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 #include <sys/types.h>
  41 #ifdef _KERNEL
  42 #include <sys/kernel.h>
  43 #endif
  44 #include <sys/conf.h>
  45 #ifdef _KERNEL
  46 #include <sys/systm.h>
  47 #endif
  48 #include <sys/tree.h>
  49 #include <sys/malloc.h>
  50 #include <sys/mount.h>
  51 #include <sys/mountctl.h>
  52 #include <sys/vnode.h>
  53 #include <sys/proc.h>
  54 #include <sys/priv.h>
  55 #include <sys/stat.h>
  56 #include <sys/globaldata.h>
  57 #include <sys/lockf.h>
  58 #include <sys/buf.h>
  59 #include <sys/queue.h>
  60 #include <sys/ktr.h>
  61 #include <sys/limits.h>
  62 #include <vm/vm_extern.h>
  63
  64 #include <sys/buf2.h>
  65 #ifdef _KERNEL
  66 #include <sys/signal2.h>
  67 #include <vm/vm_page2.h>
  68 #endif
  69
  70 #include "hammer_disk.h"
  71 #include "hammer_mount.h"
  72 #include "hammer_ioctl.h"
  73
  74 #if defined(_KERNEL) || defined(_KERNEL_STRUCTURES)
  75
  76 MALLOC_DECLARE(M_HAMMER);
  77
  78 /*
  79  * Kernel trace
  80  */
  81 #if !defined(KTR_HAMMER)
  82 #define KTR_HAMMER      KTR_ALL
  83 #endif
  84 /* KTR_INFO_MASTER_EXTERN(hammer); */
  85
  86 /*
  87  * Misc structures
  88  */
  89 struct hammer_mount;
  90
  91 /*
  92  * Key structure used for custom RB tree inode lookups.  This prototypes
  93  * the function hammer_ino_rb_tree_RB_LOOKUP_INFO(root, info).
  94  */
  95 typedef struct hammer_inode_info {
  96         int64_t         obj_id;         /* (key) object identifier */
  97         hammer_tid_t    obj_asof;       /* (key) snapshot transid or 0 */
  98         u_int32_t       obj_localization; /* (key) pseudo-fs */
  99         union {
 100                 struct hammer_btree_leaf_elm *leaf;
 101         } u;
 102 } *hammer_inode_info_t;
 103
 104 typedef enum hammer_transaction_type {
 105         HAMMER_TRANS_RO,
 106         HAMMER_TRANS_STD,
 107         HAMMER_TRANS_FLS
 108 } hammer_transaction_type_t;
 109
 110 /*
 111  * HAMMER Transaction tracking
 112  */
 113 struct hammer_transaction {
 114         hammer_transaction_type_t type;
 115         struct hammer_mount *hmp;
 116         hammer_tid_t    tid;
 117         u_int64_t       time;
 118         u_int32_t       time32;
 119         int             sync_lock_refs;
 120         int             flags;
 121         struct hammer_volume *rootvol;
 122 };
 123
 124 typedef struct hammer_transaction *hammer_transaction_t;
 125
 126 #define HAMMER_TRANSF_NEWINODE  0x0001
 127 #define HAMMER_TRANSF_DIDIO     0x0002
 128 #define HAMMER_TRANSF_CRCDOM    0x0004  /* EDOM on CRC error, less critical */
 129
 130 /*
 131  * HAMMER locks
 132  */
 133 struct hammer_lock {
 134         volatile u_int  refs;           /* active references */
 135         volatile u_int  lockval;        /* lock count and control bits */
 136         struct thread   *lowner;        /* owner if exclusively held */
 137         struct thread   *rowner;        /* owner if exclusively held */
 138 };
 139
 140 #define HAMMER_REFS_LOCKED      0x40000000      /* transition check */
 141 #define HAMMER_REFS_WANTED      0x20000000      /* transition check */
 142 #define HAMMER_REFS_CHECK       0x10000000      /* transition check */
 143
 144 #define HAMMER_REFS_FLAGS       (HAMMER_REFS_LOCKED | \
 145                                  HAMMER_REFS_WANTED | \
 146                                  HAMMER_REFS_CHECK)
 147
 148 #define HAMMER_LOCKF_EXCLUSIVE  0x40000000
 149 #define HAMMER_LOCKF_WANTED     0x20000000
 150
 151 static __inline int
 152 hammer_notlocked(struct hammer_lock *lock)
 153 {
 154         return(lock->lockval == 0);
 155 }
 156
 157 static __inline int
 158 hammer_islocked(struct hammer_lock *lock)
 159 {
 160         return(lock->lockval != 0);
 161 }
 162
 163 /*
 164  * Returns the number of refs on the object.
 165  */
 166 static __inline int
 167 hammer_isactive(struct hammer_lock *lock)
 168 {
 169         return(lock->refs & ~HAMMER_REFS_FLAGS);
 170 }
 171
 172 static __inline int
 173 hammer_oneref(struct hammer_lock *lock)
 174 {
 175         return((lock->refs & ~HAMMER_REFS_FLAGS) == 1);
 176 }
 177
 178 static __inline int
 179 hammer_norefs(struct hammer_lock *lock)
 180 {
 181         return((lock->refs & ~HAMMER_REFS_FLAGS) == 0);
 182 }
 183
 184 static __inline int
 185 hammer_norefsorlock(struct hammer_lock *lock)
 186 {
 187         return(lock->refs == 0);
 188 }
 189
 190 static __inline int
 191 hammer_refsorlock(struct hammer_lock *lock)
 192 {
 193         return(lock->refs != 0);
 194 }
 195
 196 /*
 197  * Return if we specifically own the lock exclusively.
 198  */
 199 static __inline int
 200 hammer_lock_excl_owned(struct hammer_lock *lock, thread_t td)
 201 {
 202         if ((lock->lockval & HAMMER_LOCKF_EXCLUSIVE) &&
 203             lock->lowner == td) {
 204                 return(1);
 205         }
 206         return(0);
 207 }
 208
 209 /*
 210  * Flush state, used by various structures
 211  */
 212 typedef enum hammer_inode_state {
 213         HAMMER_FST_IDLE,
 214         HAMMER_FST_SETUP,
 215         HAMMER_FST_FLUSH
 216 } hammer_inode_state_t;
 217
 218 TAILQ_HEAD(hammer_record_list, hammer_record);
 219
 220 /*
 221  * Pseudo-filesystem extended data tracking
 222  */
 223 struct hammer_pfs_rb_tree;
 224 struct hammer_pseudofs_inmem;
 225 RB_HEAD(hammer_pfs_rb_tree, hammer_pseudofs_inmem);
 226 RB_PROTOTYPE2(hammer_pfs_rb_tree, hammer_pseudofs_inmem, rb_node,
 227               hammer_pfs_rb_compare, u_int32_t);
 228
 229 struct hammer_pseudofs_inmem {
 230         RB_ENTRY(hammer_pseudofs_inmem) rb_node;
 231         struct hammer_lock      lock;
 232         u_int32_t               localization;
 233         hammer_tid_t            create_tid;
 234         int                     flags;
 235         udev_t                  fsid_udev;
 236         struct hammer_pseudofs_data pfsd;
 237 };
 238
 239 typedef struct hammer_pseudofs_inmem *hammer_pseudofs_inmem_t;
 240
 241 #define HAMMER_PFSM_DELETED     0x0001
 242
 243 /*
 244  * Cache object ids.  A fixed number of objid cache structures are
 245  * created to reserve object id's for newly created files in multiples
 246  * of 100,000, localized to a particular directory, and recycled as
 247  * needed.  This allows parallel create operations in different
 248  * directories to retain fairly localized object ids which in turn
 249  * improves reblocking performance and layout.
 250  */
 251 #define OBJID_CACHE_SIZE        2048
 252 #define OBJID_CACHE_BULK_BITS   10              /* 10 bits (1024)       */
 253 #define OBJID_CACHE_BULK        (32 * 32)       /* two level (1024)     */
 254 #define OBJID_CACHE_BULK_MASK   (OBJID_CACHE_BULK - 1)
 255 #define OBJID_CACHE_BULK_MASK64 ((u_int64_t)(OBJID_CACHE_BULK - 1))
 256
 257 typedef struct hammer_objid_cache {
 258         TAILQ_ENTRY(hammer_objid_cache) entry;
 259         struct hammer_inode             *dip;
 260         hammer_tid_t                    base_tid;
 261         int                             count;
 262         u_int32_t                       bm0;
 263         u_int32_t                       bm1[32];
 264 } *hammer_objid_cache_t;
 265
 266 /*
 267  * Associate an inode with a B-Tree node to cache search start positions
 268  */
 269 typedef struct hammer_node_cache {
 270         TAILQ_ENTRY(hammer_node_cache)  entry;
 271         struct hammer_node              *node;
 272         struct hammer_inode             *ip;
 273 } *hammer_node_cache_t;
 274
 275 TAILQ_HEAD(hammer_node_cache_list, hammer_node_cache);
 276
 277 /*
 278  * Live dedup cache
 279  */
 280 struct hammer_dedup_crc_rb_tree;
 281 RB_HEAD(hammer_dedup_crc_rb_tree, hammer_dedup_cache);
 282 RB_PROTOTYPE2(hammer_dedup_crc_rb_tree, hammer_dedup_cache, crc_entry,
 283                 hammer_dedup_crc_rb_compare, hammer_crc_t);
 284
 285 struct hammer_dedup_off_rb_tree;
 286 RB_HEAD(hammer_dedup_off_rb_tree, hammer_dedup_cache);
 287 RB_PROTOTYPE2(hammer_dedup_off_rb_tree, hammer_dedup_cache, off_entry,
 288                 hammer_dedup_off_rb_compare, hammer_off_t);
 289
 290 #define DEDUP_CACHE_SIZE        4096 /* XXX make it a dynamic tunable */
 291
 292 typedef struct hammer_dedup_cache {
 293         RB_ENTRY(hammer_dedup_cache) crc_entry;
 294         RB_ENTRY(hammer_dedup_cache) off_entry;
 295         TAILQ_ENTRY(hammer_dedup_cache) lru_entry;
 296         struct hammer_mount *hmp;
 297         int64_t obj_id;
 298         u_int32_t localization;
 299         off_t file_offset;
 300         int bytes;
 301         hammer_off_t data_offset;
 302         hammer_crc_t crc;
 303 } *hammer_dedup_cache_t;
 304
 305 /*
 306  * Structure used to organize flush groups.  Flush groups must be
 307  * organized into chunks in order to avoid blowing out the UNDO FIFO.
