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 #include <sys/kernel.h>
  42 #include <sys/conf.h>
  43 #include <sys/systm.h>
  44 #include <sys/tree.h>
  45 #include <sys/malloc.h>
  46 #include <sys/mount.h>
  47 #include <sys/mountctl.h>
  48 #include <sys/vnode.h>
  49 #include <sys/proc.h>
  50 #include <sys/priv.h>
  51 #include <sys/stat.h>
  52 #include <sys/globaldata.h>
  53 #include <sys/lockf.h>
  54 #include <sys/buf.h>
  55 #include <sys/queue.h>
  56 #include <sys/ktr.h>
  57 #include <sys/globaldata.h>
  58 #include <sys/limits.h>
  59 #include <vm/vm_extern.h>
  60
  61 #include <sys/buf2.h>
  62 #include <sys/signal2.h>
  63 #include <vm/vm_page2.h>
  64
  65 #include "hammer_disk.h"
  66 #include "hammer_mount.h"
  67 #include "hammer_ioctl.h"
  68
  69 #if defined(_KERNEL) || defined(_KERNEL_STRUCTURES)
  70
  71 MALLOC_DECLARE(M_HAMMER);
  72
  73 /*
  74  * Kernel trace
  75  */
  76 #if !defined(KTR_HAMMER)
  77 #define KTR_HAMMER      KTR_ALL
  78 #endif
  79 /* KTR_INFO_MASTER_EXTERN(hammer); */
  80
  81 /*
  82  * Misc structures
  83  */
  84 struct hammer_mount;
  85
  86 /*
  87  * Key structure used for custom RB tree inode lookups.  This prototypes
  88  * the function hammer_ino_rb_tree_RB_LOOKUP_INFO(root, info).
  89  */
  90 typedef struct hammer_inode_info {
  91         int64_t         obj_id;         /* (key) object identifier */
  92         hammer_tid_t    obj_asof;       /* (key) snapshot transid or 0 */
  93         u_int32_t       obj_localization; /* (key) pseudo-fs */
  94         union {
  95                 struct hammer_btree_leaf_elm *leaf;
  96         } u;
  97 } *hammer_inode_info_t;
  98
  99 typedef enum hammer_transaction_type {
 100         HAMMER_TRANS_RO,
 101         HAMMER_TRANS_STD,
 102         HAMMER_TRANS_FLS
 103 } hammer_transaction_type_t;
 104
 105 /*
 106  * HAMMER Transaction tracking
 107  */
 108 struct hammer_transaction {
 109         hammer_transaction_type_t type;
 110         struct hammer_mount *hmp;
 111         hammer_tid_t    tid;
 112         u_int64_t       time;
 113         u_int32_t       time32;
 114         int             sync_lock_refs;
 115         int             flags;
 116         struct hammer_volume *rootvol;
 117 };
 118
 119 typedef struct hammer_transaction *hammer_transaction_t;
 120
 121 #define HAMMER_TRANSF_NEWINODE  0x0001
 122 #define HAMMER_TRANSF_DIDIO     0x0002
 123 #define HAMMER_TRANSF_CRCDOM    0x0004  /* EDOM on CRC error, less critical */
 124
 125 /*
 126  * HAMMER locks
 127  */
 128 struct hammer_lock {
 129         volatile u_int  refs;           /* active references */
 130         volatile u_int  lockval;        /* lock count and control bits */
 131         struct thread   *lowner;        /* owner if exclusively held */
 132         struct thread   *rowner;        /* owner if exclusively held */
 133 };
 134
 135 #define HAMMER_REFS_LOCKED      0x40000000      /* transition check */
 136 #define HAMMER_REFS_WANTED      0x20000000      /* transition check */
 137 #define HAMMER_REFS_CHECK       0x10000000      /* transition check */
 138
 139 #define HAMMER_REFS_FLAGS       (HAMMER_REFS_LOCKED | \
 140                                  HAMMER_REFS_WANTED | \
 141                                  HAMMER_REFS_CHECK)
 142
 143 #define HAMMER_LOCKF_EXCLUSIVE  0x40000000
 144 #define HAMMER_LOCKF_WANTED     0x20000000
 145
 146 static __inline int
 147 hammer_notlocked(struct hammer_lock *lock)
 148 {
 149         return(lock->lockval == 0);
 150 }
 151
 152 static __inline int
 153 hammer_islocked(struct hammer_lock *lock)
 154 {
 155         return(lock->lockval != 0);
 156 }
 157
 158 /*
 159  * Returns the number of refs on the object.
 160  */
 161 static __inline int
 162 hammer_isactive(struct hammer_lock *lock)
 163 {
 164         return(lock->refs & ~HAMMER_REFS_FLAGS);
 165 }
 166
 167 static __inline int
 168 hammer_oneref(struct hammer_lock *lock)
 169 {
 170         return((lock->refs & ~HAMMER_REFS_FLAGS) == 1);
 171 }
 172
 173 static __inline int
 174 hammer_norefs(struct hammer_lock *lock)
 175 {
 176         return((lock->refs & ~HAMMER_REFS_FLAGS) == 0);
 177 }
 178
 179 static __inline int
 180 hammer_norefsorlock(struct hammer_lock *lock)
 181 {
 182         return(lock->refs == 0);
 183 }
 184
 185 static __inline int
 186 hammer_refsorlock(struct hammer_lock *lock)
 187 {
 188         return(lock->refs != 0);
 189 }
 190
 191 /*
 192  * Return if we specifically own the lock exclusively.
 193  */
 194 static __inline int
 195 hammer_lock_excl_owned(struct hammer_lock *lock, thread_t td)
 196 {
 197         if ((lock->lockval & HAMMER_LOCKF_EXCLUSIVE) &&
 198             lock->lowner == td) {
 199                 return(1);
 200         }
 201         return(0);
 202 }
 203
 204 /*
 205  * Flush state, used by various structures
 206  */
 207 typedef enum hammer_inode_state {
 208         HAMMER_FST_IDLE,
 209         HAMMER_FST_SETUP,
 210         HAMMER_FST_FLUSH
 211 } hammer_inode_state_t;
 212
 213 TAILQ_HEAD(hammer_record_list, hammer_record);
 214
 215 /*
 216  * Pseudo-filesystem extended data tracking
 217  */
 218 struct hammer_pfs_rb_tree;
 219 struct hammer_pseudofs_inmem;
 220 RB_HEAD(hammer_pfs_rb_tree, hammer_pseudofs_inmem);
 221 RB_PROTOTYPE2(hammer_pfs_rb_tree, hammer_pseudofs_inmem, rb_node,
 222               hammer_pfs_rb_compare, u_int32_t);
 223
 224 struct hammer_pseudofs_inmem {
 225         RB_ENTRY(hammer_pseudofs_inmem) rb_node;
 226         struct hammer_lock      lock;
 227         u_int32_t               localization;
 228         hammer_tid_t            create_tid;
 229         int                     flags;
 230         udev_t                  fsid_udev;
 231         struct hammer_pseudofs_data pfsd;
 232 };
 233
 234 typedef struct hammer_pseudofs_inmem *hammer_pseudofs_inmem_t;
 235
 236 #define HAMMER_PFSM_DELETED     0x0001
 237
 238 /*
 239  * Cache object ids.  A fixed number of objid cache structures are
 240  * created to reserve object id's for newly created files in multiples
 241  * of 100,000, localized to a particular directory, and recycled as
 242  * needed.  This allows parallel create operations in different
 243  * directories to retain fairly localized object ids which in turn
 244  * improves reblocking performance and layout.
 245  */
 246 #define OBJID_CACHE_SIZE        2048
 247 #define OBJID_CACHE_BULK_BITS   10              /* 10 bits (1024)       */
 248 #define OBJID_CACHE_BULK        (32 * 32)       /* two level (1024)     */
 249 #define OBJID_CACHE_BULK_MASK   (OBJID_CACHE_BULK - 1)
 250 #define OBJID_CACHE_BULK_MASK64 ((u_int64_t)(OBJID_CACHE_BULK - 1))
 251
 252 typedef struct hammer_objid_cache {
 253         TAILQ_ENTRY(hammer_objid_cache) entry;
 254         struct hammer_inode             *dip;
 255         hammer_tid_t                    base_tid;
 256         int                             count;
 257         u_int32_t                       bm0;
 258         u_int32_t                       bm1[32];
 259 } *hammer_objid_cache_t;
 260
 261 /*
 262  * Associate an inode with a B-Tree node to cache search start positions
 263  */
 264 typedef struct hammer_node_cache {
 265         TAILQ_ENTRY(hammer_node_cache)  entry;
 266         struct hammer_node              *node;
 267         struct hammer_inode             *ip;
 268 } *hammer_node_cache_t;
 269
 270 TAILQ_HEAD(hammer_node_cache_list, hammer_node_cache);
 271
 272 /*
 273  * Live dedup cache
 274  */
 275 struct hammer_dedup_crc_rb_tree;
 276 RB_HEAD(hammer_dedup_crc_rb_tree, hammer_dedup_cache);
 277 RB_PROTOTYPE2(hammer_dedup_crc_rb_tree, hammer_dedup_cache, crc_entry,
 278                 hammer_dedup_crc_rb_compare, hammer_crc_t);
 279
 280 struct hammer_dedup_off_rb_tree;
 281 RB_HEAD(hammer_dedup_off_rb_tree, hammer_dedup_cache);
 282 RB_PROTOTYPE2(hammer_dedup_off_rb_tree, hammer_dedup_cache, off_entry,
 283                 hammer_dedup_off_rb_compare, hammer_off_t);
 284
 285 #define DEDUP_CACHE_SIZE        4096 /* XXX make it a dynamic tunable */
 286
 287 typedef struct hammer_dedup_cache {
 288         RB_ENTRY(hammer_dedup_cache) crc_entry;
 289         RB_ENTRY(hammer_dedup_cache) off_entry;
 290         TAILQ_ENTRY(hammer_dedup_cache) lru_entry;
 291         struct hammer_mount *hmp;
 292         int64_t obj_id;
 293         u_int32_t localization;
 294         off_t file_offset;
 295         int bytes;
 296         hammer_off_t data_offset;
 297         hammer_crc_t crc;
 298 } *hammer_dedup_cache_t;
 299
 300 /*
 301  * Structure used to organize flush groups.  Flush groups must be
 302  * organized into chunks in order to avoid blowing out the UNDO FIFO.
 303  * Without this a 'sync' could end up flushing 50,000 inodes in a single
 304  * transaction.
