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  * $DragonFly: src/sys/vfs/hammer/hammer.h,v 1.113 2008/07/12 02:47:39 dillon Exp $
  35  */
  36 /*
  37  * This header file contains structures used internally by the HAMMERFS
  38  * implementation.  See hammer_disk.h for on-disk structures.
  39  */
  40
  41 #include <sys/param.h>
  42 #include <sys/types.h>
  43 #include <sys/kernel.h>
  44 #include <sys/conf.h>
  45 #include <sys/systm.h>
  46 #include <sys/tree.h>
  47 #include <sys/malloc.h>
  48 #include <sys/mount.h>
  49 #include <sys/mountctl.h>
  50 #include <sys/vnode.h>
  51 #include <sys/proc.h>
  52 #include <sys/stat.h>
  53 #include <sys/globaldata.h>
  54 #include <sys/lockf.h>
  55 #include <sys/buf.h>
  56 #include <sys/queue.h>
  57 #include <sys/ktr.h>
  58 #include <sys/globaldata.h>
  59
  60 #include <sys/buf2.h>
  61 #include <sys/signal2.h>
  62 #include "hammer_disk.h"
  63 #include "hammer_mount.h"
  64 #include "hammer_ioctl.h"
  65
  66 #if defined(_KERNEL) || defined(_KERNEL_STRUCTURES)
  67
  68 MALLOC_DECLARE(M_HAMMER);
  69
  70 /*
  71  * Kernel trace
  72  */
  73 #if !defined(KTR_HAMMER)
  74 #define KTR_HAMMER      KTR_ALL
  75 #endif
  76 KTR_INFO_MASTER_EXTERN(hammer);
  77
  78 /*
  79  * Misc structures
  80  */
  81 struct hammer_mount;
  82
  83 /*
  84  * Key structure used for custom RB tree inode lookups.  This prototypes
  85  * the function hammer_ino_rb_tree_RB_LOOKUP_INFO(root, info).
  86  */
  87 typedef struct hammer_inode_info {
  88         int64_t         obj_id;         /* (key) object identifier */
  89         hammer_tid_t    obj_asof;       /* (key) snapshot transid or 0 */
  90         u_int32_t       obj_localization; /* (key) pseudo-fs */
  91         union {
  92                 struct hammer_btree_leaf_elm *leaf;
  93         } u;
  94 } *hammer_inode_info_t;
  95
  96 typedef enum hammer_transaction_type {
  97         HAMMER_TRANS_RO,
  98         HAMMER_TRANS_STD,
  99         HAMMER_TRANS_FLS
 100 } hammer_transaction_type_t;
 101
 102 /*
 103  * HAMMER Transaction tracking
 104  */
 105 struct hammer_transaction {
 106         hammer_transaction_type_t type;
 107         struct hammer_mount *hmp;
 108         hammer_tid_t    tid;
 109         u_int64_t       time;
 110         u_int32_t       time32;
 111         int             sync_lock_refs;
 112         struct hammer_volume *rootvol;
 113 };
 114
 115 typedef struct hammer_transaction *hammer_transaction_t;
 116
 117 /*
 118  * HAMMER locks
 119  */
 120 struct hammer_lock {
 121         int     refs;           /* active references delay writes */
 122         int     lockcount;      /* lock count for exclusive/shared access */
 123         int     wanted;
 124         int     exwanted;       /* number of threads waiting for ex lock */
 125         struct thread *locktd;
 126 };
 127
 128 static __inline int
 129 hammer_islocked(struct hammer_lock *lock)
 130 {
 131         return(lock->lockcount != 0);
 132 }
 133
 134 static __inline int
 135 hammer_isactive(struct hammer_lock *lock)
 136 {
 137         return(lock->refs != 0);
 138 }
 139
 140 static __inline int
 141 hammer_islastref(struct hammer_lock *lock)
 142 {
 143         return(lock->refs == 1);
 144 }
 145
 146 /*
 147  * Return if we specifically own the lock exclusively.
 148  */
 149 static __inline int
 150 hammer_lock_excl_owned(struct hammer_lock *lock, thread_t td)
 151 {
 152         if (lock->lockcount > 0 && lock->locktd == td)
 153                 return(1);
 154         return(0);
 155 }
 156
 157 /*
 158  * Flush state, used by various structures
 159  */
 160 typedef enum hammer_inode_state {
 161         HAMMER_FST_IDLE,
 162         HAMMER_FST_SETUP,
 163         HAMMER_FST_FLUSH
 164 } hammer_inode_state_t;
 165
 166 TAILQ_HEAD(hammer_record_list, hammer_record);
 167
 168 /*
 169  * Pseudo-filesystem extended data tracking
 170  */
 171 struct hammer_pfs_rb_tree;
 172 struct hammer_pseudofs_inmem;
 173 RB_HEAD(hammer_pfs_rb_tree, hammer_pseudofs_inmem);
 174 RB_PROTOTYPE2(hammer_pfs_rb_tree, hammer_pseudofs_inmem, rb_node,
 175               hammer_pfs_rb_compare, u_int32_t);
 176
 177 struct hammer_pseudofs_inmem {
 178         RB_ENTRY(hammer_pseudofs_inmem) rb_node;
 179         struct hammer_lock      lock;
 180         u_int32_t               localization;
 181         hammer_tid_t            create_tid;
 182         int                     flags;
 183         udev_t                  fsid_udev;
 184         struct hammer_pseudofs_data pfsd;
 185 };
 186
 187 typedef struct hammer_pseudofs_inmem *hammer_pseudofs_inmem_t;
 188
 189 #define HAMMER_PFSM_DELETED     0x0001
 190
 191 /*
 192  * Cache object ids.  A fixed number of objid cache structures are
 193  * created to reserve object id's for newly created files in multiples
 194  * of 100,000, localized to a particular directory, and recycled as
 195  * needed.  This allows parallel create operations in different
 196  * directories to retain fairly localized object ids which in turn
 197  * improves reblocking performance and layout.