 308  * Without this a 'sync' could end up flushing 50,000 inodes in a single
 309  * transaction.
 310  */
 311 struct hammer_fls_rb_tree;
 312 RB_HEAD(hammer_fls_rb_tree, hammer_inode);
 313 RB_PROTOTYPE(hammer_fls_rb_tree, hammer_inode, rb_flsnode,
 314               hammer_ino_rb_compare);
 315
 316 struct hammer_flush_group {
 317         TAILQ_ENTRY(hammer_flush_group) flush_entry;
 318         struct hammer_fls_rb_tree       flush_tree;
 319         int                             seq;            /* our seq no */
 320         int                             total_count;    /* record load */
 321         int                             running;        /* group is running */
 322         int                             closed;
 323         int                             refs;
 324 };
 325
 326 typedef struct hammer_flush_group *hammer_flush_group_t;
 327
 328 TAILQ_HEAD(hammer_flush_group_list, hammer_flush_group);
 329
 330 /*
 331  * Structure used to represent an inode in-memory.
 332  *
 333  * The record and data associated with an inode may be out of sync with
 334  * the disk (xDIRTY flags), or not even on the disk at all (ONDISK flag
 335  * clear).
 336  *
 337  * An inode may also hold a cache of unsynchronized records, used for
 338  * database and directories only.  Unsynchronized regular file data is
 339  * stored in the buffer cache.
 340  *
 341  * NOTE: A file which is created and destroyed within the initial
 342  * synchronization period can wind up not doing any disk I/O at all.
 343  *
 344  * Finally, an inode may cache numerous disk-referencing B-Tree cursors.
 345  */
 346 struct hammer_ino_rb_tree;
 347 struct hammer_inode;
 348 RB_HEAD(hammer_ino_rb_tree, hammer_inode);
 349 RB_PROTOTYPEX(hammer_ino_rb_tree, INFO, hammer_inode, rb_node,
 350               hammer_ino_rb_compare, hammer_inode_info_t);
 351
 352 struct hammer_redo_rb_tree;
 353 RB_HEAD(hammer_redo_rb_tree, hammer_inode);
 354 RB_PROTOTYPE2(hammer_redo_rb_tree, hammer_inode, rb_redonode,
 355               hammer_redo_rb_compare, hammer_off_t);
 356
 357 struct hammer_rec_rb_tree;
 358 struct hammer_record;
 359 RB_HEAD(hammer_rec_rb_tree, hammer_record);
 360 RB_PROTOTYPEX(hammer_rec_rb_tree, INFO, hammer_record, rb_node,
 361               hammer_rec_rb_compare, hammer_btree_leaf_elm_t);
 362
 363 TAILQ_HEAD(hammer_node_list, hammer_node);
 364
 365 struct hammer_inode {
 366         RB_ENTRY(hammer_inode)  rb_node;
 367         hammer_inode_state_t    flush_state;
 368         hammer_flush_group_t    flush_group;
 369         RB_ENTRY(hammer_inode)  rb_flsnode;     /* when on flush list */
 370         RB_ENTRY(hammer_inode)  rb_redonode;    /* when INODE_RDIRTY is set */
 371         struct hammer_record_list target_list;  /* target of dependant recs */
 372         int64_t                 obj_id;         /* (key) object identifier */
 373         hammer_tid_t            obj_asof;       /* (key) snapshot or 0 */
 374         u_int32_t               obj_localization; /* (key) pseudo-fs */
 375         struct hammer_mount     *hmp;
 376         hammer_objid_cache_t    objid_cache;
 377         int                     flags;
 378         int                     error;          /* flush error */
 379         int                     cursor_ip_refs; /* sanity */
 380         int                     cursor_exclreq_count;
 381         int                     rsv_recs;
 382         struct vnode            *vp;
 383         hammer_pseudofs_inmem_t pfsm;
 384         struct lockf            advlock;
 385         struct hammer_lock      lock;           /* sync copy interlock */
 386         off_t                   trunc_off;
 387         struct hammer_btree_leaf_elm ino_leaf;  /* in-memory cache */
 388         struct hammer_inode_data ino_data;      /* in-memory cache */
 389         struct hammer_rec_rb_tree rec_tree;     /* in-memory cache */
 390         int                     rec_generation;
 391         struct hammer_node_cache cache[4];      /* search initiate cache */
 392
 393         /*
 394          * When a demark is created to synchronize an inode to
 395          * disk, certain fields are copied so the front-end VOPs
 396          * can continue to run in parallel with the synchronization
 397          * occuring in the background.
 398          */
 399         int             sync_flags;             /* to-sync flags cache */
 400         off_t           sync_trunc_off;         /* to-sync truncation */
 401         off_t           save_trunc_off;         /* write optimization */
 402         struct hammer_btree_leaf_elm sync_ino_leaf; /* to-sync cache */
 403         struct hammer_inode_data sync_ino_data; /* to-sync cache */
 404         size_t          redo_count;
 405
 406         /*
 407          * Track the earliest offset in the UNDO/REDO FIFO containing
 408          * REDO records.  This is staged to the backend during flush
 409          * sequences.  While the inode is staged redo_fifo_next is used
 410          * to track the earliest offset for rotation into redo_fifo_start
 411          * on completion of the flush.
 412          */
 413         hammer_off_t    redo_fifo_start;
 414         hammer_off_t    redo_fifo_next;
 415 };
 416
 417 typedef struct hammer_inode *hammer_inode_t;
 418
 419 #define VTOI(vp)        ((struct hammer_inode *)(vp)->v_data)
 420
 421 /*
 422  * NOTE: DDIRTY does not include atime or mtime and does not include
 423  *       write-append size changes.  SDIRTY handles write-append size
 424  *       changes.
 425  *
 426  *       REDO indicates that REDO logging is active, creating a definitive
 427  *       stream of REDO records in the UNDO/REDO log for writes and
 428  *       truncations, including boundary records when/if REDO is turned off.
 429  *       REDO is typically enabled by fsync() and turned off if excessive
 430  *       writes without an fsync() occurs.
 431  *
 432  *       RDIRTY indicates that REDO records were laid down in the UNDO/REDO
 433  *       FIFO (even if REDO is turned off some might still be active) and
 434  *       still being tracked for this inode.  See hammer_redo.c
 435  */
 436                                         /* (not including atime/mtime) */
 437 #define HAMMER_INODE_DDIRTY     0x0001  /* in-memory ino_data is dirty */
 438 #define HAMMER_INODE_RSV_INODES 0x0002  /* hmp->rsv_inodes bumped */
 439 #define HAMMER_INODE_CONN_DOWN  0x0004  /* include in downward recursion */
 440 #define HAMMER_INODE_XDIRTY     0x0008  /* in-memory records */
 441 #define HAMMER_INODE_ONDISK     0x0010  /* inode is on-disk (else not yet) */
 442 #define HAMMER_INODE_FLUSH      0x0020  /* flush on last ref */
 443 #define HAMMER_INODE_DELETED    0x0080  /* inode delete (backend) */
 444 #define HAMMER_INODE_DELONDISK  0x0100  /* delete synchronized to disk */
 445 #define HAMMER_INODE_RO         0x0200  /* read-only (because of as-of) */
 446 #define HAMMER_INODE_RECSW      0x0400  /* waiting on data record flush */
 447 #define HAMMER_INODE_DONDISK    0x0800  /* data records may be on disk */
 448 #define HAMMER_INODE_BUFS       0x1000  /* dirty high level bps present */
 449 #define HAMMER_INODE_REFLUSH    0x2000  /* flush on dependancy / reflush */
 450 #define HAMMER_INODE_RECLAIM    0x4000  /* trying to reclaim */
 451 #define HAMMER_INODE_FLUSHW     0x8000  /* Someone waiting for flush */
 452
 453 #define HAMMER_INODE_TRUNCATED  0x00010000
 454 #define HAMMER_INODE_DELETING   0x00020000 /* inode delete request (frontend)*/
 455 #define HAMMER_INODE_RESIGNAL   0x00040000 /* re-signal on re-flush */
 456 #define HAMMER_INODE_ATIME      0x00100000 /* in-memory atime modified */
 457 #define HAMMER_INODE_MTIME      0x00200000 /* in-memory mtime modified */
 458 #define HAMMER_INODE_WOULDBLOCK 0x00400000 /* re-issue to new flush group */
 459 #define HAMMER_INODE_DUMMY      0x00800000 /* dummy inode covering bad file */
 460 #define HAMMER_INODE_SDIRTY     0x01000000 /* in-memory ino_data.size is dirty*/
 461 #define HAMMER_INODE_REDO       0x02000000 /* REDO logging active */
 462 #define HAMMER_INODE_RDIRTY     0x04000000 /* REDO records active in fifo */
 463 #define HAMMER_INODE_SLAVEFLUSH 0x08000000 /* being flushed by slave */
 464
 465 #define HAMMER_INODE_MODMASK    (HAMMER_INODE_DDIRTY|HAMMER_INODE_SDIRTY|   \
 466                                  HAMMER_INODE_XDIRTY|HAMMER_INODE_BUFS|     \
 467                                  HAMMER_INODE_ATIME|HAMMER_INODE_MTIME|     \
 468                                  HAMMER_INODE_TRUNCATED|HAMMER_INODE_DELETING)
 469
 470 #define HAMMER_INODE_MODMASK_NOXDIRTY   \
 471                                 (HAMMER_INODE_MODMASK & ~HAMMER_INODE_XDIRTY)
 472
 473 #define HAMMER_INODE_MODMASK_NOREDO     \
 474                                 (HAMMER_INODE_DDIRTY|                       \
 475                                  HAMMER_INODE_XDIRTY|                       \
 476                                  HAMMER_INODE_TRUNCATED|HAMMER_INODE_DELETING)
 477
 478 #define HAMMER_FLUSH_SIGNAL     0x0001
 479 #define HAMMER_FLUSH_RECURSION  0x0002
 480
 481 /*
 482  * Used by the inode reclaim code to pipeline reclaims and avoid
 483  * blowing out kernel memory or letting the flusher get too far
 484  * behind.  The reclaim wakes up when count reaches 0 or the
 485  * timer expires.
 486  */
 487 struct hammer_reclaim {
 488         TAILQ_ENTRY(hammer_reclaim) entry;
 489         int     count;
 490 };
 491
 492 /*
 493  * Track who is creating the greatest burden on the
 494  * inode cache.