 305  */
 306 struct hammer_fls_rb_tree;
 307 RB_HEAD(hammer_fls_rb_tree, hammer_inode);
 308 RB_PROTOTYPE(hammer_fls_rb_tree, hammer_inode, rb_flsnode,
 309               hammer_ino_rb_compare);
 310
 311 struct hammer_flush_group {
 312         TAILQ_ENTRY(hammer_flush_group) flush_entry;
 313         struct hammer_fls_rb_tree       flush_tree;
 314         int                             seq;            /* our seq no */
 315         int                             total_count;    /* record load */
 316         int                             running;        /* group is running */
 317         int                             closed;
 318         int                             refs;
 319 };
 320
 321 typedef struct hammer_flush_group *hammer_flush_group_t;
 322
 323 TAILQ_HEAD(hammer_flush_group_list, hammer_flush_group);
 324
 325 /*
 326  * Structure used to represent an inode in-memory.
 327  *
 328  * The record and data associated with an inode may be out of sync with
 329  * the disk (xDIRTY flags), or not even on the disk at all (ONDISK flag
 330  * clear).
 331  *
 332  * An inode may also hold a cache of unsynchronized records, used for
 333  * database and directories only.  Unsynchronized regular file data is
 334  * stored in the buffer cache.
 335  *
 336  * NOTE: A file which is created and destroyed within the initial
 337  * synchronization period can wind up not doing any disk I/O at all.
 338  *
 339  * Finally, an inode may cache numerous disk-referencing B-Tree cursors.
 340  */
 341 struct hammer_ino_rb_tree;
 342 struct hammer_inode;
 343 RB_HEAD(hammer_ino_rb_tree, hammer_inode);
 344 RB_PROTOTYPEX(hammer_ino_rb_tree, INFO, hammer_inode, rb_node,
 345               hammer_ino_rb_compare, hammer_inode_info_t);
 346
 347 struct hammer_redo_rb_tree;
 348 RB_HEAD(hammer_redo_rb_tree, hammer_inode);
 349 RB_PROTOTYPE2(hammer_redo_rb_tree, hammer_inode, rb_redonode,
 350               hammer_redo_rb_compare, hammer_off_t);
 351
 352 struct hammer_rec_rb_tree;
 353 struct hammer_record;
 354 RB_HEAD(hammer_rec_rb_tree, hammer_record);
 355 RB_PROTOTYPEX(hammer_rec_rb_tree, INFO, hammer_record, rb_node,
 356               hammer_rec_rb_compare, hammer_btree_leaf_elm_t);
 357
 358 TAILQ_HEAD(hammer_node_list, hammer_node);
 359
 360 struct hammer_inode {
 361         RB_ENTRY(hammer_inode)  rb_node;
 362         hammer_inode_state_t    flush_state;
 363         hammer_flush_group_t    flush_group;
 364         RB_ENTRY(hammer_inode)  rb_flsnode;     /* when on flush list */
 365         RB_ENTRY(hammer_inode)  rb_redonode;    /* when INODE_RDIRTY is set */
 366         struct hammer_record_list target_list;  /* target of dependant recs */
 367         int64_t                 obj_id;         /* (key) object identifier */
 368         hammer_tid_t            obj_asof;       /* (key) snapshot or 0 */
 369         u_int32_t               obj_localization; /* (key) pseudo-fs */
 370         struct hammer_mount     *hmp;
 371         hammer_objid_cache_t    objid_cache;
 372         int                     flags;
 373         int                     error;          /* flush error */
 374         int                     cursor_ip_refs; /* sanity */
 375         int                     cursor_exclreq_count;
 376         int                     rsv_recs;
 377         struct vnode            *vp;
 378         hammer_pseudofs_inmem_t pfsm;
 379         struct lockf            advlock;
 380         struct hammer_lock      lock;           /* sync copy interlock */
 381         off_t                   trunc_off;
 382         struct hammer_btree_leaf_elm ino_leaf;  /* in-memory cache */
 383         struct hammer_inode_data ino_data;      /* in-memory cache */
 384         struct hammer_rec_rb_tree rec_tree;     /* in-memory cache */
 385         int                     rec_generation;
 386         struct hammer_node_cache cache[4];      /* search initiate cache */
 387
 388         /*
 389          * When a demark is created to synchronize an inode to
 390          * disk, certain fields are copied so the front-end VOPs
 391          * can continue to run in parallel with the synchronization
 392          * occuring in the background.
 393          */
 394         int             sync_flags;             /* to-sync flags cache */
 395         off_t           sync_trunc_off;         /* to-sync truncation */
 396         off_t           save_trunc_off;         /* write optimization */
 397         struct hammer_btree_leaf_elm sync_ino_leaf; /* to-sync cache */
 398         struct hammer_inode_data sync_ino_data; /* to-sync cache */
 399         size_t          redo_count;
 400
 401         /*
 402          * Track the earliest offset in the UNDO/REDO FIFO containing
 403          * REDO records.  This is staged to the backend during flush
 404          * sequences.  While the inode is staged redo_fifo_next is used
 405          * to track the earliest offset for rotation into redo_fifo_start
 406          * on completion of the flush.
 407          */
 408         hammer_off_t    redo_fifo_start;
 409         hammer_off_t    redo_fifo_next;
 410 };
 411
 412 typedef struct hammer_inode *hammer_inode_t;
 413
 414 #define VTOI(vp)        ((struct hammer_inode *)(vp)->v_data)
 415
 416 /*
 417  * NOTE: DDIRTY does not include atime or mtime and does not include
 418  *       write-append size changes.  SDIRTY handles write-append size
 419  *       changes.
 420  *
 421  *       REDO indicates that REDO logging is active, creating a definitive
 422  *       stream of REDO records in the UNDO/REDO log for writes and
 423  *       truncations, including boundary records when/if REDO is turned off.
 424  *       REDO is typically enabled by fsync() and turned off if excessive
 425  *       writes without an fsync() occurs.
 426  *
 427  *       RDIRTY indicates that REDO records were laid down in the UNDO/REDO
 428  *       FIFO (even if REDO is turned off some might still be active) and
 429  *       still being tracked for this inode.  See hammer_redo.c
 430  */
 431                                         /* (not including atime/mtime) */
 432 #define HAMMER_INODE_DDIRTY     0x0001  /* in-memory ino_data is dirty */
 433 #define HAMMER_INODE_RSV_INODES 0x0002  /* hmp->rsv_inodes bumped */
 434 #define HAMMER_INODE_CONN_DOWN  0x0004  /* include in downward recursion */
 435 #define HAMMER_INODE_XDIRTY     0x0008  /* in-memory records */
 436 #define HAMMER_INODE_ONDISK     0x0010  /* inode is on-disk (else not yet) */
 437 #define HAMMER_INODE_FLUSH      0x0020  /* flush on last ref */
 438 #define HAMMER_INODE_DELETED    0x0080  /* inode delete (backend) */
 439 #define HAMMER_INODE_DELONDISK  0x0100  /* delete synchronized to disk */
 440 #define HAMMER_INODE_RO         0x0200  /* read-only (because of as-of) */
 441 #define HAMMER_INODE_RECSW      0x0400  /* waiting on data record flush */
 442 #define HAMMER_INODE_DONDISK    0x0800  /* data records may be on disk */
 443 #define HAMMER_INODE_BUFS       0x1000  /* dirty high level bps present */
 444 #define HAMMER_INODE_REFLUSH    0x2000  /* flush on dependancy / reflush */
 445 #define HAMMER_INODE_RECLAIM    0x4000  /* trying to reclaim */
 446 #define HAMMER_INODE_FLUSHW     0x8000  /* Someone waiting for flush */
 447
 448 #define HAMMER_INODE_TRUNCATED  0x00010000
 449 #define HAMMER_INODE_DELETING   0x00020000 /* inode delete request (frontend)*/
 450 #define HAMMER_INODE_RESIGNAL   0x00040000 /* re-signal on re-flush */
 451 #define HAMMER_INODE_ATIME      0x00100000 /* in-memory atime modified */
 452 #define HAMMER_INODE_MTIME      0x00200000 /* in-memory mtime modified */
 453 #define HAMMER_INODE_WOULDBLOCK 0x00400000 /* re-issue to new flush group */
 454 #define HAMMER_INODE_DUMMY      0x00800000 /* dummy inode covering bad file */
 455 #define HAMMER_INODE_SDIRTY     0x01000000 /* in-memory ino_data.size is dirty*/
 456 #define HAMMER_INODE_REDO       0x02000000 /* REDO logging active */
 457 #define HAMMER_INODE_RDIRTY     0x04000000 /* REDO records active in fifo */
 458 #define HAMMER_INODE_SLAVEFLUSH 0x08000000 /* being flushed by slave */
 459
 460 #define HAMMER_INODE_MODMASK    (HAMMER_INODE_DDIRTY|HAMMER_INODE_SDIRTY|   \
 461                                  HAMMER_INODE_XDIRTY|HAMMER_INODE_BUFS|     \
 462                                  HAMMER_INODE_ATIME|HAMMER_INODE_MTIME|     \
 463                                  HAMMER_INODE_TRUNCATED|HAMMER_INODE_DELETING)
 464
 465 #define HAMMER_INODE_MODMASK_NOXDIRTY   \
 466                                 (HAMMER_INODE_MODMASK & ~HAMMER_INODE_XDIRTY)
 467
 468 #define HAMMER_INODE_MODMASK_NOREDO     \
 469                                 (HAMMER_INODE_DDIRTY|                       \
 470                                  HAMMER_INODE_XDIRTY|                       \
 471                                  HAMMER_INODE_TRUNCATED|HAMMER_INODE_DELETING)
 472
 473 #define HAMMER_FLUSH_SIGNAL     0x0001
 474 #define HAMMER_FLUSH_RECURSION  0x0002
 475
 476 /*
 477  * Used by the inode reclaim code to pipeline reclaims and avoid
 478  * blowing out kernel memory or letting the flusher get too far
 479  * behind.  The reclaim wakes up when count reaches 0 or the
 480  * timer expires.
 481  */
 482 struct hammer_reclaim {
 483         TAILQ_ENTRY(hammer_reclaim) entry;
 484         int     count;
 485 };
 486
 487 /*
 488  * Track who is creating the greatest burden on the
 489  * inode cache.
