t ondisk);
1542 int hammer_crc_test_btree(hammer_node_ondisk_t ondisk);
1543 int hammer_crc_test_leaf(void *data, hammer_btree_leaf_elm_t leaf);
1544 void hkprintf(const char *ctl, ...) __printflike(1, 2);
1545 udev_t hammer_fsid_to_udev(uuid_t *uuid);
1546
1547
1548 int hammer_blocksize(int64_t file_offset);
1549 int hammer_blockoff(int64_t file_offset);
1550 int64_t hammer_blockdemarc(int64_t file_offset1, int64_t file_offset2);
1551
1552 /*
1553  * Shortcut for _hammer_checkspace(), used all over the code.
1554  */
1555 static __inline int
1556 hammer_checkspace(hammer_mount_t hmp, int slop)
1557 {
1558         return(_hammer_checkspace(hmp, slop, NULL));
1559 }
1560
1561 static __inline void
1562 hammer_wait_mem_record(hammer_record_t record)
1563 {
1564         hammer_wait_mem_record_ident(record, "hmmwai");
1565 }
1566
1567 static __inline void
1568 hammer_lock_ex(struct hammer_lock *lock)
1569 {
1570         hammer_lock_ex_ident(lock, "hmrlck");
1571 }
1572
1573 static __inline void
1574 hammer_modify_volume_noundo(hammer_transaction_t trans, hammer_volume_t volume)
1575 {
1576         hammer_modify_volume(trans, volume, NULL, 0);
1577 }
1578
1579 static __inline void
1580 hammer_modify_buffer_noundo(hammer_transaction_t trans, hammer_buffer_t buffer)
1581 {
1582         hammer_modify_buffer(trans, buffer, NULL, 0);
1583 }
1584
1585 /*
1586  * Indicate that a B-Tree node is being modified.
1587  */
1588 static __inline void
1589 hammer_modify_node_noundo(hammer_transaction_t trans, hammer_node_t node)
1590 {
1591         KKASSERT((node->flags & HAMMER_NODE_CRCBAD) == 0);
1592         hammer_modify_buffer(trans, node->buffer, NULL, 0);
1593 }
1594
1595 static __inline void
1596 hammer_modify_node_all(hammer_transaction_t trans, struct hammer_node *node)
1597 {
1598         KKASSERT((node->flags & HAMMER_NODE_CRCBAD) == 0);
1599         hammer_modify_buffer(trans, node->buffer,
1600                              node->ondisk, sizeof(*node->ondisk));
1601 }
1602
1603 static __inline void
1604 hammer_modify_node(hammer_transaction_t trans, hammer_node_t node,
1605                    void *base, int len)
1606 {
1607         hammer_crc_t *crcptr;
1608
1609         KKASSERT((char *)base >= (char *)node->ondisk &&
1610                  (char *)base + len <=
1611                     (char *)node->ondisk + sizeof(*node->ondisk));
1612         KKASSERT((node->flags & HAMMER_NODE_CRCBAD) == 0);
1613
1614         if (hammer_btree_full_undo) {
1615                 hammer_modify_node_all(trans, node);
1616         } else {
1617                 hammer_modify_buffer(trans, node->buffer, base, len);
1618                 crcptr = &node->ondisk->crc;
1619                 hammer_modify_buffer(trans, node->buffer,
1620                                      crcptr, sizeof(hammer_crc_t));
1621                 --node->buffer->io.modify_refs; /* only want one ref */
1622         }
1623 }
1624
1625 /*
1626  * Indicate that the specified modifications have been completed.
1627  *
1628  * Do not try to generate the crc here, it's very expensive to do and a
1629  * sequence of insertions or deletions can result in many calls to this
1630  * function on the same node.
1631  */
1632 static __inline void
1633 hammer_modify_node_done(hammer_node_t node)
1634 {
1635         node->flags |= HAMMER_NODE_CRCGOOD;
1636         if ((node->flags & HAMMER_NODE_NEEDSCRC) == 0) {
1637                 node->flags |= HAMMER_NODE_NEEDSCRC;
1638                 node->buffer->io.gencrc = 1;
1639                 hammer_ref_node(node);
1640         }
1641         hammer_modify_buffer_done(node->buffer);
1642 }
1643
1644 /*
1645  * Lookup a blockmap offset.
1646  */
1647 static __inline hammer_off_t
1648 hammer_blockmap_lookup(hammer_mount_t hmp, hammer_off_t zone_offset,
1649                         int *errorp)
1650 {
1651 #if defined INVARIANTS
1652         int zone = HAMMER_ZONE_DECODE(zone_offset);
1653         KKASSERT(zone >= HAMMER_ZONE2_MAPPED_INDEX && zone < HAMMER_MAX_ZONES);
1654 #endif
1655
1656         /*
1657          * We can actually skip blockmap verify by default,
1658          * as normal blockmaps are now direct-mapped onto the freemap
1659          * and so represent zone-2 addresses.
1660          */
1661         if (hammer_verify_zone == 0) {
1662                 *errorp = 0;
1663                 return hammer_xlate_to_zone2(zone_offset);
1664         }
1665
1666         return hammer_blockmap_lookup_verify(hmp, zone_offset, errorp);
1667 }
1668
1669 #define hammer_modify_volume_field(trans, vol, field)           \
1670         hammer_modify_volume(trans, vol, &(vol)->ondisk->field, \
1671                              sizeof((vol)->ondisk->field))
1672
1673 #define hammer_modify_node_field(trans, node, field)            \
1674         hammer_modify_node(trans, node, &(node)->ondisk->field, \
1675                              sizeof((node)->ondisk->field))
1676
1677 /*
1678  * The HAMMER_INODE_CAP_DIR_LOCAL_INO capability is set on newly
1679  * created directories for HAMMER version 2 or greater and causes
1680  * directory entries to be placed the inode localization zone in
1681  * the B-Tree instead of the misc zone.
1682  *
1683  * This greatly improves localization between directory entries and
1684  * inodes
1685  */
1686 static __inline u_int32_t
1687 hammer_dir_localization(hammer_inode_t dip)
1688 {
1689         if (dip->ino_data.cap_flags & HAMMER_INODE_CAP_DIR_LOCAL_INO)
1690                 return(HAMMER_LOCALIZE_INODE);
1691         else
1692                 return(HAMMER_LOCALIZE_MISC);
1693 }
1694 #endif  /* _KERNEL */