 495  */
 496 struct hammer_inostats {
 497         pid_t           pid;    /* track user process */
 498         int             ltick;  /* last tick */
 499         int             count;  /* count (degenerates) */
 500 };
 501
 502 #define HAMMER_INOSTATS_HSIZE   32
 503 #define HAMMER_INOSTATS_HMASK   (HAMMER_INOSTATS_HSIZE - 1)
 504
 505 /*
 506  * Structure used to represent an unsynchronized record in-memory.  These
 507  * records typically represent directory entries.  Only non-historical
 508  * records are kept in-memory.
 509  *
 510  * Records are organized as a per-inode RB-Tree.  If the inode is not
 511  * on disk then neither are any records and the in-memory record tree
 512  * represents the entire contents of the inode.  If the inode is on disk
 513  * then the on-disk B-Tree is scanned in parallel with the in-memory
 514  * RB-Tree to synthesize the current state of the file.
 515  *
 516  * Records are also used to enforce the ordering of directory create/delete
 517  * operations.  A new inode will not be flushed to disk unless its related
 518  * directory entry is also being flushed at the same time.  A directory entry
 519  * will not be removed unless its related inode is also being removed at the
 520  * same time.
 521  */
 522 typedef enum hammer_record_type {
 523         HAMMER_MEM_RECORD_GENERAL,      /* misc record */
 524         HAMMER_MEM_RECORD_INODE,        /* inode record */
 525         HAMMER_MEM_RECORD_ADD,          /* positive memory cache record */
 526         HAMMER_MEM_RECORD_DEL,          /* negative delete-on-disk record */
 527         HAMMER_MEM_RECORD_DATA          /* bulk-data record w/on-disk ref */
 528 } hammer_record_type_t;
 529
 530 struct hammer_record {
 531         RB_ENTRY(hammer_record)         rb_node;
 532         TAILQ_ENTRY(hammer_record)      target_entry;
 533         hammer_inode_state_t            flush_state;
 534         hammer_flush_group_t            flush_group;
 535         hammer_record_type_t            type;
 536         struct hammer_lock              lock;
 537         struct hammer_reserve           *resv;
 538         struct hammer_inode             *ip;
 539         struct hammer_inode             *target_ip;
 540         struct hammer_btree_leaf_elm    leaf;
 541         union hammer_data_ondisk        *data;
 542         int                             flags;
 543         int                             gflags;
 544         hammer_off_t                    zone2_offset;   /* direct-write only */
 545 };
 546
 547 typedef struct hammer_record *hammer_record_t;
 548
 549 /*
 550  * Record flags.  Note that FE can only be set by the frontend if the
 551  * record has not been interlocked by the backend w/ BE.
 552  */
 553 #define HAMMER_RECF_ALLOCDATA           0x0001
 554 #define HAMMER_RECF_ONRBTREE            0x0002
 555 #define HAMMER_RECF_DELETED_FE          0x0004  /* deleted (frontend) */
 556 #define HAMMER_RECF_DELETED_BE          0x0008  /* deleted (backend) */
 557 #define HAMMER_RECF_COMMITTED           0x0010  /* committed to the B-Tree */
 558 #define HAMMER_RECF_INTERLOCK_BE        0x0020  /* backend interlock */
 559 #define HAMMER_RECF_WANTED              0x0040  /* wanted by the frontend */
 560 #define HAMMER_RECF_DEDUPED             0x0080  /* will be live-dedup'ed */
 561 #define HAMMER_RECF_CONVERT_DELETE      0x0100  /* special case */
 562 #define HAMMER_RECF_REDO                0x1000  /* REDO was laid down */
 563
 564 /*
 565  * These flags must be separate to deal with SMP races
 566  */
 567 #define HAMMER_RECG_DIRECT_IO           0x0001  /* related direct I/O running*/
 568 #define HAMMER_RECG_DIRECT_WAIT         0x0002  /* related direct I/O running*/
 569 #define HAMMER_RECG_DIRECT_INVAL        0x0004  /* buffer alias invalidation */
 570 /*
 571  * hammer_create_at_cursor() and hammer_delete_at_cursor() flags.
 572  */
 573 #define HAMMER_CREATE_MODE_UMIRROR      0x0001
 574 #define HAMMER_CREATE_MODE_SYS          0x0002
 575
 576 #define HAMMER_DELETE_ADJUST            0x0001
 577 #define HAMMER_DELETE_DESTROY           0x0002
 578
 579 /*
 580  * In-memory structures representing on-disk structures.
 581  */
 582 struct hammer_volume;
 583 struct hammer_buffer;
 584 struct hammer_node;
 585 struct hammer_undo;
 586 struct hammer_reserve;
 587
 588 RB_HEAD(hammer_vol_rb_tree, hammer_volume);
 589 RB_HEAD(hammer_buf_rb_tree, hammer_buffer);
 590 RB_HEAD(hammer_nod_rb_tree, hammer_node);
 591 RB_HEAD(hammer_und_rb_tree, hammer_undo);
 592 RB_HEAD(hammer_res_rb_tree, hammer_reserve);
 593 RB_HEAD(hammer_mod_rb_tree, hammer_io);
 594
 595 RB_PROTOTYPE2(hammer_vol_rb_tree, hammer_volume, rb_node,
 596               hammer_vol_rb_compare, int32_t);
 597 RB_PROTOTYPE2(hammer_buf_rb_tree, hammer_buffer, rb_node,
 598               hammer_buf_rb_compare, hammer_off_t);
 599 RB_PROTOTYPE2(hammer_nod_rb_tree, hammer_node, rb_node,
 600               hammer_nod_rb_compare, hammer_off_t);
 601 RB_PROTOTYPE2(hammer_und_rb_tree, hammer_undo, rb_node,
 602               hammer_und_rb_compare, hammer_off_t);
 603 RB_PROTOTYPE2(hammer_res_rb_tree, hammer_reserve, rb_node,
 604               hammer_res_rb_compare, hammer_off_t);
 605 RB_PROTOTYPE2(hammer_mod_rb_tree, hammer_io, rb_node,
 606               hammer_mod_rb_compare, hammer_off_t);
 607
 608 /*
 609  * IO management - embedded at the head of various in-memory structures
 610  *
 611  * VOLUME       - hammer_volume containing meta-data
 612  * META_BUFFER  - hammer_buffer containing meta-data
 613  * DATA_BUFFER  - hammer_buffer containing pure-data
 614  *
 615  * Dirty volume headers and dirty meta-data buffers are locked until the
 616  * flusher can sequence them out.  Dirty pure-data buffers can be written.
 617  * Clean buffers can be passively released.
 618  */
 619 typedef enum hammer_io_type {
 620         HAMMER_STRUCTURE_VOLUME,
 621         HAMMER_STRUCTURE_META_BUFFER,
 622         HAMMER_STRUCTURE_UNDO_BUFFER,
 623         HAMMER_STRUCTURE_DATA_BUFFER,
 624         HAMMER_STRUCTURE_DUMMY
 625 } hammer_io_type_t;
 626
 627 union hammer_io_structure;
 628 struct hammer_io;
 629
 630 struct worklist {
 631         LIST_ENTRY(worklist) node;
 632 };
 633
 634 TAILQ_HEAD(hammer_io_list, hammer_io);
 635 typedef struct hammer_io_list *hammer_io_list_t;
 636
 637 struct hammer_io {
 638         struct worklist         worklist;
 639         struct hammer_lock      lock;
 640         enum hammer_io_type     type;
 641         struct hammer_mount     *hmp;
 642         struct hammer_volume    *volume;
 643         RB_ENTRY(hammer_io)     rb_node;     /* if modified */
 644         TAILQ_ENTRY(hammer_io)  iorun_entry; /* iorun_list */
 645         struct hammer_mod_rb_tree *mod_root;
 646         struct buf              *bp;
 647         int64_t                 offset;    /* zone-2 offset */
 648         int                     bytes;     /* buffer cache buffer size */
 649         int                     modify_refs;
 650
 651         /*
 652          * These can be modified at any time by the backend while holding
 653          * io_token, due to bio_done and hammer_io_complete() callbacks.
 654          */
 655         u_int           running : 1;    /* bp write IO in progress */
 656         u_int           waiting : 1;    /* someone is waiting on us */
 657         u_int           ioerror : 1;    /* abort on io-error */
 658         u_int           unusedA : 29;
 659
 660         /*
 661          * These can only be modified by the frontend while holding
 662          * fs_token, or by the backend while holding the io interlocked
 663          * with no references (which will block the frontend when it
 664          * tries to reference it).
 665          *
 666          * WARNING! SMP RACES will create havoc if the callbacks ever tried
 667          *          to modify any of these outside the above restrictions.
 668          */
 669         u_int           modified : 1;   /* bp's data was modified */
 670         u_int           released : 1;   /* bp released (w/ B_LOCKED set) */
 671         u_int           validated : 1;  /* ondisk has been validated */
 672         u_int           waitdep : 1;    /* flush waits for dependancies */
 673         u_int           recovered : 1;  /* has recovery ref */
 674         u_int           waitmod : 1;    /* waiting for modify_refs */
 675         u_int           reclaim : 1;    /* reclaim requested */
 676         u_int           gencrc : 1;     /* crc needs to be generated */
 677         u_int           unusedB : 24;
 678 };
 679
 680 typedef struct hammer_io *hammer_io_t;
 681
 682 #define HAMMER_CLUSTER_SIZE     (64 * 1024)
 683 #if HAMMER_CLUSTER_SIZE > MAXBSIZE
 684 #undef  HAMMER_CLUSTER_SIZE
 685 #define HAMMER_CLUSTER_SIZE     MAXBSIZE
 686 #endif
 687 #define HAMMER_CLUSTER_BUFS     (HAMMER_CLUSTER_SIZE / HAMMER_BUFSIZE)
 688
 689 /*
 690  * In-memory volume representing on-disk buffer
 691  */
 692 struct hammer_volume {
 693         struct hammer_io io;
 694         RB_ENTRY(hammer_volume) rb_node;
 695         struct hammer_volume_ondisk *ondisk;
 696         int32_t vol_no;
 697         int64_t nblocks;        /* note: special calculation for statfs */
 698         int64_t buffer_base;    /* base offset of buffer 0 */
 699         hammer_off_t maxbuf_off; /* Maximum buffer offset (zone-2) */
 700         hammer_off_t maxraw_off; /* Maximum raw offset for device */
 701         char    *vol_name;
 702         struct vnode *devvp;
 703         int     vol_flags;
 704 };
 705
 706 typedef struct hammer_volume *hammer_volume_t;
 707
 708 /*
 709  * In-memory buffer (other then volume, super-cluster, or cluster),
 710  * representing an on-disk buffer.