 490  */
 491 struct hammer_inostats {
 492         pid_t           pid;    /* track user process */
 493         int             ltick;  /* last tick */
 494         int             count;  /* count (degenerates) */
 495 };
 496
 497 #define HAMMER_INOSTATS_HSIZE   32
 498 #define HAMMER_INOSTATS_HMASK   (HAMMER_INOSTATS_HSIZE - 1)
 499
 500 /*
 501  * Structure used to represent an unsynchronized record in-memory.  These
 502  * records typically represent directory entries.  Only non-historical
 503  * records are kept in-memory.
 504  *
 505  * Records are organized as a per-inode RB-Tree.  If the inode is not
 506  * on disk then neither are any records and the in-memory record tree
 507  * represents the entire contents of the inode.  If the inode is on disk
 508  * then the on-disk B-Tree is scanned in parallel with the in-memory
 509  * RB-Tree to synthesize the current state of the file.
 510  *
 511  * Records are also used to enforce the ordering of directory create/delete
 512  * operations.  A new inode will not be flushed to disk unless its related
 513  * directory entry is also being flushed at the same time.  A directory entry
 514  * will not be removed unless its related inode is also being removed at the
 515  * same time.
 516  */
 517 typedef enum hammer_record_type {
 518         HAMMER_MEM_RECORD_GENERAL,      /* misc record */
 519         HAMMER_MEM_RECORD_INODE,        /* inode record */
 520         HAMMER_MEM_RECORD_ADD,          /* positive memory cache record */
 521         HAMMER_MEM_RECORD_DEL,          /* negative delete-on-disk record */
 522         HAMMER_MEM_RECORD_DATA          /* bulk-data record w/on-disk ref */
 523 } hammer_record_type_t;
 524
 525 struct hammer_record {
 526         RB_ENTRY(hammer_record)         rb_node;
 527         TAILQ_ENTRY(hammer_record)      target_entry;
 528         hammer_inode_state_t            flush_state;
 529         hammer_flush_group_t            flush_group;
 530         hammer_record_type_t            type;
 531         struct hammer_lock              lock;
 532         struct hammer_reserve           *resv;
 533         struct hammer_inode             *ip;
 534         struct hammer_inode             *target_ip;
 535         struct hammer_btree_leaf_elm    leaf;
 536         union hammer_data_ondisk        *data;
 537         int                             flags;
 538         int                             gflags;
 539         hammer_off_t                    zone2_offset;   /* direct-write only */
 540 };
 541
 542 typedef struct hammer_record *hammer_record_t;
 543
 544 /*
 545  * Record flags.  Note that FE can only be set by the frontend if the
 546  * record has not been interlocked by the backend w/ BE.
 547  */
 548 #define HAMMER_RECF_ALLOCDATA           0x0001
 549 #define HAMMER_RECF_ONRBTREE            0x0002
 550 #define HAMMER_RECF_DELETED_FE          0x0004  /* deleted (frontend) */
 551 #define HAMMER_RECF_DELETED_BE          0x0008  /* deleted (backend) */
 552 #define HAMMER_RECF_COMMITTED           0x0010  /* committed to the B-Tree */
 553 #define HAMMER_RECF_INTERLOCK_BE        0x0020  /* backend interlock */
 554 #define HAMMER_RECF_WANTED              0x0040  /* wanted by the frontend */
 555 #define HAMMER_RECF_DEDUPED             0x0080  /* will be live-dedup'ed */
 556 #define HAMMER_RECF_CONVERT_DELETE      0x0100  /* special case */
 557 #define HAMMER_RECF_REDO                0x1000  /* REDO was laid down */
 558
 559 /*
 560  * These flags must be separate to deal with SMP races
 561  */
 562 #define HAMMER_RECG_DIRECT_IO           0x0001  /* related direct I/O running*/
 563 #define HAMMER_RECG_DIRECT_WAIT         0x0002  /* related direct I/O running*/
 564 #define HAMMER_RECG_DIRECT_INVAL        0x0004  /* buffer alias invalidation */
 565 /*
 566  * hammer_create_at_cursor() and hammer_delete_at_cursor() flags.
 567  */
 568 #define HAMMER_CREATE_MODE_UMIRROR      0x0001
 569 #define HAMMER_CREATE_MODE_SYS          0x0002
 570
 571 #define HAMMER_DELETE_ADJUST            0x0001
 572 #define HAMMER_DELETE_DESTROY           0x0002
 573
 574 /*
 575  * In-memory structures representing on-disk structures.
 576  */
 577 struct hammer_volume;
 578 struct hammer_buffer;
 579 struct hammer_node;
 580 struct hammer_undo;
 581 struct hammer_reserve;
 582
 583 RB_HEAD(hammer_vol_rb_tree, hammer_volume);
 584 RB_HEAD(hammer_buf_rb_tree, hammer_buffer);
 585 RB_HEAD(hammer_nod_rb_tree, hammer_node);
 586 RB_HEAD(hammer_und_rb_tree, hammer_undo);
 587 RB_HEAD(hammer_res_rb_tree, hammer_reserve);
 588 RB_HEAD(hammer_mod_rb_tree, hammer_io);
 589
 590 RB_PROTOTYPE2(hammer_vol_rb_tree, hammer_volume, rb_node,
 591               hammer_vol_rb_compare, int32_t);
 592 RB_PROTOTYPE2(hammer_buf_rb_tree, hammer_buffer, rb_node,
 593               hammer_buf_rb_compare, hammer_off_t);
 594 RB_PROTOTYPE2(hammer_nod_rb_tree, hammer_node, rb_node,
 595               hammer_nod_rb_compare, hammer_off_t);
 596 RB_PROTOTYPE2(hammer_und_rb_tree, hammer_undo, rb_node,
 597               hammer_und_rb_compare, hammer_off_t);
 598 RB_PROTOTYPE2(hammer_res_rb_tree, hammer_reserve, rb_node,
 599               hammer_res_rb_compare, hammer_off_t);
 600 RB_PROTOTYPE2(hammer_mod_rb_tree, hammer_io, rb_node,
 601               hammer_mod_rb_compare, hammer_off_t);
 602
 603 /*
 604  * IO management - embedded at the head of various in-memory structures
 605  *
 606  * VOLUME       - hammer_volume containing meta-data
 607  * META_BUFFER  - hammer_buffer containing meta-data
 608  * DATA_BUFFER  - hammer_buffer containing pure-data
 609  *
 610  * Dirty volume headers and dirty meta-data buffers are locked until the
 611  * flusher can sequence them out.  Dirty pure-data buffers can be written.
 612  * Clean buffers can be passively released.
 613  */
 614 typedef enum hammer_io_type {
 615         HAMMER_STRUCTURE_VOLUME,
 616         HAMMER_STRUCTURE_META_BUFFER,
 617         HAMMER_STRUCTURE_UNDO_BUFFER,
 618         HAMMER_STRUCTURE_DATA_BUFFER,
 619         HAMMER_STRUCTURE_DUMMY
 620 } hammer_io_type_t;
 621
 622 union hammer_io_structure;
 623 struct hammer_io;
 624
 625 struct worklist {
 626         LIST_ENTRY(worklist) node;
 627 };
 628
 629 TAILQ_HEAD(hammer_io_list, hammer_io);
 630 typedef struct hammer_io_list *hammer_io_list_t;
 631
 632 struct hammer_io {
 633         struct worklist         worklist;
 634         struct hammer_lock      lock;
 635         enum hammer_io_type     type;
 636         struct hammer_mount     *hmp;
 637         struct hammer_volume    *volume;
 638         RB_ENTRY(hammer_io)     rb_node;     /* if modified */
 639         TAILQ_ENTRY(hammer_io)  iorun_entry; /* iorun_list */
 640         struct hammer_mod_rb_tree *mod_root;
 641         struct buf              *bp;
 642         int64_t                 offset;    /* zone-2 offset */
 643         int                     bytes;     /* buffer cache buffer size */
 644         int                     modify_refs;
 645
 646         /*
 647          * These can be modified at any time by the backend while holding
 648          * io_token, due to bio_done and hammer_io_complete() callbacks.
 649          */
 650         u_int           running : 1;    /* bp write IO in progress */
 651         u_int           waiting : 1;    /* someone is waiting on us */
 652         u_int           ioerror : 1;    /* abort on io-error */
 653         u_int           unusedA : 29;
 654
 655         /*
 656          * These can only be modified by the frontend while holding
 657          * fs_token, or by the backend while holding the io interlocked
 658          * with no references (which will block the frontend when it
 659          * tries to reference it).
 660          *
 661          * WARNING! SMP RACES will create havoc if the callbacks ever tried
 662          *          to modify any of these outside the above restrictions.
 663          */
 664         u_int           modified : 1;   /* bp's data was modified */
 665         u_int           released : 1;   /* bp released (w/ B_LOCKED set) */
 666         u_int           validated : 1;  /* ondisk has been validated */
 667         u_int           waitdep : 1;    /* flush waits for dependancies */
 668         u_int           recovered : 1;  /* has recovery ref */
 669         u_int           waitmod : 1;    /* waiting for modify_refs */
 670         u_int           reclaim : 1;    /* reclaim requested */
 671         u_int           gencrc : 1;     /* crc needs to be generated */
 672         u_int           unusedB : 24;
 673 };
 674
 675 typedef struct hammer_io *hammer_io_t;
 676
 677 #define HAMMER_CLUSTER_SIZE     (64 * 1024)
 678 #if HAMMER_CLUSTER_SIZE > MAXBSIZE
 679 #undef  HAMMER_CLUSTER_SIZE
 680 #define HAMMER_CLUSTER_SIZE     MAXBSIZE
 681 #endif
 682 #define HAMMER_CLUSTER_BUFS     (HAMMER_CLUSTER_SIZE / HAMMER_BUFSIZE)
 683
 684 /*
 685  * In-memory volume representing on-disk buffer
 686  */
 687 struct hammer_volume {
 688         struct hammer_io io;
 689         RB_ENTRY(hammer_volume) rb_node;
 690         struct hammer_volume_ondisk *ondisk;
 691         int32_t vol_no;
 692         int64_t nblocks;        /* note: special calculation for statfs */
 693         int64_t buffer_base;    /* base offset of buffer 0 */
 694         hammer_off_t maxbuf_off; /* Maximum buffer offset (zone-2) */
 695         hammer_off_t maxraw_off; /* Maximum raw offset for device */
 696         char    *vol_name;
 697         struct vnode *devvp;
 698         int     vol_flags;
 699 };
 700
 701 typedef struct hammer_volume *hammer_volume_t;
 702
 703 /*
 704  * In-memory buffer (other then volume, super-cluster, or cluster),
 705  * representing an on-disk buffer.