 711  */
 712 struct hammer_buffer {
 713         struct hammer_io io;
 714         RB_ENTRY(hammer_buffer) rb_node;
 715         void *ondisk;
 716         hammer_off_t zoneX_offset;
 717         hammer_off_t zone2_offset;
 718         struct hammer_reserve *resv;
 719         struct hammer_node_list clist;
 720 };
 721
 722 typedef struct hammer_buffer *hammer_buffer_t;
 723
 724 /*
 725  * In-memory B-Tree node, representing an on-disk B-Tree node.
 726  *
 727  * This is a hang-on structure which is backed by a hammer_buffer,
 728  * indexed by a hammer_cluster, and used for fine-grained locking of
 729  * B-Tree nodes in order to properly control lock ordering.  A hammer_buffer
 730  * can contain multiple nodes representing wildly disassociated portions
 731  * of the B-Tree so locking cannot be done on a buffer-by-buffer basis.
 732  *
 733  * This structure uses a cluster-relative index to reduce the number
 734  * of layers required to access it, and also because all on-disk B-Tree
 735  * references are cluster-relative offsets.
 736  */
 737 struct hammer_node {
 738         struct hammer_lock      lock;           /* node-by-node lock */
 739         TAILQ_ENTRY(hammer_node) entry;         /* per-buffer linkage */
 740         RB_ENTRY(hammer_node)   rb_node;        /* per-cluster linkage */
 741         hammer_off_t            node_offset;    /* full offset spec */
 742         struct hammer_mount     *hmp;
 743         struct hammer_buffer    *buffer;        /* backing buffer */
 744         hammer_node_ondisk_t    ondisk;         /* ptr to on-disk structure */
 745         TAILQ_HEAD(, hammer_cursor) cursor_list;  /* deadlock recovery */
 746         struct hammer_node_cache_list cache_list; /* passive caches */
 747         int                     flags;
 748         int                     cursor_exclreq_count;
 749 };
 750
 751 #define HAMMER_NODE_DELETED     0x0001
 752 #define HAMMER_NODE_FLUSH       0x0002
 753 #define HAMMER_NODE_CRCGOOD     0x0004
 754 #define HAMMER_NODE_NEEDSCRC    0x0008
 755 #define HAMMER_NODE_NEEDSMIRROR 0x0010
 756 #define HAMMER_NODE_CRCBAD      0x0020
 757 #define HAMMER_NODE_NONLINEAR   0x0040          /* linear heuristic */
 758
 759 #define HAMMER_NODE_CRCANY      (HAMMER_NODE_CRCGOOD | HAMMER_NODE_CRCBAD)
 760
 761 typedef struct hammer_node      *hammer_node_t;
 762
 763 /*
 764  * List of locked nodes.  This structure is used to lock potentially large
 765  * numbers of nodes as an aid for complex B-Tree operations.
 766  */
 767 struct hammer_node_lock;
 768 TAILQ_HEAD(hammer_node_lock_list, hammer_node_lock);
 769
 770 struct hammer_node_lock {
 771         TAILQ_ENTRY(hammer_node_lock) entry;
 772         struct hammer_node_lock_list  list;
 773         struct hammer_node_lock       *parent;
 774         hammer_node_t   node;
 775         hammer_node_ondisk_t copy;      /* copy of on-disk data */
 776         int             index;          /* index of this node in parent */
 777         int             count;          /* count children */
 778         int             flags;
 779 };
 780
 781 typedef struct hammer_node_lock *hammer_node_lock_t;
 782
 783 #define HAMMER_NODE_LOCK_UPDATED        0x0001
 784 #define HAMMER_NODE_LOCK_LCACHE         0x0002
 785
 786 /*
 787  * Common I/O management structure - embedded in in-memory structures
 788  * which are backed by filesystem buffers.
 789  */
 790 union hammer_io_structure {
 791         struct hammer_io        io;
 792         struct hammer_volume    volume;
 793         struct hammer_buffer    buffer;
 794 };
 795
 796 typedef union hammer_io_structure *hammer_io_structure_t;
 797
 798 /*
 799  * The reserve structure prevents the blockmap from allocating
 800  * out of a reserved bigblock.  Such reservations are used by
 801  * the direct-write mechanism.
 802  *
 803  * The structure is also used to hold off on reallocations of
 804  * big blocks from the freemap until flush dependancies have
 805  * been dealt with.
 806  */
 807 struct hammer_reserve {
 808         RB_ENTRY(hammer_reserve) rb_node;
 809         TAILQ_ENTRY(hammer_reserve) delay_entry;
 810         int             flush_group;
 811         int             flags;
 812         int             refs;
 813         int             zone;
 814         int             append_off;
 815         int32_t         bytes_free;
 816         hammer_off_t    zone_offset;
 817 };
 818
 819 typedef struct hammer_reserve *hammer_reserve_t;
 820
 821 #define HAMMER_RESF_ONDELAY     0x0001
 822 #define HAMMER_RESF_LAYER2FREE  0x0002
 823
 824 #include "hammer_cursor.h"
 825
 826 /*
 827  * The undo structure tracks recent undos to avoid laying down duplicate
 828  * undos within a flush group, saving us a significant amount of overhead.
 829  *
 830  * This is strictly a heuristic.
 831  */
 832 #define HAMMER_MAX_UNDOS                1024
 833 #define HAMMER_MAX_FLUSHERS             4
 834
 835 struct hammer_undo {
 836         RB_ENTRY(hammer_undo)   rb_node;
 837         TAILQ_ENTRY(hammer_undo) lru_entry;
 838         hammer_off_t            offset;
 839         int                     bytes;
 840 };
 841
 842 typedef struct hammer_undo *hammer_undo_t;
 843
 844 struct hammer_flusher_info;
 845 TAILQ_HEAD(hammer_flusher_info_list, hammer_flusher_info);
 846
 847 struct hammer_flusher {
 848         int             signal;         /* flusher thread sequencer */
 849         int             done;           /* last completed flush group */
 850         int             next;           /* next unallocated flg seqno */
 851         int             group_lock;     /* lock sequencing of the next flush */
 852         int             exiting;        /* request master exit */
 853         thread_t        td;             /* master flusher thread */
 854         hammer_tid_t    tid;            /* last flushed transaction id */
 855         int             finalize_want;          /* serialize finalization */
 856         struct hammer_lock finalize_lock;       /* serialize finalization */
 857         struct hammer_transaction trans;        /* shared transaction */
 858         struct hammer_flusher_info_list run_list;
 859         struct hammer_flusher_info_list ready_list;
 860 };
 861
 862 #define HAMMER_FLUSH_UNDOS_RELAXED      0
 863 #define HAMMER_FLUSH_UNDOS_FORCED       1
 864 #define HAMMER_FLUSH_UNDOS_AUTO         2
 865 /*
 866  * Internal hammer mount data structure
 867  */
 868 struct hammer_mount {
 869         struct mount *mp;
 870         /*struct vnode *rootvp;*/
 871         struct hammer_ino_rb_tree rb_inos_root;
 872         struct hammer_redo_rb_tree rb_redo_root;
 873         struct hammer_vol_rb_tree rb_vols_root;
 874         struct hammer_nod_rb_tree rb_nods_root;
 875         struct hammer_und_rb_tree rb_undo_root;
 876         struct hammer_res_rb_tree rb_resv_root;
 877         struct hammer_buf_rb_tree rb_bufs_root;
 878         struct hammer_pfs_rb_tree rb_pfsm_root;
 879
 880         struct hammer_dedup_crc_rb_tree rb_dedup_crc_root;
 881         struct hammer_dedup_off_rb_tree rb_dedup_off_root;
 882
 883         struct hammer_volume *rootvol;
 884         struct hammer_base_elm root_btree_beg;
 885         struct hammer_base_elm root_btree_end;
 886
 887         struct malloc_type      *m_misc;
 888         struct malloc_type      *m_inodes;
 889
 890         int     flags;          /* HAMMER_MOUNT_xxx flags */
 891         int     hflags;
 892         int     ronly;
 893         int     nvolumes;
 894         int     volume_iterator;
 895         int     master_id;      /* -1 or 0-15 - clustering and mirroring */
 896         int     version;        /* hammer filesystem version to use */
 897         int     rsv_inodes;     /* reserved space due to dirty inodes */
 898         int64_t rsv_databytes;  /* reserved space due to record data */
 899         int     rsv_recs;       /* reserved space due to dirty records */
 900         int     rsv_fromdelay;  /* bigblocks reserved due to flush delay */
 901         int     undo_rec_limit; /* based on size of undo area */
 902         int     last_newrecords;
 903         int     count_newrecords;
 904
 905         int     volume_to_remove; /* volume that is currently being removed */
 906
 907         int     count_inodes;   /* total number of inodes */
 908         int     count_iqueued;  /* inodes queued to flusher */
 909         int     count_reclaims; /* inodes pending reclaim by flusher */
 910
 911         struct hammer_flusher flusher;
 912
 913         u_int   check_interrupt;
 914         u_int   check_yield;
 915         uuid_t  fsid;
 916         struct hammer_mod_rb_tree volu_root;    /* dirty undo buffers */
 917         struct hammer_mod_rb_tree undo_root;    /* dirty undo buffers */
 918         struct hammer_mod_rb_tree data_root;    /* dirty data buffers */
 919         struct hammer_mod_rb_tree meta_root;    /* dirty meta bufs    */
 920         struct hammer_mod_rb_tree lose_root;    /* loose buffers      */
 921         long    locked_dirty_space;             /* meta/volu count    */
 922         long    io_running_space;               /* io_token */
 923         int     unused01;
 924         int     objid_cache_count;
 925         int     dedup_cache_count;
 926         int     error;                          /* critical I/O error */
 927         struct krate    krate;                  /* 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 bigblocks */
 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 };
 963
 964 typedef struct hammer_mount     *hammer_mount_t;
 965
 966 #define HAMMER_MOUNT_CRITICAL_ERROR     0x0001
 967 #define HAMMER_MOUNT_FLUSH_RECOVERY     0x0002
 968 #define HAMMER_MOUNT_REDO_SYNC          0x0004
 969 #define HAMMER_MOUNT_REDO_RECOVERY_REQ  0x0008
 970 #define HAMMER_MOUNT_REDO_RECOVERY_RUN  0x0010
 971
 972 struct hammer_sync_info {
 973         int error;
 974         int waitfor;
 975 };
 976
 977 /*
 978  * Minium buffer cache bufs required to rebalance the B-Tree.