 706  */
 707 struct hammer_buffer {
 708         struct hammer_io io;
 709         RB_ENTRY(hammer_buffer) rb_node;
 710         void *ondisk;
 711         hammer_off_t zoneX_offset;
 712         hammer_off_t zone2_offset;
 713         struct hammer_reserve *resv;
 714         struct hammer_node_list clist;
 715 };
 716
 717 typedef struct hammer_buffer *hammer_buffer_t;
 718
 719 /*
 720  * In-memory B-Tree node, representing an on-disk B-Tree node.
 721  *
 722  * This is a hang-on structure which is backed by a hammer_buffer,
 723  * indexed by a hammer_cluster, and used for fine-grained locking of
 724  * B-Tree nodes in order to properly control lock ordering.  A hammer_buffer
 725  * can contain multiple nodes representing wildly disassociated portions
 726  * of the B-Tree so locking cannot be done on a buffer-by-buffer basis.
 727  *
 728  * This structure uses a cluster-relative index to reduce the number
 729  * of layers required to access it, and also because all on-disk B-Tree
 730  * references are cluster-relative offsets.
 731  */
 732 struct hammer_node {
 733         struct hammer_lock      lock;           /* node-by-node lock */
 734         TAILQ_ENTRY(hammer_node) entry;         /* per-buffer linkage */
 735         RB_ENTRY(hammer_node)   rb_node;        /* per-cluster linkage */
 736         hammer_off_t            node_offset;    /* full offset spec */
 737         struct hammer_mount     *hmp;
 738         struct hammer_buffer    *buffer;        /* backing buffer */
 739         hammer_node_ondisk_t    ondisk;         /* ptr to on-disk structure */
 740         TAILQ_HEAD(, hammer_cursor) cursor_list;  /* deadlock recovery */
 741         struct hammer_node_cache_list cache_list; /* passive caches */
 742         int                     flags;
 743         int                     cursor_exclreq_count;
 744 };
 745
 746 #define HAMMER_NODE_DELETED     0x0001
 747 #define HAMMER_NODE_FLUSH       0x0002
 748 #define HAMMER_NODE_CRCGOOD     0x0004
 749 #define HAMMER_NODE_NEEDSCRC    0x0008
 750 #define HAMMER_NODE_NEEDSMIRROR 0x0010
 751 #define HAMMER_NODE_CRCBAD      0x0020
 752 #define HAMMER_NODE_NONLINEAR   0x0040          /* linear heuristic */
 753
 754 #define HAMMER_NODE_CRCANY      (HAMMER_NODE_CRCGOOD | HAMMER_NODE_CRCBAD)
 755
 756 typedef struct hammer_node      *hammer_node_t;
 757
 758 /*
 759  * List of locked nodes.  This structure is used to lock potentially large
 760  * numbers of nodes as an aid for complex B-Tree operations.
 761  */
 762 struct hammer_node_lock;
 763 TAILQ_HEAD(hammer_node_lock_list, hammer_node_lock);
 764
 765 struct hammer_node_lock {
 766         TAILQ_ENTRY(hammer_node_lock) entry;
 767         struct hammer_node_lock_list  list;
 768         struct hammer_node_lock       *parent;
 769         hammer_node_t   node;
 770         hammer_node_ondisk_t copy;      /* copy of on-disk data */
 771         int             index;          /* index of this node in parent */
 772         int             count;          /* count children */
 773         int             flags;
 774 };
 775
 776 typedef struct hammer_node_lock *hammer_node_lock_t;
 777
 778 #define HAMMER_NODE_LOCK_UPDATED        0x0001
 779 #define HAMMER_NODE_LOCK_LCACHE         0x0002
 780
 781 /*
 782  * Common I/O management structure - embedded in in-memory structures
 783  * which are backed by filesystem buffers.
 784  */
 785 union hammer_io_structure {
 786         struct hammer_io        io;
 787         struct hammer_volume    volume;
 788         struct hammer_buffer    buffer;
 789 };
 790
 791 typedef union hammer_io_structure *hammer_io_structure_t;
 792
 793 /*
 794  * The reserve structure prevents the blockmap from allocating
 795  * out of a reserved bigblock.  Such reservations are used by
 796  * the direct-write mechanism.
 797  *
 798  * The structure is also used to hold off on reallocations of
 799  * big blocks from the freemap until flush dependancies have
 800  * been dealt with.
 801  */
 802 struct hammer_reserve {
 803         RB_ENTRY(hammer_reserve) rb_node;
 804         TAILQ_ENTRY(hammer_reserve) delay_entry;
 805         int             flush_group;
 806         int             flags;
 807         int             refs;
 808         int             zone;
 809         int             append_off;
 810         int32_t         bytes_free;
 811         hammer_off_t    zone_offset;
 812 };
 813
 814 typedef struct hammer_reserve *hammer_reserve_t;
 815
 816 #define HAMMER_RESF_ONDELAY     0x0001
 817 #define HAMMER_RESF_LAYER2FREE  0x0002
 818
 819 #include "hammer_cursor.h"
 820
 821 /*
 822  * The undo structure tracks recent undos to avoid laying down duplicate
 823  * undos within a flush group, saving us a significant amount of overhead.
 824  *
 825  * This is strictly a heuristic.
 826  */
 827 #define HAMMER_MAX_UNDOS                1024
 828 #define HAMMER_MAX_FLUSHERS             4
 829
 830 struct hammer_undo {
 831         RB_ENTRY(hammer_undo)   rb_node;
 832         TAILQ_ENTRY(hammer_undo) lru_entry;
 833         hammer_off_t            offset;
 834         int                     bytes;
 835 };
 836
 837 typedef struct hammer_undo *hammer_undo_t;
 838
 839 struct hammer_flusher_info;
 840 TAILQ_HEAD(hammer_flusher_info_list, hammer_flusher_info);
 841
 842 struct hammer_flusher {
 843         int             signal;         /* flusher thread sequencer */
 844         int             done;           /* last completed flush group */
 845         int             next;           /* next unallocated flg seqno */
 846         int             group_lock;     /* lock sequencing of the next flush */
 847         int             exiting;        /* request master exit */
 848         thread_t        td;             /* master flusher thread */
 849         hammer_tid_t    tid;            /* last flushed transaction id */
 850         int             finalize_want;          /* serialize finalization */
 851         struct hammer_lock finalize_lock;       /* serialize finalization */
 852         struct hammer_transaction trans;        /* shared transaction */
 853         struct hammer_flusher_info_list run_list;
 854         struct hammer_flusher_info_list ready_list;
 855 };
 856
 857 #define HAMMER_FLUSH_UNDOS_RELAXED      0
 858 #define HAMMER_FLUSH_UNDOS_FORCED       1
 859 #define HAMMER_FLUSH_UNDOS_AUTO         2
 860 /*
 861  * Internal hammer mount data structure
 862  */
 863 struct hammer_mount {
 864         struct mount *mp;
 865         /*struct vnode *rootvp;*/
 866         struct hammer_ino_rb_tree rb_inos_root;
 867         struct hammer_redo_rb_tree rb_redo_root;
 868         struct hammer_vol_rb_tree rb_vols_root;
 869         struct hammer_nod_rb_tree rb_nods_root;
 870         struct hammer_und_rb_tree rb_undo_root;
 871         struct hammer_res_rb_tree rb_resv_root;
 872         struct hammer_buf_rb_tree rb_bufs_root;
 873         struct hammer_pfs_rb_tree rb_pfsm_root;
 874
 875         struct hammer_dedup_crc_rb_tree rb_dedup_crc_root;
 876         struct hammer_dedup_off_rb_tree rb_dedup_off_root;
 877
 878         struct hammer_volume *rootvol;
 879         struct hammer_base_elm root_btree_beg;
 880         struct hammer_base_elm root_btree_end;
 881
 882         struct malloc_type      *m_misc;
 883         struct malloc_type      *m_inodes;
 884
 885         int     flags;          /* HAMMER_MOUNT_xxx flags */
 886         int     hflags;
 887         int     ronly;
 888         int     nvolumes;
 889         int     volume_iterator;
 890         int     master_id;      /* -1 or 0-15 - clustering and mirroring */
 891         int     version;        /* hammer filesystem version to use */
 892         int     rsv_inodes;     /* reserved space due to dirty inodes */
 893         int64_t rsv_databytes;  /* reserved space due to record data */
 894         int     rsv_recs;       /* reserved space due to dirty records */
 895         int     rsv_fromdelay;  /* bigblocks reserved due to flush delay */
 896         int     undo_rec_limit; /* based on size of undo area */
 897         int     last_newrecords;
 898         int     count_newrecords;
 899
 900         int     volume_to_remove; /* volume that is currently being removed */
 901
 902         int     count_inodes;   /* total number of inodes */
 903         int     count_iqueued;  /* inodes queued to flusher */
 904         int     count_reclaims; /* inodes pending reclaim by flusher */
 905
 906         struct hammer_flusher flusher;
 907
 908         u_int   check_interrupt;
 909         u_int   check_yield;
 910         uuid_t  fsid;
 911         struct hammer_mod_rb_tree volu_root;    /* dirty undo buffers */
 912         struct hammer_mod_rb_tree undo_root;    /* dirty undo buffers */
 913         struct hammer_mod_rb_tree data_root;    /* dirty data buffers */
 914         struct hammer_mod_rb_tree meta_root;    /* dirty meta bufs    */
 915         struct hammer_mod_rb_tree lose_root;    /* loose buffers      */
 916         long    locked_dirty_space;             /* meta/volu count    */
 917         long    io_running_space;               /* io_token */
 918         int     unused01;
 919         int     objid_cache_count;
 920         int     dedup_cache_count;
 921         int     error;                          /* critical I/O error */
 922         struct krate    krate;                  /* rate limited kprintf */
 923         hammer_tid_t    asof;                   /* snapshot mount */
 924         hammer_tid_t    next_tid;
 925         hammer_tid_t    flush_tid1;             /* flusher tid sequencing */
 926         hammer_tid_t    flush_tid2;             /* flusher tid sequencing */
 927         int64_t copy_stat_freebigblocks;        /* number of free bigblocks */
 928         u_int32_t       undo_seqno;             /* UNDO/REDO FIFO seqno */
 929         u_int32_t       recover_stage2_seqno;   /* REDO recovery seqno */
 930         hammer_off_t    recover_stage2_offset;  /* REDO recovery offset */
 931
 932         struct netexport export;
 933         struct hammer_lock sync_lock;
 934         struct hammer_lock free_lock;
 935         struct hammer_lock undo_lock;
 936         struct hammer_lock blkmap_lock;
 937         struct hammer_lock snapshot_lock;
 938         struct hammer_lock volume_lock;
 939         struct hammer_blockmap  blockmap[HAMMER_MAX_ZONES];
 940         struct hammer_undo      undos[HAMMER_MAX_UNDOS];
 941         int                     undo_alloc;
 942         TAILQ_HEAD(, hammer_undo)  undo_lru_list;
 943         TAILQ_HEAD(, hammer_reserve) delay_list;
 944         struct hammer_flush_group_list  flush_group_list;
 945         hammer_flush_group_t    fill_flush_group;
 946         hammer_flush_group_t    next_flush_group;
 947         TAILQ_HEAD(, hammer_objid_cache) objid_cache_list;
 948         TAILQ_HEAD(, hammer_dedup_cache) dedup_lru_list;
 949         hammer_dedup_cache_t    dedup_free_cache;
 950         TAILQ_HEAD(, hammer_reclaim) reclaim_list;
 951         TAILQ_HEAD(, hammer_io) iorun_list;
 952
 953         struct lwkt_token       fs_token;       /* high level */
 954         struct lwkt_token       io_token;       /* low level (IO callback) */
 955
 956         struct hammer_inostats  inostats[HAMMER_INOSTATS_HSIZE];
 957 };
 958
 959 typedef struct hammer_mount     *hammer_mount_t;
 960
 961 #define HAMMER_MOUNT_CRITICAL_ERROR     0x0001
 962 #define HAMMER_MOUNT_FLUSH_RECOVERY     0x0002
 963 #define HAMMER_MOUNT_REDO_SYNC          0x0004
 964 #define HAMMER_MOUNT_REDO_RECOVERY_REQ  0x0008
 965 #define HAMMER_MOUNT_REDO_RECOVERY_RUN  0x0010
 966
 967 struct hammer_sync_info {
 968         int error;
 969         int waitfor;
 970 };
 971
 972 /*
 973  * Minium buffer cache bufs required to rebalance the B-Tree.
 974  * This is because we must hold the children and the children's children
 975  * locked.  Even this might not be enough if things are horribly out
 976  * of balance.
 977  */
 978 #define HAMMER_REBALANCE_MIN_BUFS       \
 979         (HAMMER_BTREE_LEAF_ELMS * HAMMER_BTREE_LEAF_ELMS)
 980
 981
 982 #endif
 983
 984 /*
 985  * checkspace slop (8MB chunks), higher numbers are more conservative.
 986  */
 987 #define HAMMER_CHKSPC_REBLOCK   25
 988 #define HAMMER_CHKSPC_MIRROR    20
 989 #define HAMMER_CHKSPC_WRITE     20
 990 #define HAMMER_CHKSPC_CREATE    20
 991 #define HAMMER_CHKSPC_REMOVE    10
 992 #define HAMMER_CHKSPC_EMERGENCY 0
 993
 994 #if defined(_KERNEL)
 995
 996 extern struct vop_ops hammer_vnode_vops;
 997 extern struct vop_ops hammer_spec_vops;
 998 extern struct vop_ops hammer_fifo_vops;
 999 extern struct bio_ops hammer_bioops;
1000
1001 extern int hammer_debug_io;
1002 extern int hammer_debug_general;
1003 extern int hammer_debug_debug;
1004 extern int hammer_debug_inode;
1005 extern int hammer_debug_locks;
1006 extern int hammer_debug_btree;
1007 extern int hammer_debug_tid;
1008 extern int hammer_debug_recover;
1009 extern int hammer_debug_recover_faults;
1010 extern int hammer_debug_critical;
1011 extern int hammer_cluster_enable;
1012 extern int hammer_live_dedup;
1013 extern int hammer_tdmux_ticks;
1014 extern int hammer_count_fsyncs;
1015 extern int hammer_count_inodes;
1016 extern int hammer_count_iqueued;
1017 extern int hammer_count_reclaims;
1018 extern int hammer_count_records;
1019 extern int hammer_count_record_datas;
1020 extern int hammer_count_volumes;
1021 extern int hammer_count_buffers;
1022 extern int hammer_count_nodes;
1023 extern int64_t hammer_count_extra_space_used;
1024 extern int64_t hammer_stats_btree_lookups;
1025 extern int64_t hammer_stats_btree_searches;
1026 extern int64_t hammer_stats_btree_inserts;
1027 extern int64_t hammer_stats_btree_deletes;
1028 extern int64_t hammer_stats_btree_elements;
1029 extern int64_t hammer_stats_btree_splits;
1030 extern int64_t hammer_stats_btree_iterations;
1031 extern int64_t hammer_stats_btree_root_iterations;
1032 extern int64_t hammer_stats_record_iterations;
1033 extern int64_t hammer_stats_file_read;
1034 extern int64_t hammer_stats_file_write;
1035 extern int64_t hammer_stats_file_iopsr;
1036 extern int64_t hammer_stats_file_iopsw;
1037 extern int64_t hammer_stats_disk_read;
1038 extern int64_t hammer_stats_disk_write;
1039 extern int64_t hammer_stats_inode_flushes;
1040 extern int64_t hammer_stats_commits;
1041 extern int64_t hammer_stats_undo;
1042 extern int64_t hammer_stats_redo;
1043 extern long hammer_count_dirtybufspace;
1044 extern int hammer_count_refedbufs;
1045 extern int hammer_count_reservations;
1046 extern long hammer_count_io_running_read;
1047 extern long hammer_count_io_running_write;
1048 extern int hammer_count_io_locked;
1049 extern long hammer_limit_dirtybufspace;
1050 extern int hammer_limit_recs;
1051 extern int hammer_limit_inode_recs;
1052 extern int hammer_limit_reclaims;
1053 extern int hammer_live_dedup_cache_size;
1054 extern int hammer_limit_redo;
1055 extern int hammer_bio_count;
1056 extern int hammer_verify_zone;
1057 extern int hammer_verify_data;
1058 extern int hammer_write_mode;
1059 extern int hammer_double_buffer;
1060 extern int hammer_btree_full_undo;
1061 extern int hammer_yield_check;
1062 extern int hammer_fsync_mode;
1063 extern int hammer_autoflush;
1064 extern int64_t hammer_contention_count;
1065
1066 extern int64_t hammer_live_dedup_vnode_bcmps;
1067 extern int64_t hammer_live_dedup_device_bcmps;
1068 extern int64_t hammer_live_dedup_findblk_failures;
1069 extern int64_t hammer_live_dedup_bmap_saves;
1070
1071 void    hammer_critical_error(hammer_mount_t hmp, hammer_inode_t ip,
1072                         int error, const char *msg);
1073 int     hammer_vop_inactive(struct vop_inactive_args *);
1074 int     hammer_vop_reclaim(struct vop_reclaim_args *);
1075 int     hammer_get_vnode(struct hammer_inode *ip, struct vnode **vpp);
1076 struct hammer_inode *hammer_get_inode(hammer_transaction_t trans,
1077                         hammer_inode_t dip, int64_t obj_id,
1078                         hammer_tid_t asof, u_int32_t localization,
1079                         int flags, int *errorp);
1080 struct hammer_inode *hammer_get_dummy_inode(hammer_transaction_t trans,
1081                         hammer_inode_t dip, int64_t obj_id,
1082                         hammer_tid_t asof, u_int32_t localization,
1083                         int flags, int *errorp);
1084 struct hammer_inode *hammer_find_inode(hammer_transaction_t trans,
1085                         int64_t obj_id, hammer_tid_t asof,
1086                         u_int32_t localization);
1087 void    hammer_scan_inode_snapshots(hammer_mount_t hmp,
1088                         hammer_inode_info_t iinfo,
1089                         int (*callback)(hammer_inode_t ip, void *data),
1090                         void *data);
1091 void    hammer_put_inode(struct hammer_inode *ip);
1092 void    hammer_put_inode_ref(struct hammer_inode *ip);
1093 void    hammer_inode_waitreclaims(hammer_transaction_t trans);
1094
1095 int     hammer_unload_volume(hammer_volume_t volume, void *data __unused);
1096 int     hammer_adjust_volume_mode(hammer_volume_t volume, void *data __unused);
1097
1098 int     hammer_unload_buffer(hammer_buffer_t buffer, void *data);
1099 int     hammer_install_volume(hammer_mount_t hmp, const char *volname,
1100                         struct vnode *devvp);
1101 int     hammer_mountcheck_volumes(hammer_mount_t hmp);
1102
1103 int     hammer_mem_add(hammer_record_t record);
1104 int     hammer_ip_lookup(hammer_cursor_t cursor);
1105 int     hammer_ip_first(hammer_cursor_t cursor);
1106 int     hammer_ip_next(hammer_cursor_t cursor);
1107 int     hammer_ip_resolve_data(hammer_cursor_t cursor);
1108 int     hammer_ip_delete_record(hammer_cursor_t cursor, hammer_inode_t ip,
1109                         hammer_tid_t tid);
1110 int     hammer_create_at_cursor(hammer_cursor_t cursor,
1111                         hammer_btree_leaf_elm_t leaf, void *udata, int mode);
1112 int     hammer_delete_at_cursor(hammer_cursor_t cursor, int delete_flags,
1113                         hammer_tid_t delete_tid, u_int32_t delete_ts,
1114                         int track, int64_t *stat_bytes);
1115 int     hammer_ip_check_directory_empty(hammer_transaction_t trans,
1116                         hammer_inode_t ip);
1117 int     hammer_sync_hmp(hammer_mount_t hmp, int waitfor);
1118 int     hammer_queue_inodes_flusher(hammer_mount_t hmp, int waitfor);
1119
1120 hammer_record_t
1121         hammer_alloc_mem_record(hammer_inode_t ip, int data_len);
1122 void    hammer_flush_record_done(hammer_record_t record, int error);
1123 void    hammer_wait_mem_record_ident(hammer_record_t record, const char *ident);
1124 void    hammer_rel_mem_record(hammer_record_t record);
1125
1126 int     hammer_cursor_up(hammer_cursor_t cursor);
1127 int     hammer_cursor_up_locked(hammer_cursor_t cursor);
1128 int     hammer_cursor_down(hammer_cursor_t cursor);
1129 int     hammer_cursor_upgrade(hammer_cursor_t cursor);
1130 int     hammer_cursor_upgrade_node(hammer_cursor_t cursor);
1131 void    hammer_cursor_downgrade(hammer_cursor_t cursor);
1132 int     hammer_cursor_upgrade2(hammer_cursor_t c1, hammer_cursor_t c2);
1133 void    hammer_cursor_downgrade2(hammer_cursor_t c1, hammer_cursor_t c2);
1134 int     hammer_cursor_seek(hammer_cursor_t cursor, hammer_node_t node,
1135                         int index);
1136 void    hammer_lock_ex_ident(struct hammer_lock *lock, const char *ident);
1137 int     hammer_lock_ex_try(struct hammer_lock *lock);
1138 void    hammer_lock_sh(struct hammer_lock *lock);
1139 int     hammer_lock_sh_try(struct hammer_lock *lock);
1140 int     hammer_lock_upgrade(struct hammer_lock *lock, int shcount);
1141 void    hammer_lock_downgrade(struct hammer_lock *lock, int shcount);
1142 int     hammer_lock_status(struct hammer_lock *lock);
1143 void    hammer_unlock(struct hammer_lock *lock);
1144 void    hammer_ref(struct hammer_lock *lock);
1145 int     hammer_ref_interlock(struct hammer_lock *lock);
1146 int     hammer_ref_interlock_true(struct hammer_lock *lock);
1147 void    hammer_ref_interlock_done(struct hammer_lock *lock);
1148 void    hammer_rel(struct hammer_lock *lock);
1149 int     hammer_rel_interlock(struct hammer_lock *lock, int locked);
1150 void    hammer_rel_interlock_done(struct hammer_lock *lock, int orig_locked);
1151 int     hammer_get_interlock(struct hammer_lock *lock);
1152 int     hammer_try_interlock_norefs(struct hammer_lock *lock);
1153 void    hammer_put_interlock(struct hammer_lock *lock, int error);
1154