 979  * This is because we must hold the children and the children's children
 980  * locked.  Even this might not be enough if things are horribly out
 981  * of balance.
 982  */
 983 #define HAMMER_REBALANCE_MIN_BUFS       \
 984         (HAMMER_BTREE_LEAF_ELMS * HAMMER_BTREE_LEAF_ELMS)
 985
 986
 987 #endif
 988
 989 /*
 990  * checkspace slop (8MB chunks), higher numbers are more conservative.
 991  */
 992 #define HAMMER_CHKSPC_REBLOCK   25
 993 #define HAMMER_CHKSPC_MIRROR    20
 994 #define HAMMER_CHKSPC_WRITE     20
 995 #define HAMMER_CHKSPC_CREATE    20
 996 #define HAMMER_CHKSPC_REMOVE    10
 997 #define HAMMER_CHKSPC_EMERGENCY 0
 998
 999 #if defined(_KERNEL)
1000
1001 extern struct vop_ops hammer_vnode_vops;
1002 extern struct vop_ops hammer_spec_vops;
1003 extern struct vop_ops hammer_fifo_vops;
1004 extern struct bio_ops hammer_bioops;
1005
1006 extern int hammer_debug_io;
1007 extern int hammer_debug_general;
1008 extern int hammer_debug_debug;
1009 extern int hammer_debug_inode;
1010 extern int hammer_debug_locks;
1011 extern int hammer_debug_btree;
1012 extern int hammer_debug_tid;
1013 extern int hammer_debug_recover;
1014 extern int hammer_debug_recover_faults;
1015 extern int hammer_debug_critical;
1016 extern int hammer_cluster_enable;
1017 extern int hammer_live_dedup;
1018 extern int hammer_tdmux_ticks;
1019 extern int hammer_count_fsyncs;
1020 extern int hammer_count_inodes;
1021 extern int hammer_count_iqueued;
1022 extern int hammer_count_reclaims;
1023 extern int hammer_count_records;
1024 extern int hammer_count_record_datas;
1025 extern int hammer_count_volumes;
1026 extern int hammer_count_buffers;
1027 extern int hammer_count_nodes;
1028 extern int64_t hammer_count_extra_space_used;
1029 extern int64_t hammer_stats_btree_lookups;
1030 extern int64_t hammer_stats_btree_searches;
1031 extern int64_t hammer_stats_btree_inserts;
1032 extern int64_t hammer_stats_btree_deletes;
1033 extern int64_t hammer_stats_btree_elements;
1034 extern int64_t hammer_stats_btree_splits;
1035 extern int64_t hammer_stats_btree_iterations;
1036 extern int64_t hammer_stats_btree_root_iterations;
1037 extern int64_t hammer_stats_record_iterations;
1038 extern int64_t hammer_stats_file_read;
1039 extern int64_t hammer_stats_file_write;
1040 extern int64_t hammer_stats_file_iopsr;
1041 extern int64_t hammer_stats_file_iopsw;
1042 extern int64_t hammer_stats_disk_read;
1043 extern int64_t hammer_stats_disk_write;
1044 extern int64_t hammer_stats_inode_flushes;
1045 extern int64_t hammer_stats_commits;
1046 extern int64_t hammer_stats_undo;
1047 extern int64_t hammer_stats_redo;
1048 extern long hammer_count_dirtybufspace;
1049 extern int hammer_count_refedbufs;
1050 extern int hammer_count_reservations;
1051 extern long hammer_count_io_running_read;
1052 extern long hammer_count_io_running_write;
1053 extern int hammer_count_io_locked;
1054 extern long hammer_limit_dirtybufspace;
1055 extern int hammer_limit_recs;
1056 extern int hammer_limit_inode_recs;
1057 extern int hammer_limit_reclaims;
1058 extern int hammer_live_dedup_cache_size;
1059 extern int hammer_limit_redo;
1060 extern int hammer_bio_count;
1061 extern int hammer_verify_zone;
1062 extern int hammer_verify_data;
1063 extern int hammer_write_mode;
1064 extern int hammer_double_buffer;
1065 extern int hammer_btree_full_undo;
1066 extern int hammer_yield_check;
1067 extern int hammer_fsync_mode;
1068 extern int hammer_autoflush;
1069 extern int64_t hammer_contention_count;
1070
1071 extern int64_t hammer_live_dedup_vnode_bcmps;
1072 extern int64_t hammer_live_dedup_device_bcmps;
1073 extern int64_t hammer_live_dedup_findblk_failures;
1074 extern int64_t hammer_live_dedup_bmap_saves;
1075
1076 void    hammer_critical_error(hammer_mount_t hmp, hammer_inode_t ip,
1077                         int error, const char *msg);
1078 int     hammer_vop_inactive(struct vop_inactive_args *);
1079 int     hammer_vop_reclaim(struct vop_reclaim_args *);
1080 int     hammer_get_vnode(struct hammer_inode *ip, struct vnode **vpp);
1081 struct hammer_inode *hammer_get_inode(hammer_transaction_t trans,
1082                         hammer_inode_t dip, int64_t obj_id,
1083                         hammer_tid_t asof, u_int32_t localization,
1084                         int flags, int *errorp);
1085 struct hammer_inode *hammer_get_dummy_inode(hammer_transaction_t trans,
1086                         hammer_inode_t dip, int64_t obj_id,
1087                         hammer_tid_t asof, u_int32_t localization,
1088                         int flags, int *errorp);
1089 struct hammer_inode *hammer_find_inode(hammer_transaction_t trans,
1090                         int64_t obj_id, hammer_tid_t asof,
1091                         u_int32_t localization);
1092 void    hammer_scan_inode_snapshots(hammer_mount_t hmp,
1093                         hammer_inode_info_t iinfo,
1094                         int (*callback)(hammer_inode_t ip, void *data),
1095                         void *data);
1096 void    hammer_put_inode(struct hammer_inode *ip);
1097 void    hammer_put_inode_ref(struct hammer_inode *ip);
1098 void    hammer_inode_waitreclaims(hammer_transaction_t trans);
1099 void    hammer_inode_dirty(struct hammer_inode *ip);
1100
1101 int     hammer_unload_volume(hammer_volume_t volume, void *data __unused);
1102 int     hammer_adjust_volume_mode(hammer_volume_t volume, void *data __unused);
1103
1104 int     hammer_unload_buffer(hammer_buffer_t buffer, void *data);
1105 int     hammer_install_volume(hammer_mount_t hmp, const char *volname,
1106                         struct vnode *devvp);
1107 int     hammer_mountcheck_volumes(hammer_mount_t hmp);
1108
1109 int     hammer_mem_add(hammer_record_t record);
1110 int     hammer_ip_lookup(hammer_cursor_t cursor);
1111 int     hammer_ip_first(hammer_cursor_t cursor);
1112 int     hammer_ip_next(hammer_cursor_t cursor);
1113 int     hammer_ip_resolve_data(hammer_cursor_t cursor);
1114 int     hammer_ip_delete_record(hammer_cursor_t cursor, hammer_inode_t ip,
1115                         hammer_tid_t tid);
1116 int     hammer_create_at_cursor(hammer_cursor_t cursor,
1117                         hammer_btree_leaf_elm_t leaf, void *udata, int mode);
1118 int     hammer_delete_at_cursor(hammer_cursor_t cursor, int delete_flags,
1119                         hammer_tid_t delete_tid, u_int32_t delete_ts,
1120                         int track, int64_t *stat_bytes);
1121 int     hammer_ip_check_directory_empty(hammer_transaction_t trans,
1122                         hammer_inode_t ip);
1123 int     hammer_sync_hmp(hammer_mount_t hmp, int waitfor);
1124 int     hammer_queue_inodes_flusher(hammer_mount_t hmp, int waitfor);
1125
1126 hammer_record_t
1127         hammer_alloc_mem_record(hammer_inode_t ip, int data_len);
1128 void    hammer_flush_record_done(hammer_record_t record, int error);
1129 void    hammer_wait_mem_record_ident(hammer_record_t record, const char *ident);
1130 void    hammer_rel_mem_record(hammer_record_t record);
1131
1132 int     hammer_cursor_up(hammer_cursor_t cursor);
1133 int     hammer_cursor_up_locked(hammer_cursor_t cursor);
1134 int     hammer_cursor_down(hammer_cursor_t cursor);
1135 int     hammer_cursor_upgrade(hammer_cursor_t cursor);
1136 int     hammer_cursor_upgrade_node(hammer_cursor_t cursor);
1137 void    hammer_cursor_downgrade(hammer_cursor_t cursor);
1138 int     hammer_cursor_upgrade2(hammer_cursor_t c1, hammer_cursor_t c2);
1139 void    hammer_cursor_downgrade2(hammer_cursor_t c1, hammer_cursor_t c2);
1140 int     hammer_cursor_seek(hammer_cursor_t cursor, hammer_node_t node,
1141                         int index);
1142 void    hammer_lock_ex_ident(struct hammer_lock *lock, const char *ident);
1143 int     hammer_lock_ex_try(struct hammer_lock *lock);
1144 void    hammer_lock_sh(struct hammer_lock *lock);
1145 int     hammer_lock_sh_try(struct hammer_lock *lock);
1146 int     hammer_lock_upgrade(struct hammer_lock *lock, int shcount);
1147 void    hammer_lock_downgrade(struct hammer_lock *lock, int shcount);
1148 int     hammer_lock_status(struct hammer_lock *lock);
1149 void    hammer_unlock(struct hammer_lock *lock);
1150 void    hammer_ref(struct hammer_lock *lock);
1151 int     hammer_ref_interlock(struct hammer_lock *lock);
1152 int     hammer_ref_interlock_true(struct hammer_lock *lock);
1153 void    hammer_ref_interlock_done(struct hammer_lock *lock);
1154 void    hammer_rel(struct hammer_lock *lock);
1155 int     hammer_rel_interlock(struct hammer_lock *lock, int locked);
1156 void    hammer_rel_interlock_done(struct hammer_lock *lock, int orig_locked);
1157 int     hammer_get_interlock(struct hammer_lock *lock);
1158 int     hammer_try_interlock_norefs(struct hammer_lock *lock);
1159 void    hammer_put_interlock(struct hammer_lock *lock, int error);
1160
1161 void    hammer_sync_lock_ex(hammer_transaction_t trans);
1162 void    hammer_sync_lock_sh(hammer_transaction_t trans);
1163 int     hammer_sync_lock_sh_try(hammer_transaction_t trans);
1164 void    hammer_sync_unlock(hammer_transaction_t trans);
1165