1155 void    hammer_sync_lock_ex(hammer_transaction_t trans);
1156 void    hammer_sync_lock_sh(hammer_transaction_t trans);
1157 int     hammer_sync_lock_sh_try(hammer_transaction_t trans);
1158 void    hammer_sync_unlock(hammer_transaction_t trans);
1159
1160 u_int32_t hammer_to_unix_xid(uuid_t *uuid);
1161 void hammer_guid_to_uuid(uuid_t *uuid, u_int32_t guid);
1162 void    hammer_time_to_timespec(u_int64_t xtime, struct timespec *ts);
1163 u_int64_t hammer_timespec_to_time(struct timespec *ts);
1164 int     hammer_str_to_tid(const char *str, int *ispfsp,
1165                         hammer_tid_t *tidp, u_int32_t *localizationp);
1166 int     hammer_is_atatext(const char *name, int len);
1167 hammer_tid_t hammer_alloc_objid(hammer_mount_t hmp, hammer_inode_t dip,
1168                         int64_t namekey);
1169 void hammer_clear_objid(hammer_inode_t dip);
1170 void hammer_destroy_objid_cache(hammer_mount_t hmp);
1171
1172 int hammer_dedup_crc_rb_compare(hammer_dedup_cache_t dc1,
1173                         hammer_dedup_cache_t dc2);
1174 int hammer_dedup_off_rb_compare(hammer_dedup_cache_t dc1,
1175                         hammer_dedup_cache_t dc2);
1176 hammer_dedup_cache_t hammer_dedup_cache_add(hammer_inode_t ip,
1177                         hammer_btree_leaf_elm_t leaf);
1178 hammer_dedup_cache_t hammer_dedup_cache_lookup(hammer_mount_t hmp,
1179                         hammer_crc_t crc);
1180 void hammer_dedup_cache_inval(hammer_mount_t hmp, hammer_off_t base_offset);
1181 void hammer_destroy_dedup_cache(hammer_mount_t hmp);
1182 void hammer_dump_dedup_cache(hammer_mount_t hmp);
1183 int hammer_dedup_validate(hammer_dedup_cache_t dcp, int zone, int bytes,
1184                         void *data);
1185
1186 int hammer_enter_undo_history(hammer_mount_t hmp, hammer_off_t offset,
1187                         int bytes);
1188 void hammer_clear_undo_history(hammer_mount_t hmp);
1189 enum vtype hammer_get_vnode_type(u_int8_t obj_type);
1190 int hammer_get_dtype(u_int8_t obj_type);
1191 u_int8_t hammer_get_obj_type(enum vtype vtype);
1192 int64_t hammer_directory_namekey(hammer_inode_t dip, const void *name, int len,
1193                         u_int32_t *max_iterationsp);
1194 int     hammer_nohistory(hammer_inode_t ip);
1195
1196 int     hammer_init_cursor(hammer_transaction_t trans, hammer_cursor_t cursor,
1197                         hammer_node_cache_t cache, hammer_inode_t ip);
1198 void    hammer_normalize_cursor(hammer_cursor_t cursor);
1199 void    hammer_done_cursor(hammer_cursor_t cursor);
1200 int     hammer_recover_cursor(hammer_cursor_t cursor);
1201 void    hammer_unlock_cursor(hammer_cursor_t cursor);
1202 int     hammer_lock_cursor(hammer_cursor_t cursor);
1203 hammer_cursor_t hammer_push_cursor(hammer_cursor_t ocursor);
1204 void    hammer_pop_cursor(hammer_cursor_t ocursor, hammer_cursor_t ncursor);
1205
1206 void    hammer_cursor_replaced_node(hammer_node_t onode, hammer_node_t nnode);
1207 void    hammer_cursor_removed_node(hammer_node_t onode, hammer_node_t parent,
1208                         int index);
1209 void    hammer_cursor_split_node(hammer_node_t onode, hammer_node_t nnode,
1210                         int index);
1211 void    hammer_cursor_moved_element(hammer_node_t oparent, int pindex,
1212                         hammer_node_t onode, int oindex,
1213                         hammer_node_t nnode, int nindex);
1214 void    hammer_cursor_parent_changed(hammer_node_t node, hammer_node_t oparent,
1215                         hammer_node_t nparent, int nindex);
1216 void    hammer_cursor_inserted_element(hammer_node_t node, int index);
1217 void    hammer_cursor_deleted_element(hammer_node_t node, int index);
1218 void    hammer_cursor_invalidate_cache(hammer_cursor_t cursor);
1219
1220 int     hammer_btree_lookup(hammer_cursor_t cursor);
1221 int     hammer_btree_first(hammer_cursor_t cursor);
1222 int     hammer_btree_last(hammer_cursor_t cursor);
1223 int     hammer_btree_extract(hammer_cursor_t cursor, int flags);
1224 int     hammer_btree_iterate(hammer_cursor_t cursor);
1225 int     hammer_btree_iterate_reverse(hammer_cursor_t cursor);
1226 int     hammer_btree_insert(hammer_cursor_t cursor,
1227                             hammer_btree_leaf_elm_t elm, int *doprop);
1228 int     hammer_btree_delete(hammer_cursor_t cursor);
1229 void    hammer_btree_do_propagation(hammer_cursor_t cursor,
1230                             hammer_pseudofs_inmem_t pfsm,
1231                             hammer_btree_leaf_elm_t leaf);
1232 int     hammer_btree_cmp(hammer_base_elm_t key1, hammer_base_elm_t key2);
1233 int     hammer_btree_chkts(hammer_tid_t ts, hammer_base_elm_t key);
1234 int     hammer_btree_correct_rhb(hammer_cursor_t cursor, hammer_tid_t tid);
1235 int     hammer_btree_correct_lhb(hammer_cursor_t cursor, hammer_tid_t tid);
1236
1237 int     btree_set_parent(hammer_transaction_t trans, hammer_node_t node,
1238                         hammer_btree_elm_t elm);
1239 void    hammer_node_lock_init(hammer_node_lock_t parent, hammer_node_t node);
1240 void    hammer_btree_lcache_init(hammer_mount_t hmp, hammer_node_lock_t lcache,
1241                         int depth);
1242 void    hammer_btree_lcache_free(hammer_mount_t hmp, hammer_node_lock_t lcache);
1243 int     hammer_btree_lock_children(hammer_cursor_t cursor, int depth,
1244                         hammer_node_lock_t parent,
1245                         hammer_node_lock_t lcache);
1246 void    hammer_btree_lock_copy(hammer_cursor_t cursor,
1247                         hammer_node_lock_t parent);
1248 int     hammer_btree_sync_copy(hammer_cursor_t cursor,
1249                         hammer_node_lock_t parent);
1250 void    hammer_btree_unlock_children(hammer_mount_t hmp,
1251                         hammer_node_lock_t parent,
1252                         hammer_node_lock_t lcache);
1253 int     hammer_btree_search_node(hammer_base_elm_t elm, hammer_node_ondisk_t node);
1254 hammer_node_t hammer_btree_get_parent(hammer_transaction_t trans,
1255                         hammer_node_t node, int *parent_indexp,
1256                         int *errorp, int try_exclusive);
1257
1258 void    hammer_print_btree_node(hammer_node_ondisk_t ondisk);
1259 void    hammer_print_btree_elm(hammer_btree_elm_t elm, u_int8_t type, int i);
1260
1261 void    *hammer_bread(struct hammer_mount *hmp, hammer_off_t off,
1262                         int *errorp, struct hammer_buffer **bufferp);
1263 void    *hammer_bnew(struct hammer_mount *hmp, hammer_off_t off,
1264                         int *errorp, struct hammer_buffer **bufferp);
1265 void    *hammer_bread_ext(struct hammer_mount *hmp, hammer_off_t off, int bytes,
1266                         int *errorp, struct hammer_buffer **bufferp);
1267 void    *hammer_bnew_ext(struct hammer_mount *hmp, hammer_off_t off, int bytes,
1268                         int *errorp, struct hammer_buffer **bufferp);
1269
1270 hammer_volume_t hammer_get_root_volume(hammer_mount_t hmp, int *errorp);
1271
1272 hammer_volume_t hammer_get_volume(hammer_mount_t hmp,
1273                         int32_t vol_no, int *errorp);
1274 hammer_buffer_t hammer_get_buffer(hammer_mount_t hmp, hammer_off_t buf_offset,
1275                         int bytes, int isnew, int *errorp);
1276 void            hammer_sync_buffers(hammer_mount_t hmp,
1277                         hammer_off_t base_offset, int bytes);
1278 int             hammer_del_buffers(hammer_mount_t hmp,
1279                         hammer_off_t base_offset,
1280                         hammer_off_t zone2_offset, int bytes,
1281                         int report_conflicts);
1282
1283 int             hammer_ref_volume(hammer_volume_t volume);
1284 int             hammer_ref_buffer(hammer_buffer_t buffer);
1285 void            hammer_flush_buffer_nodes(hammer_buffer_t buffer);
1286
1287 void            hammer_rel_volume(hammer_volume_t volume, int locked);
1288 void            hammer_rel_buffer(hammer_buffer_t buffer, int locked);
1289
1290 int             hammer_vfs_export(struct mount *mp, int op,
1291                         const struct export_args *export);
1292 hammer_node_t   hammer_get_node(hammer_transaction_t trans,
1293                         hammer_off_t node_offset, int isnew, int *errorp);
1294 void            hammer_ref_node(hammer_node_t node);
1295 hammer_node_t   hammer_ref_node_safe(hammer_transaction_t trans,
1296                         hammer_node_cache_t cache, int *errorp);
1297 void            hammer_rel_node(hammer_node_t node);
1298 void            hammer_delete_node(hammer_transaction_t trans,
1299                         hammer_node_t node);
1300 void            hammer_cache_node(hammer_node_cache_t cache,
1301                         hammer_node_t node);
1302 void            hammer_uncache_node(hammer_node_cache_t cache);