1166 u_int32_t hammer_to_unix_xid(uuid_t *uuid);
1167 void hammer_guid_to_uuid(uuid_t *uuid, u_int32_t guid);
1168 void    hammer_time_to_timespec(u_int64_t xtime, struct timespec *ts);
1169 u_int64_t hammer_timespec_to_time(struct timespec *ts);
1170 int     hammer_str_to_tid(const char *str, int *ispfsp,
1171                         hammer_tid_t *tidp, u_int32_t *localizationp);
1172 int     hammer_is_atatext(const char *name, int len);
1173 hammer_tid_t hammer_alloc_objid(hammer_mount_t hmp, hammer_inode_t dip,
1174                         int64_t namekey);
1175 void hammer_clear_objid(hammer_inode_t dip);
1176 void hammer_destroy_objid_cache(hammer_mount_t hmp);
1177
1178 int hammer_dedup_crc_rb_compare(hammer_dedup_cache_t dc1,
1179                         hammer_dedup_cache_t dc2);
1180 int hammer_dedup_off_rb_compare(hammer_dedup_cache_t dc1,
1181                         hammer_dedup_cache_t dc2);
1182 hammer_dedup_cache_t hammer_dedup_cache_add(hammer_inode_t ip,
1183                         hammer_btree_leaf_elm_t leaf);
1184 hammer_dedup_cache_t hammer_dedup_cache_lookup(hammer_mount_t hmp,
1185                         hammer_crc_t crc);
1186 void hammer_dedup_cache_inval(hammer_mount_t hmp, hammer_off_t base_offset);
1187 void hammer_destroy_dedup_cache(hammer_mount_t hmp);
1188 void hammer_dump_dedup_cache(hammer_mount_t hmp);
1189 int hammer_dedup_validate(hammer_dedup_cache_t dcp, int zone, int bytes,
1190                         void *data);
1191
1192 int hammer_enter_undo_history(hammer_mount_t hmp, hammer_off_t offset,
1193                         int bytes);
1194 void hammer_clear_undo_history(hammer_mount_t hmp);
1195 enum vtype hammer_get_vnode_type(u_int8_t obj_type);
1196 int hammer_get_dtype(u_int8_t obj_type);
1197 u_int8_t hammer_get_obj_type(enum vtype vtype);
1198 int64_t hammer_directory_namekey(hammer_inode_t dip, const void *name, int len,
1199                         u_int32_t *max_iterationsp);
1200 int     hammer_nohistory(hammer_inode_t ip);
1201
1202 int     hammer_init_cursor(hammer_transaction_t trans, hammer_cursor_t cursor,
1203                         hammer_node_cache_t cache, hammer_inode_t ip);
1204 void    hammer_normalize_cursor(hammer_cursor_t cursor);
1205 void    hammer_done_cursor(hammer_cursor_t cursor);
1206 int     hammer_recover_cursor(hammer_cursor_t cursor);
1207 void    hammer_unlock_cursor(hammer_cursor_t cursor);
1208 int     hammer_lock_cursor(hammer_cursor_t cursor);
1209 hammer_cursor_t hammer_push_cursor(hammer_cursor_t ocursor);
1210 void    hammer_pop_cursor(hammer_cursor_t ocursor, hammer_cursor_t ncursor);
1211
1212 void    hammer_cursor_replaced_node(hammer_node_t onode, hammer_node_t nnode);
1213 void    hammer_cursor_removed_node(hammer_node_t onode, hammer_node_t parent,
1214                         int index);
1215 void    hammer_cursor_split_node(hammer_node_t onode, hammer_node_t nnode,
1216                         int index);
1217 void    hammer_cursor_moved_element(hammer_node_t oparent, int pindex,
1218                         hammer_node_t onode, int oindex,
1219                         hammer_node_t nnode, int nindex);
1220 void    hammer_cursor_parent_changed(hammer_node_t node, hammer_node_t oparent,
1221                         hammer_node_t nparent, int nindex);
1222 void    hammer_cursor_inserted_element(hammer_node_t node, int index);
1223 void    hammer_cursor_deleted_element(hammer_node_t node, int index);
1224 void    hammer_cursor_invalidate_cache(hammer_cursor_t cursor);
1225
1226 int     hammer_btree_lookup(hammer_cursor_t cursor);
1227 int     hammer_btree_first(hammer_cursor_t cursor);
1228 int     hammer_btree_last(hammer_cursor_t cursor);
1229 int     hammer_btree_extract(hammer_cursor_t cursor, int flags);
1230 int     hammer_btree_iterate(hammer_cursor_t cursor);
1231 int     hammer_btree_iterate_reverse(hammer_cursor_t cursor);
1232 int     hammer_btree_insert(hammer_cursor_t cursor,
1233                             hammer_btree_leaf_elm_t elm, int *doprop);
1234 int     hammer_btree_delete(hammer_cursor_t cursor);
1235 void    hammer_btree_do_propagation(hammer_cursor_t cursor,
1236                             hammer_pseudofs_inmem_t pfsm,
1237                             hammer_btree_leaf_elm_t leaf);
1238 int     hammer_btree_cmp(hammer_base_elm_t key1, hammer_base_elm_t key2);
1239 int     hammer_btree_chkts(hammer_tid_t ts, hammer_base_elm_t key);
1240 int     hammer_btree_correct_rhb(hammer_cursor_t cursor, hammer_tid_t tid);
1241 int     hammer_btree_correct_lhb(hammer_cursor_t cursor, hammer_tid_t tid);
1242
1243 int     btree_set_parent(hammer_transaction_t trans, hammer_node_t node,
1244                         hammer_btree_elm_t elm);
1245 void    hammer_node_lock_init(hammer_node_lock_t parent, hammer_node_t node);
1246 void    hammer_btree_lcache_init(hammer_mount_t hmp, hammer_node_lock_t lcache,
1247                         int depth);
1248 void    hammer_btree_lcache_free(hammer_mount_t hmp, hammer_node_lock_t lcache);
1249 int     hammer_btree_lock_children(hammer_cursor_t cursor, int depth,
1250                         hammer_node_lock_t parent,
1251                         hammer_node_lock_t lcache);
1252 void    hammer_btree_lock_copy(hammer_cursor_t cursor,
1253                         hammer_node_lock_t parent);
1254 int     hammer_btree_sync_copy(hammer_cursor_t cursor,
1255                         hammer_node_lock_t parent);
1256 void    hammer_btree_unlock_children(hammer_mount_t hmp,
1257                         hammer_node_lock_t parent,
1258                         hammer_node_lock_t lcache);
1259 int     hammer_btree_search_node(hammer_base_elm_t elm, hammer_node_ondisk_t node);
1260 hammer_node_t hammer_btree_get_parent(hammer_transaction_t trans,
1261                         hammer_node_t node, int *parent_indexp,
1262                         int *errorp, int try_exclusive);
1263
1264 void    hammer_print_btree_node(hammer_node_ondisk_t ondisk);
1265 void    hammer_print_btree_elm(hammer_btree_elm_t elm, u_int8_t type, int i);
1266
1267 void    *hammer_bread(struct hammer_mount *hmp, hammer_off_t off,
1268                         int *errorp, struct hammer_buffer **bufferp);
1269 void    *hammer_bnew(struct hammer_mount *hmp, hammer_off_t off,
1270                         int *errorp, struct hammer_buffer **bufferp);
1271 void    *hammer_bread_ext(struct hammer_mount *hmp, hammer_off_t off, int bytes,
1272                         int *errorp, struct hammer_buffer **bufferp);
1273 void    *hammer_bnew_ext(struct hammer_mount *hmp, hammer_off_t off, int bytes,
1274                         int *errorp, struct hammer_buffer **bufferp);
1275
1276 hammer_volume_t hammer_get_root_volume(hammer_mount_t hmp, int *errorp);
1277
1278 hammer_volume_t hammer_get_volume(hammer_mount_t hmp,
1279                         int32_t vol_no, int *errorp);
1280 hammer_buffer_t hammer_get_buffer(hammer_mount_t hmp, hammer_off_t buf_offset,
1281                         int bytes, int isnew, int *errorp);
1282 void            hammer_sync_buffers(hammer_mount_t hmp,
1283                         hammer_off_t base_offset, int bytes);
1284 int             hammer_del_buffers(hammer_mount_t hmp,
1285                         hammer_off_t base_offset,
1286                         hammer_off_t zone2_offset, int bytes,
1287                         int report_conflicts);
1288
1289 int             hammer_ref_volume(hammer_volume_t volume);
1290 int             hammer_ref_buffer(hammer_buffer_t buffer);
1291 void            hammer_flush_buffer_nodes(hammer_buffer_t buffer);
1292
1293 void            hammer_rel_volume(hammer_volume_t volume, int locked);
1294 void            hammer_rel_buffer(hammer_buffer_t buffer, int locked);
1295
1296 int             hammer_vfs_export(struct mount *mp, int op,
1297                         const struct export_args *export);
1298 hammer_node_t   hammer_get_node(hammer_transaction_t trans,
1299                         hammer_off_t node_offset, int isnew, int *errorp);
1300 void            hammer_ref_node(hammer_node_t node);
1301 hammer_node_t   hammer_ref_node_safe(hammer_transaction_t trans,
1302                         hammer_node_cache_t cache, int *errorp);
1303 void            hammer_rel_node(hammer_node_t node);
1304 void            hammer_delete_node(hammer_transaction_t trans,
1305                         hammer_node_t node);
1306 void            hammer_cache_node(hammer_node_cache_t cache,
1307                         hammer_node_t node);
1308 void            hammer_uncache_node(hammer_node_cache_t cache);
1309 void            hammer_flush_node(hammer_node_t node, int locked);
1310
1311 void hammer_dup_buffer(struct hammer_buffer **bufferp,
1312                         struct hammer_buffer *buffer);
1313 hammer_node_t hammer_alloc_btree(hammer_transaction_t trans,
1314                         hammer_off_t hint, int *errorp);
1315 void *hammer_alloc_data(hammer_transaction_t trans, int32_t data_len,
1316                         u_int16_t rec_type, hammer_off_t *data_offsetp,
1317                         struct hammer_buffer **data_bufferp,