1303 void            hammer_flush_node(hammer_node_t node, int locked);
1304
1305 void hammer_dup_buffer(struct hammer_buffer **bufferp,
1306                         struct hammer_buffer *buffer);
1307 hammer_node_t hammer_alloc_btree(hammer_transaction_t trans,
1308                         hammer_off_t hint, int *errorp);
1309 void *hammer_alloc_data(hammer_transaction_t trans, int32_t data_len,
1310                         u_int16_t rec_type, hammer_off_t *data_offsetp,
1311                         struct hammer_buffer **data_bufferp,
1312                         hammer_off_t hint, int *errorp);
1313
1314 int hammer_generate_undo(hammer_transaction_t trans,
1315                         hammer_off_t zone1_offset, void *base, int len);
1316 int hammer_generate_redo(hammer_transaction_t trans, hammer_inode_t ip,
1317                         hammer_off_t file_offset, u_int32_t flags,
1318                         void *base, int len);
1319 void hammer_generate_redo_sync(hammer_transaction_t trans);
1320 void hammer_redo_fifo_start_flush(hammer_inode_t ip);
1321 void hammer_redo_fifo_end_flush(hammer_inode_t ip);
1322
1323 void hammer_format_undo(void *base, u_int32_t seqno);
1324 int hammer_upgrade_undo_4(hammer_transaction_t trans);
1325
1326 void hammer_put_volume(struct hammer_volume *volume, int flush);
1327 void hammer_put_buffer(struct hammer_buffer *buffer, int flush);
1328
1329 hammer_off_t hammer_freemap_alloc(hammer_transaction_t trans,
1330                         hammer_off_t owner, int *errorp);
1331 void hammer_freemap_free(hammer_transaction_t trans, hammer_off_t phys_offset,
1332                         hammer_off_t owner, int *errorp);
1333 int _hammer_checkspace(hammer_mount_t hmp, int slop, int64_t *resp);
1334 hammer_off_t hammer_blockmap_alloc(hammer_transaction_t trans, int zone,
1335                         int bytes, hammer_off_t hint, int *errorp);
1336 hammer_reserve_t hammer_blockmap_reserve(hammer_mount_t hmp, int zone,
1337                         int bytes, hammer_off_t *zone_offp, int *errorp);
1338 hammer_reserve_t hammer_blockmap_reserve_dedup(hammer_mount_t hmp, int zone,
1339                         int bytes, hammer_off_t zone_offset, int *errorp);
1340 void hammer_blockmap_reserve_complete(hammer_mount_t hmp,
1341                         hammer_reserve_t resv);
1342 void hammer_reserve_clrdelay(hammer_mount_t hmp, hammer_reserve_t resv);
1343 void hammer_blockmap_free(hammer_transaction_t trans,
1344                         hammer_off_t bmap_off, int bytes);
1345 int hammer_blockmap_dedup(hammer_transaction_t trans,
1346                         hammer_off_t bmap_off, int bytes);
1347 int hammer_blockmap_finalize(hammer_transaction_t trans,
1348                         hammer_reserve_t resv,
1349                         hammer_off_t bmap_off, int bytes);
1350 int hammer_blockmap_getfree(hammer_mount_t hmp, hammer_off_t bmap_off,
1351                         int *curp, int *errorp);
1352 hammer_off_t hammer_blockmap_lookup(hammer_mount_t hmp, hammer_off_t bmap_off,
1353                         int *errorp);
1354 hammer_off_t hammer_undo_lookup(hammer_mount_t hmp, hammer_off_t bmap_off,
1355                         int *errorp);
1356 int64_t hammer_undo_used(hammer_transaction_t trans);
1357 int64_t hammer_undo_space(hammer_transaction_t trans);
1358 int64_t hammer_undo_max(hammer_mount_t hmp);
1359 int hammer_undo_reclaim(hammer_io_t io);
1360
1361 void hammer_start_transaction(struct hammer_transaction *trans,
1362                               struct hammer_mount *hmp);
1363 void hammer_simple_transaction(struct hammer_transaction *trans,
1364                               struct hammer_mount *hmp);
1365 void hammer_start_transaction_fls(struct hammer_transaction *trans,
1366                                   struct hammer_mount *hmp);
1367 void hammer_done_transaction(struct hammer_transaction *trans);
1368 hammer_tid_t hammer_alloc_tid(hammer_mount_t hmp, int count);
1369
1370 void hammer_modify_inode(hammer_transaction_t trans, hammer_inode_t ip, int flags);
1371 void hammer_flush_inode(hammer_inode_t ip, int flags);
1372 void hammer_flush_inode_done(hammer_inode_t ip, int error);
1373 void hammer_wait_inode(hammer_inode_t ip);
1374
1375 int  hammer_create_inode(struct hammer_transaction *trans, struct vattr *vap,
1376                         struct ucred *cred, struct hammer_inode *dip,
1377                         const char *name, int namelen,
1378                         hammer_pseudofs_inmem_t pfsm,
1379                         struct hammer_inode **ipp);
1380 void hammer_rel_inode(hammer_inode_t ip, int flush);
1381 int hammer_reload_inode(hammer_inode_t ip, void *arg __unused);
1382 int hammer_ino_rb_compare(hammer_inode_t ip1, hammer_inode_t ip2);
1383 int hammer_redo_rb_compare(hammer_inode_t ip1, hammer_inode_t ip2);
1384 int hammer_destroy_inode_callback(hammer_inode_t ip, void *data __unused);
1385
1386 int hammer_sync_inode(hammer_transaction_t trans, hammer_inode_t ip);
1387 void hammer_test_inode(hammer_inode_t dip);
1388 void hammer_inode_unloadable_check(hammer_inode_t ip, int getvp);
1389 int hammer_update_atime_quick(hammer_inode_t ip);
1390
1391 int  hammer_ip_add_directory(struct hammer_transaction *trans,
1392                         hammer_inode_t dip, const char *name, int bytes,
1393                         hammer_inode_t nip);
1394 int  hammer_ip_del_directory(struct hammer_transaction *trans,
1395                         hammer_cursor_t cursor, hammer_inode_t dip,
1396                         hammer_inode_t ip);
1397 void hammer_ip_replace_bulk(hammer_mount_t hmp, hammer_record_t record);
1398 hammer_record_t hammer_ip_add_bulk(hammer_inode_t ip, off_t file_offset,
1399                         void *data, int bytes, int *errorp);
1400 int  hammer_ip_frontend_trunc(struct hammer_inode *ip, off_t file_size);
1401 int  hammer_ip_add_record(struct hammer_transaction *trans,
1402                         hammer_record_t record);
1403 int  hammer_ip_delete_range(hammer_cursor_t cursor, hammer_inode_t ip,
1404                         int64_t ran_beg, int64_t ran_end, int truncating);
1405 int  hammer_ip_delete_clean(hammer_cursor_t cursor, hammer_inode_t ip,
1406                         int *countp);
1407 int  hammer_ip_sync_data(hammer_cursor_t cursor, hammer_inode_t ip,
1408                         int64_t offset, void *data, int bytes);
1409 int  hammer_ip_sync_record(hammer_transaction_t trans, hammer_record_t rec);
1410 int  hammer_ip_sync_record_cursor(hammer_cursor_t cursor, hammer_record_t rec);
1411 hammer_pseudofs_inmem_t  hammer_load_pseudofs(hammer_transaction_t trans,
1412                         u_int32_t localization, int *errorp);
1413 int  hammer_mkroot_pseudofs(hammer_transaction_t trans, struct ucred *cred,
1414                         hammer_pseudofs_inmem_t pfsm);
1415 int  hammer_save_pseudofs(hammer_transaction_t trans,
1416                         hammer_pseudofs_inmem_t pfsm);
1417 int  hammer_unload_pseudofs(hammer_transaction_t trans, u_int32_t localization);
1418 void hammer_rel_pseudofs(hammer_mount_t hmp, hammer_pseudofs_inmem_t pfsm);
1419 int hammer_ioctl(hammer_inode_t ip, u_long com, caddr_t data, int fflag,
1420                         struct ucred *cred);
1421
1422 void hammer_io_init(hammer_io_t io, hammer_volume_t volume,
1423                         enum hammer_io_type type);
1424 int hammer_io_read(struct vnode *devvp, struct hammer_io *io, int limit);
1425 void hammer_io_advance(struct hammer_io *io);
1426 int hammer_io_new(struct vnode *devvp, struct hammer_io *io);
1427 int hammer_io_inval(hammer_volume_t volume, hammer_off_t zone2_offset);
1428 struct buf *hammer_io_release(struct hammer_io *io, int flush);
1429 void hammer_io_flush(struct hammer_io *io, int reclaim);
1430 void hammer_io_wait(struct hammer_io *io);
1431 void hammer_io_waitdep(struct hammer_io *io);
1432 void hammer_io_wait_all(hammer_mount_t hmp, const char *ident, int doflush);
1433 int hammer_io_direct_read(hammer_mount_t hmp, struct bio *bio,
1434                         hammer_btree_leaf_elm_t leaf);
1435 int hammer_io_indirect_read(hammer_mount_t hmp, struct bio *bio,
1436                         hammer_btree_leaf_elm_t leaf);
1437 int hammer_io_direct_write(hammer_mount_t hmp, struct bio *bio,
1438                         hammer_record_t record);
1439 void hammer_io_direct_wait(hammer_record_t record);
1440 void hammer_io_direct_uncache(hammer_mount_t hmp, hammer_btree_leaf_elm_t leaf);
1441 void hammer_io_write_interlock(hammer_io_t io);
1442 void hammer_io_done_interlock(hammer_io_t io);
1443 void hammer_io_clear_modify(struct hammer_io *io, int inval);
1444 void hammer_io_clear_modlist(struct hammer_io *io);
1445 void hammer_io_flush_sync(hammer_mount_t hmp);
1446 void hammer_io_clear_error(struct hammer_io *io);
1447 void hammer_io_clear_error_noassert(struct hammer_io *io);
1448 void hammer_io_notmeta(hammer_buffer_t buffer);
1449 void hammer_io_limit_backlog(hammer_mount_t hmp);
1450
1451 void hammer_modify_volume(hammer_transaction_t trans, hammer_volume_t volume,
1452                         void *base, int len);
1453 void hammer_modify_buffer(hammer_transaction_t trans, hammer_buffer_t buffer,
1454                         void *base, int len);
1455 void hammer_modify_volume_done(hammer_volume_t volume);
1456 void hammer_modify_buffer_done(hammer_buffer_t buffer);
1457
1458 int hammer_ioc_reblock(hammer_transaction_t trans, hammer_inode_t ip,
1459                         struct hammer_ioc_reblock *reblock);
1460 int hammer_ioc_rebalance(hammer_transaction_t trans, hammer_inode_t ip,
1461                         struct hammer_ioc_rebalance *rebal);
1462 int hammer_ioc_prune(hammer_transaction_t trans, hammer_inode_t ip,
1463                         struct hammer_ioc_prune *prune);
1464 int hammer_ioc_mirror_read(hammer_transaction_t trans, hammer_inode_t ip,
1465                         struct hammer_ioc_mirror_rw *mirror);
1466 int hammer_ioc_mirror_write(hammer_transaction_t trans, hammer_inode_t ip,
1467                         struct hammer_ioc_mirror_rw *mirror);
1468 int hammer_ioc_set_pseudofs(hammer_transaction_t trans, hammer_inode_t ip,
1469                         struct ucred *cred, struct hammer_ioc_pseudofs_rw *pfs);
1470 int hammer_ioc_get_pseudofs(hammer_transaction_t trans, hammer_inode_t ip,
1471                         struct hammer_ioc_pseudofs_rw *pfs);
1472 int hammer_ioc_destroy_pseudofs(hammer_transaction_t trans, hammer_inode_t ip,
1473                         struct hammer_ioc_pseudofs_rw *pfs);
1474 int hammer_ioc_downgrade_pseudofs(hammer_transaction_t trans, hammer_inode_t ip,
1475                         struct hammer_ioc_pseudofs_rw *pfs);
1476 int hammer_ioc_upgrade_pseudofs(hammer_transaction_t trans, hammer_inode_t ip,
1477                         struct hammer_ioc_pseudofs_rw *pfs);
1478 int hammer_ioc_wait_pseudofs(hammer_transaction_t trans, hammer_inode_t ip,
1479                         struct hammer_ioc_pseudofs_rw *pfs);
1480 int hammer_ioc_volume_add(hammer_transaction_t trans, hammer_inode_t ip,
1481                         struct hammer_ioc_volume *ioc);
1482 int hammer_ioc_volume_del(hammer_transaction_t trans, hammer_inode_t ip,
1483                         struct hammer_ioc_volume *ioc);
1484 int hammer_ioc_volume_list(hammer_transaction_t trans, hammer_inode_t ip,
1485                         struct hammer_ioc_volume_list *ioc);
1486 int hammer_ioc_dedup(hammer_transaction_t trans, hammer_inode_t ip,
1487                         struct hammer_ioc_dedup *dedup);
1488
1489 int hammer_signal_check(hammer_mount_t hmp);
1490
1491 void hammer_flusher_create(hammer_mount_t hmp);
1492 void hammer_flusher_destroy(hammer_mount_t hmp);
1493 void hammer_flusher_sync(hammer_mount_t hmp);
1494 int  hammer_flusher_async(hammer_mount_t hmp, hammer_flush_group_t flg);
1495 int  hammer_flusher_async_one(hammer_mount_t hmp);
1496 int hammer_flusher_running(hammer_mount_t hmp);
1497 void hammer_flusher_wait(hammer_mount_t hmp, int seq);
1498 void hammer_flusher_wait_next(hammer_mount_t hmp);
1499 int  hammer_flusher_meta_limit(hammer_mount_t hmp);
1500 int  hammer_flusher_meta_halflimit(hammer_mount_t hmp);
1501 int  hammer_flusher_undo_exhausted(hammer_transaction_t trans, int quarter);
1502 void hammer_flusher_clean_loose_ios(hammer_mount_t hmp);
1503 void hammer_flusher_finalize(hammer_transaction_t trans, int final);
1504 int  hammer_flusher_haswork(hammer_mount_t hmp);
1505 void hammer_flusher_flush_undos(hammer_mount_t hmp, int already_flushed);
1506
1507 int hammer_recover_stage1(hammer_mount_t hmp, hammer_volume_t rootvol);
1508 int hammer_recover_stage2(hammer_mount_t hmp, hammer_volume_t rootvol);
1509 void hammer_recover_flush_buffers(hammer_mount_t hmp,
1510                         hammer_volume_t root_volume, int final);
1511
1512 void hammer_crc_set_blockmap(hammer_blockmap_t blockmap);
1513 void hammer_crc_set_volume(hammer_volume_ondisk_t ondisk);
1514 void hammer_crc_set_leaf(void *data, hammer_btree_leaf_elm_t leaf);
1515
1516 int hammer_crc_test_blockmap(hammer_blockmap_t blockmap);
1517 int hammer_crc_test_volume(hammer_volume_ondisk_t ondisk);
1518 int hammer_crc_test_btree(hammer_node_ondisk_t ondisk);
1519 int hammer_crc_test_leaf(void *data, hammer_btree_leaf_elm_t leaf);
1520 void hkprintf(const char *ctl, ...) __printflike(1, 2);
1521 udev_t hammer_fsid_to_udev(uuid_t *uuid);
1522
1523
1524 int hammer_blocksize(int64_t file_offset);
1525 int hammer_blockoff(int64_t file_offset);
1526 int64_t hammer_blockdemarc(int64_t file_offset1, int64_t file_offset2);
1527
1528 /*
1529  * Shortcut for _hammer_checkspace(), used all over the code.
1530  */
1531 static __inline int
1532 hammer_checkspace(hammer_mount_t hmp, int slop)
1533 {
1534         return(_hammer_checkspace(hmp, slop, NULL));
1535 }
1536
1537 #endif
1538
1539 static __inline void
1540 hammer_wait_mem_record(hammer_record_t record)
1541 {
1542         hammer_wait_mem_record_ident(record, "hmmwai");
1543 }
1544
1545 static __inline void
1546 hammer_lock_ex(struct hammer_lock *lock)
1547 {
1548         hammer_lock_ex_ident(lock, "hmrlck");
1549 }
1550
1551 /*
1552  * Indicate that a B-Tree node is being modified.
1553  */
1554 static __inline void
1555 hammer_modify_node_noundo(hammer_transaction_t trans, hammer_node_t node)
1556 {
1557         KKASSERT((node->flags & HAMMER_NODE_CRCBAD) == 0);
1558         hammer_modify_buffer(trans, node->buffer, NULL, 0);
1559 }
1560
1561 static __inline void
1562 hammer_modify_node_all(hammer_transaction_t trans, struct hammer_node *node)
1563 {
1564         KKASSERT((node->flags & HAMMER_NODE_CRCBAD) == 0);
1565         hammer_modify_buffer(trans, node->buffer,
1566                              node->ondisk, sizeof(*node->ondisk));
1567 }
1568
1569 static __inline void
1570 hammer_modify_node(hammer_transaction_t trans, hammer_node_t node,
1571                    void *base, int len)
1572 {
1573         hammer_crc_t *crcptr;
1574
1575         KKASSERT((char *)base >= (char *)node->ondisk &&
1576                  (char *)base + len <=
1577                     (char *)node->ondisk + sizeof(*node->ondisk));
1578         KKASSERT((node->flags & HAMMER_NODE_CRCBAD) == 0);
1579
1580         if (hammer_btree_full_undo) {
1581                 hammer_modify_node_all(trans, node);
1582         } else {
1583                 hammer_modify_buffer(trans, node->buffer, base, len);
1584                 crcptr = &node->ondisk->crc;
1585                 hammer_modify_buffer(trans, node->buffer,
1586                                      crcptr, sizeof(hammer_crc_t));
1587                 --node->buffer->io.modify_refs; /* only want one ref */
1588         }
1589 }
1590
1591 /*
1592  * Indicate that the specified modifications have been completed.
1593  *
1594  * Do not try to generate the crc here, it's very expensive to do and a
1595  * sequence of insertions or deletions can result in many calls to this
1596  * function on the same node.
1597  */
1598 static __inline void
1599 hammer_modify_node_done(hammer_node_t node)
1600 {
1601         node->flags |= HAMMER_NODE_CRCGOOD;
1602         if ((node->flags & HAMMER_NODE_NEEDSCRC) == 0) {
1603                 node->flags |= HAMMER_NODE_NEEDSCRC;
1604                 node->buffer->io.gencrc = 1;
1605                 hammer_ref_node(node);
1606         }
1607         hammer_modify_buffer_done(node->buffer);
1608 }
1609
1610 #define hammer_modify_volume_field(trans, vol, field)           \
1611         hammer_modify_volume(trans, vol, &(vol)->ondisk->field, \
1612                              sizeof((vol)->ondisk->field))
1613
1614 #define hammer_modify_node_field(trans, node, field)            \
1615         hammer_modify_node(trans, node, &(node)->ondisk->field, \
1616                              sizeof((node)->ondisk->field))
1617
1618 /*
1619  * The HAMMER_INODE_CAP_DIR_LOCAL_INO capability is set on newly
1620  * created directories for HAMMER version 2 or greater and causes
1621  * directory entries to be placed the inode localization zone in
1622  * the B-Tree instead of the misc zone.
1623  *
1624  * This greatly improves localization between directory entries and
1625  * inodes
1626  */
1627 static __inline u_int32_t
1628 hammer_dir_localization(hammer_inode_t dip)
1629 {
1630         if (dip->ino_data.cap_flags & HAMMER_INODE_CAP_DIR_LOCAL_INO)
1631                 return(HAMMER_LOCALIZE_INODE);
1632         else
1633                 return(HAMMER_LOCALIZE_MISC);
1634 }