1318                         hammer_off_t hint, int *errorp);
1319
1320 int hammer_generate_undo(hammer_transaction_t trans,
1321                         hammer_off_t zone1_offset, void *base, int len);
1322 int hammer_generate_redo(hammer_transaction_t trans, hammer_inode_t ip,
1323                         hammer_off_t file_offset, u_int32_t flags,
1324                         void *base, int len);
1325 void hammer_generate_redo_sync(hammer_transaction_t trans);
1326 void hammer_redo_fifo_start_flush(hammer_inode_t ip);
1327 void hammer_redo_fifo_end_flush(hammer_inode_t ip);
1328
1329 void hammer_format_undo(void *base, u_int32_t seqno);
1330 int hammer_upgrade_undo_4(hammer_transaction_t trans);
1331
1332 void hammer_put_volume(struct hammer_volume *volume, int flush);
1333 void hammer_put_buffer(struct hammer_buffer *buffer, int flush);
1334
1335 hammer_off_t hammer_freemap_alloc(hammer_transaction_t trans,
1336                         hammer_off_t owner, int *errorp);
1337 void hammer_freemap_free(hammer_transaction_t trans, hammer_off_t phys_offset,
1338                         hammer_off_t owner, int *errorp);
1339 int _hammer_checkspace(hammer_mount_t hmp, int slop, int64_t *resp);
1340 hammer_off_t hammer_blockmap_alloc(hammer_transaction_t trans, int zone,
1341                         int bytes, hammer_off_t hint, int *errorp);
1342 hammer_reserve_t hammer_blockmap_reserve(hammer_mount_t hmp, int zone,
1343                         int bytes, hammer_off_t *zone_offp, int *errorp);
1344 hammer_reserve_t hammer_blockmap_reserve_dedup(hammer_mount_t hmp, int zone,
1345                         int bytes, hammer_off_t zone_offset, int *errorp);
1346 void hammer_blockmap_reserve_complete(hammer_mount_t hmp,
1347                         hammer_reserve_t resv);
1348 void hammer_reserve_clrdelay(hammer_mount_t hmp, hammer_reserve_t resv);
1349 void hammer_blockmap_free(hammer_transaction_t trans,
1350                         hammer_off_t zone_offset, int bytes);
1351 int hammer_blockmap_dedup(hammer_transaction_t trans,
1352                         hammer_off_t zone_offset, int bytes);
1353 int hammer_blockmap_finalize(hammer_transaction_t trans,
1354                         hammer_reserve_t resv,
1355                         hammer_off_t zone_offset, int bytes);
1356 int hammer_blockmap_getfree(hammer_mount_t hmp, hammer_off_t zone_offset,
1357                         int *curp, int *errorp);
1358 hammer_off_t hammer_blockmap_lookup_verify(hammer_mount_t hmp,
1359                         hammer_off_t zone_offset, int *errorp);
1360
1361 hammer_off_t hammer_undo_lookup(hammer_mount_t hmp, hammer_off_t zone_offset,
1362                         int *errorp);
1363 int64_t hammer_undo_used(hammer_transaction_t trans);
1364 int64_t hammer_undo_space(hammer_transaction_t trans);
1365 int64_t hammer_undo_max(hammer_mount_t hmp);
1366 int hammer_undo_reclaim(hammer_io_t io);
1367
1368 void hammer_start_transaction(struct hammer_transaction *trans,
1369                               struct hammer_mount *hmp);
1370 void hammer_simple_transaction(struct hammer_transaction *trans,
1371                               struct hammer_mount *hmp);
1372 void hammer_start_transaction_fls(struct hammer_transaction *trans,
1373                                   struct hammer_mount *hmp);
1374 void hammer_done_transaction(struct hammer_transaction *trans);
1375 hammer_tid_t hammer_alloc_tid(hammer_mount_t hmp, int count);
1376
1377 void hammer_modify_inode(hammer_transaction_t trans, hammer_inode_t ip, int flags);
1378 void hammer_flush_inode(hammer_inode_t ip, int flags);
1379 void hammer_flush_inode_done(hammer_inode_t ip, int error);
1380 void hammer_wait_inode(hammer_inode_t ip);
1381
1382 int  hammer_create_inode(struct hammer_transaction *trans, struct vattr *vap,
1383                         struct ucred *cred, struct hammer_inode *dip,
1384                         const char *name, int namelen,
1385                         hammer_pseudofs_inmem_t pfsm,
1386                         struct hammer_inode **ipp);
1387 void hammer_rel_inode(hammer_inode_t ip, int flush);
1388 int hammer_reload_inode(hammer_inode_t ip, void *arg __unused);
1389 int hammer_ino_rb_compare(hammer_inode_t ip1, hammer_inode_t ip2);
1390 int hammer_redo_rb_compare(hammer_inode_t ip1, hammer_inode_t ip2);
1391 int hammer_destroy_inode_callback(hammer_inode_t ip, void *data __unused);
1392
1393 int hammer_sync_inode(hammer_transaction_t trans, hammer_inode_t ip);
1394 void hammer_test_inode(hammer_inode_t dip);
1395 void hammer_inode_unloadable_check(hammer_inode_t ip, int getvp);
1396 int hammer_update_atime_quick(hammer_inode_t ip);
1397
1398 int  hammer_ip_add_directory(struct hammer_transaction *trans,
1399                         hammer_inode_t dip, const char *name, int bytes,
1400                         hammer_inode_t nip);
1401 int  hammer_ip_del_directory(struct hammer_transaction *trans,
1402                         hammer_cursor_t cursor, hammer_inode_t dip,
1403                         hammer_inode_t ip);
1404 void hammer_ip_replace_bulk(hammer_mount_t hmp, hammer_record_t record);
1405 hammer_record_t hammer_ip_add_bulk(hammer_inode_t ip, off_t file_offset,
1406                         void *data, int bytes, int *errorp);
1407 int  hammer_ip_frontend_trunc(struct hammer_inode *ip, off_t file_size);
1408 int  hammer_ip_add_record(struct hammer_transaction *trans,
1409                         hammer_record_t record);
1410 int  hammer_ip_delete_range(hammer_cursor_t cursor, hammer_inode_t ip,
1411                         int64_t ran_beg, int64_t ran_end, int truncating);
1412 int  hammer_ip_delete_clean(hammer_cursor_t cursor, hammer_inode_t ip,
1413                         int *countp);
1414 int  hammer_ip_sync_data(hammer_cursor_t cursor, hammer_inode_t ip,
1415                         int64_t offset, void *data, int bytes);
1416 int  hammer_ip_sync_record(hammer_transaction_t trans, hammer_record_t rec);
1417 int  hammer_ip_sync_record_cursor(hammer_cursor_t cursor, hammer_record_t rec);
1418 hammer_pseudofs_inmem_t  hammer_load_pseudofs(hammer_transaction_t trans,
1419                         u_int32_t localization, int *errorp);
1420 int  hammer_mkroot_pseudofs(hammer_transaction_t trans, struct ucred *cred,
1421                         hammer_pseudofs_inmem_t pfsm);
1422 int  hammer_save_pseudofs(hammer_transaction_t trans,
1423                         hammer_pseudofs_inmem_t pfsm);
1424 int  hammer_unload_pseudofs(hammer_transaction_t trans, u_int32_t localization);
1425 void hammer_rel_pseudofs(hammer_mount_t hmp, hammer_pseudofs_inmem_t pfsm);
1426 int hammer_ioctl(hammer_inode_t ip, u_long com, caddr_t data, int fflag,
1427                         struct ucred *cred);
1428
1429 void hammer_io_init(hammer_io_t io, hammer_volume_t volume,
1430                         enum hammer_io_type type);
1431 int hammer_io_read(struct vnode *devvp, struct hammer_io *io, int limit);
1432 void hammer_io_advance(struct hammer_io *io);
1433 int hammer_io_new(struct vnode *devvp, struct hammer_io *io);
1434 int hammer_io_inval(hammer_volume_t volume, hammer_off_t zone2_offset);
1435 struct buf *hammer_io_release(struct hammer_io *io, int flush);
1436 void hammer_io_flush(struct hammer_io *io, int reclaim);
1437 void hammer_io_wait(struct hammer_io *io);
1438 void hammer_io_waitdep(struct hammer_io *io);
1439 void hammer_io_wait_all(hammer_mount_t hmp, const char *ident, int doflush);
1440 int hammer_io_direct_read(hammer_mount_t hmp, struct bio *bio,
1441                         hammer_btree_leaf_elm_t leaf);
1442 int hammer_io_indirect_read(hammer_mount_t hmp, struct bio *bio,
1443                         hammer_btree_leaf_elm_t leaf);
1444 int hammer_io_direct_write(hammer_mount_t hmp, struct bio *bio,
1445                         hammer_record_t record);
1446 void hammer_io_direct_wait(hammer_record_t record);
1447 void hammer_io_direct_uncache(hammer_mount_t hmp, hammer_btree_leaf_elm_t leaf);
1448 void hammer_io_write_interlock(hammer_io_t io);
1449 void hammer_io_done_interlock(hammer_io_t io);
1450 void hammer_io_clear_modify(struct hammer_io *io, int inval);
1451 void hammer_io_clear_modlist(struct hammer_io *io);
1452 void hammer_io_flush_sync(hammer_mount_t hmp);
1453 void hammer_io_clear_error(struct hammer_io *io);
1454 void hammer_io_clear_error_noassert(struct hammer_io *io);
1455 void hammer_io_notmeta(hammer_buffer_t buffer);
1456 void hammer_io_limit_backlog(hammer_mount_t hmp);
1457
1458 void hammer_modify_volume(hammer_transaction_t trans, hammer_volume_t volume,
1459                         void *base, int len);
1460 void hammer_modify_buffer(hammer_transaction_t trans, hammer_buffer_t buffer,
1461                         void *base, int len);
1462 void hammer_modify_volume_done(hammer_volume_t volume);
1463 void hammer_modify_buffer_done(hammer_buffer_t buffer);
1464
1465 int hammer_ioc_reblock(hammer_transaction_t trans, hammer_inode_t ip,
1466                         struct hammer_ioc_reblock *reblock);
1467 int hammer_ioc_rebalance(hammer_transaction_t trans, hammer_inode_t ip,
1468                         struct hammer_ioc_rebalance *rebal);
1469 int hammer_ioc_prune(hammer_transaction_t trans, hammer_inode_t ip,
1470                         struct hammer_ioc_prune *prune);
1471 int hammer_ioc_mirror_read(hammer_transaction_t trans, hammer_inode_t ip,
1472                         struct hammer_ioc_mirror_rw *mirror);
1473 int hammer_ioc_mirror_write(hammer_transaction_t trans, hammer_inode_t ip,
1474                         struct hammer_ioc_mirror_rw *mirror);
1475 int hammer_ioc_set_pseudofs(hammer_transaction_t trans, hammer_inode_t ip,
1476                         struct ucred *cred, struct hammer_ioc_pseudofs_rw *pfs);
1477 int hammer_ioc_get_pseudofs(hammer_transaction_t trans, hammer_inode_t ip,
1478                         struct hammer_ioc_pseudofs_rw *pfs);
1479 int hammer_ioc_destroy_pseudofs(hammer_transaction_t trans, hammer_inode_t ip,
1480                         struct hammer_ioc_pseudofs_rw *pfs);
1481 int hammer_ioc_downgrade_pseudofs(hammer_transaction_t trans, hammer_inode_t ip,
1482                         struct hammer_ioc_pseudofs_rw *pfs);
1483 int hammer_ioc_upgrade_pseudofs(hammer_transaction_t trans, hammer_inode_t ip,
1484                         struct hammer_ioc_pseudofs_rw *pfs);
1485 int hammer_ioc_wait_pseudofs(hammer_transaction_t trans, hammer_inode_t ip,
1486                         struct hammer_ioc_pseudofs_rw *pfs);
1487 int hammer_ioc_volume_add(hammer_transaction_t trans, hammer_inode_t ip,
1488                         struct hammer_ioc_volume *ioc);
1489 int hammer_ioc_volume_del(hammer_transaction_t trans, hammer_inode_t ip,
1490                         struct hammer_ioc_volume *ioc);
1491 int hammer_ioc_volume_list(hammer_transaction_t trans, hammer_inode_t ip,
1492                         struct hammer_ioc_volume_list *ioc);
1493 int hammer_ioc_dedup(hammer_transaction_t trans, hammer_inode_t ip,
1494                         struct hammer_ioc_dedup *dedup);
1495
1496 int hammer_signal_check(hammer_mount_t hmp);
1497
1498 void hammer_flusher_create(hammer_mount_t hmp);
1499 void hammer_flusher_destroy(hammer_mount_t hmp);
1500 void hammer_flusher_sync(hammer_mount_t hmp);
1501 int  hammer_flusher_async(hammer_mount_t hmp, hammer_flush_group_t flg);
1502 int  hammer_flusher_async_one(hammer_mount_t hmp);
1503 int hammer_flusher_running(hammer_mount_t hmp);
1504 void hammer_flusher_wait(hammer_mount_t hmp, int seq);
1505 void hammer_flusher_wait_next(hammer_mount_t hmp);
1506 int  hammer_flusher_meta_limit(hammer_mount_t hmp);
1507 int  hammer_flusher_meta_halflimit(hammer_mount_t hmp);
1508 int  hammer_flusher_undo_exhausted(hammer_transaction_t trans, int quarter);
1509 void hammer_flusher_clean_loose_ios(hammer_mount_t hmp);
1510 void hammer_flusher_finalize(hammer_transaction_t trans, int final);
1511 int  hammer_flusher_haswork(hammer_mount_t hmp);
1512 void hammer_flusher_flush_undos(hammer_mount_t hmp, int already_flushed);
1513
1514 int hammer_recover_stage1(hammer_mount_t hmp, hammer_volume_t rootvol);
1515 int hammer_recover_stage2(hammer_mount_t hmp, hammer_volume_t rootvol);
1516 void hammer_recover_flush_buffers(hammer_mount_t hmp,
1517                         hammer_volume_t root_volume, int final);
1518
1519 void hammer_crc_set_blockmap(hammer_blockmap_t blockmap);
1520 void hammer_crc_set_volume(hammer_volume_ondisk_t ondisk);
1521 void hammer_crc_set_leaf(void *data, hammer_btree_leaf_elm_t leaf);
1522
1523 int hammer_crc_test_blockmap(hammer_blockmap_t blockmap);
1524 int hammer_crc_test_volume(hammer_volume_ondisk_t ondisk);
1525 int hammer_crc_test_btree(hammer_node_ondisk_t ondisk);
1526 int hammer_crc_test_leaf(void *data, hammer_btree_leaf_elm_t leaf);
1527 void hkprintf(const char *ctl, ...) __printflike(1, 2);
1528 udev_t hammer_fsid_to_udev(uuid_t *uuid);
1529
1530
1531 int hammer_blocksize(int64_t file_offset);
1532 int hammer_blockoff(int64_t file_offset);
1533 int64_t hammer_blockdemarc(int64_t file_offset1, int64_t file_offset2);
1534
1535 /*
1536  * Shortcut for _hammer_checkspace(), used all over the code.
1537  */
1538 static __inline int
1539 hammer_checkspace(hammer_mount_t hmp, int slop)
1540 {
1541         return(_hammer_checkspace(hmp, slop, NULL));
1542 }
1543
1544 #endif
1545
1546 #ifdef _KERNEL
1547 static __inline void
1548 hammer_wait_mem_record(hammer_record_t record)
1549 {
1550         hammer_wait_mem_record_ident(record, "hmmwai");
1551 }
1552
1553 static __inline void
1554 hammer_lock_ex(struct hammer_lock *lock)
1555 {
1556         hammer_lock_ex_ident(lock, "hmrlck");
1557 }
1558
1559 /*
1560  * Indicate that a B-Tree node is being modified.
1561  */
1562 static __inline void
1563 hammer_modify_node_noundo(hammer_transaction_t trans, hammer_node_t node)
1564 {
1565         KKASSERT((node->flags & HAMMER_NODE_CRCBAD) == 0);
1566         hammer_modify_buffer(trans, node->buffer, NULL, 0);
1567 }
1568
1569 static __inline void
1570 hammer_modify_node_all(hammer_transaction_t trans, struct hammer_node *node)
1571 {
1572         KKASSERT((node->flags & HAMMER_NODE_CRCBAD) == 0);
1573         hammer_modify_buffer(trans, node->buffer,
1574                              node->ondisk, sizeof(*node->ondisk));
1575 }
1576
1577 static __inline void
1578 hammer_modify_node(hammer_transaction_t trans, hammer_node_t node,
1579                    void *base, int len)
1580 {
1581         hammer_crc_t *crcptr;
1582
1583         KKASSERT((char *)base >= (char *)node->ondisk &&
1584                  (char *)base + len <=
1585                     (char *)node->ondisk + sizeof(*node->ondisk));
1586         KKASSERT((node->flags & HAMMER_NODE_CRCBAD) == 0);
1587
1588         if (hammer_btree_full_undo) {
1589                 hammer_modify_node_all(trans, node);
1590         } else {
1591                 hammer_modify_buffer(trans, node->buffer, base, len);
1592                 crcptr = &node->ondisk->crc;
1593                 hammer_modify_buffer(trans, node->buffer,
1594                                      crcptr, sizeof(hammer_crc_t));
1595                 --node->buffer->io.modify_refs; /* only want one ref */
1596         }
1597 }
1598
1599 /*
1600  * Indicate that the specified modifications have been completed.
1601  *
1602  * Do not try to generate the crc here, it's very expensive to do and a
1603  * sequence of insertions or deletions can result in many calls to this
1604  * function on the same node.
1605  */
1606 static __inline void
1607 hammer_modify_node_done(hammer_node_t node)
1608 {
1609         node->flags |= HAMMER_NODE_CRCGOOD;
1610         if ((node->flags & HAMMER_NODE_NEEDSCRC) == 0) {
1611                 node->flags |= HAMMER_NODE_NEEDSCRC;
1612                 node->buffer->io.gencrc = 1;
1613                 hammer_ref_node(node);
1614         }
1615         hammer_modify_buffer_done(node->buffer);
1616 }
1617
1618 /*
1619  * Translate a zone address to zone-2 address.
1620  */
1621 #define hammer_xlate_to_zone2(offset) \
1622         (((offset) & ~HAMMER_OFF_ZONE_MASK) | HAMMER_ZONE_RAW_BUFFER)
1623
1624 /*
1625  * Lookup a blockmap offset.
1626  */
1627 static __inline hammer_off_t
1628 hammer_blockmap_lookup(hammer_mount_t hmp, hammer_off_t zone_offset,
1629                         int *errorp)
1630 {
1631 #if defined INVARIANTS
1632         int zone = HAMMER_ZONE_DECODE(zone_offset);
1633         KKASSERT(zone >= HAMMER_ZONE_BTREE_INDEX && zone < HAMMER_MAX_ZONES);
1634 #endif
1635
1636         /*
1637          * We can actually skip blockmap verify by default,
1638          * as normal blockmaps are now direct-mapped onto the freemap
1639          * and so represent zone-2 addresses.
1640          */
1641         if (hammer_verify_zone == 0) {
1642                 *errorp = 0;
1643                 return hammer_xlate_to_zone2(zone_offset);
1644         }
1645
1646         return hammer_blockmap_lookup_verify(hmp, zone_offset, errorp);
1647 }
1648 #endif
1649
1650 #define hammer_modify_volume_field(trans, vol, field)           \
1651         hammer_modify_volume(trans, vol, &(vol)->ondisk->field, \
1652                              sizeof((vol)->ondisk->field))
1653
1654 #define hammer_modify_node_field(trans, node, field)            \
1655         hammer_modify_node(trans, node, &(node)->ondisk->field, \
1656                              sizeof((node)->ondisk->field))
1657
1658 #ifdef _KERNEL
1659 /*
1660  * The HAMMER_INODE_CAP_DIR_LOCAL_INO capability is set on newly
1661  * created directories for HAMMER version 2 or greater and causes
1662  * directory entries to be placed the inode localization zone in
1663  * the B-Tree instead of the misc zone.
1664  *
1665  * This greatly improves localization between directory entries and
1666  * inodes
1667  */
1668 static __inline u_int32_t
1669 hammer_dir_localization(hammer_inode_t dip)
1670 {
1671         if (dip->ino_data.cap_flags & HAMMER_INODE_CAP_DIR_LOCAL_INO)
1672                 return(HAMMER_LOCALIZE_INODE);
1673         else
1674                 return(HAMMER_LOCALIZE_MISC);
1675 }
1676 #endif