 198  */
 199 #define OBJID_CACHE_SIZE        1024
 200 #define OBJID_CACHE_BULK        100000
 201
 202 typedef struct hammer_objid_cache {
 203         TAILQ_ENTRY(hammer_objid_cache) entry;
 204         struct hammer_inode             *dip;
 205         hammer_tid_t                    next_tid;
 206         int                             count;
 207 } *hammer_objid_cache_t;
 208
 209 /*
 210  * Associate an inode with a B-Tree node to cache search start positions
 211  */
 212 typedef struct hammer_node_cache {
 213         TAILQ_ENTRY(hammer_node_cache) entry;
 214         struct hammer_node              *node;
 215         struct hammer_inode             *ip;
 216 } *hammer_node_cache_t;
 217
 218 TAILQ_HEAD(hammer_node_cache_list, hammer_node_cache);
 219
 220 /*
 221  * Structure used to represent an inode in-memory.
 222  *
 223  * The record and data associated with an inode may be out of sync with
 224  * the disk (xDIRTY flags), or not even on the disk at all (ONDISK flag
 225  * clear).
 226  *
 227  * An inode may also hold a cache of unsynchronized records, used for
 228  * database and directories only.  Unsynchronized regular file data is
 229  * stored in the buffer cache.
 230  *
 231  * NOTE: A file which is created and destroyed within the initial
 232  * synchronization period can wind up not doing any disk I/O at all.
 233  *
 234  * Finally, an inode may cache numerous disk-referencing B-Tree cursors.
 235  */
 236 struct hammer_ino_rb_tree;
 237 struct hammer_inode;
 238 RB_HEAD(hammer_ino_rb_tree, hammer_inode);
 239 RB_PROTOTYPEX(hammer_ino_rb_tree, INFO, hammer_inode, rb_node,
 240               hammer_ino_rb_compare, hammer_inode_info_t);
 241
 242 struct hammer_rec_rb_tree;
 243 struct hammer_record;
 244 RB_HEAD(hammer_rec_rb_tree, hammer_record);
 245 RB_PROTOTYPEX(hammer_rec_rb_tree, INFO, hammer_record, rb_node,
 246               hammer_rec_rb_compare, hammer_btree_leaf_elm_t);
 247
 248 TAILQ_HEAD(hammer_node_list, hammer_node);
 249
 250 struct hammer_inode {
 251         RB_ENTRY(hammer_inode)  rb_node;
 252         hammer_inode_state_t    flush_state;
 253         int                     flush_group;
 254         TAILQ_ENTRY(hammer_inode) flush_entry;
 255         struct hammer_record_list target_list;  /* target of dependant recs */
 256         int64_t                 obj_id;         /* (key) object identifier */
 257         hammer_tid_t            obj_asof;       /* (key) snapshot or 0 */
 258         u_int32_t               obj_localization; /* (key) pseudo-fs */
 259         struct hammer_mount     *hmp;
 260         hammer_objid_cache_t    objid_cache;
 261         int                     flags;
 262         int                     error;          /* flush error */
 263         int                     cursor_ip_refs; /* sanity */
 264         int                     rsv_recs;
 265         struct vnode            *vp;
 266         hammer_pseudofs_inmem_t pfsm;
 267         struct lockf            advlock;
 268         struct hammer_lock      lock;           /* sync copy interlock */
 269         off_t                   trunc_off;
 270         struct hammer_btree_leaf_elm ino_leaf;  /* in-memory cache */
 271         struct hammer_inode_data ino_data;      /* in-memory cache */
 272         struct hammer_rec_rb_tree rec_tree;     /* in-memory cache */
 273         struct hammer_node_cache cache[2];      /* search initiate cache */
 274
 275         /*
 276          * When a demark is created to synchronize an inode to
 277          * disk, certain fields are copied so the front-end VOPs
 278          * can continue to run in parallel with the synchronization
 279          * occuring in the background.
 280          */
 281         int             sync_flags;             /* to-sync flags cache */
 282         off_t           sync_trunc_off;         /* to-sync truncation */
 283         off_t           save_trunc_off;         /* write optimization */
 284         struct hammer_btree_leaf_elm sync_ino_leaf; /* to-sync cache */
 285         struct hammer_inode_data sync_ino_data; /* to-sync cache */
 286 };
 287
 288 typedef struct hammer_inode *hammer_inode_t;
 289
 290 #define VTOI(vp)        ((struct hammer_inode *)(vp)->v_data)
 291
 292 #define HAMMER_INODE_DDIRTY     0x0001  /* in-memory ino_data is dirty */
 293                                         /* (not including atime/mtime) */
 294 #define HAMMER_INODE_RSV_INODES 0x0002  /* hmp->rsv_inodes bumped */
 295 #define HAMMER_INODE_UNUSED0004 0x0004
 296 #define HAMMER_INODE_XDIRTY     0x0008  /* in-memory records */
 297 #define HAMMER_INODE_ONDISK     0x0010  /* inode is on-disk (else not yet) */
 298 #define HAMMER_INODE_FLUSH      0x0020  /* flush on last ref */
 299 #define HAMMER_INODE_DELETED    0x0080  /* inode delete (backend) */
 300 #define HAMMER_INODE_DELONDISK  0x0100  /* delete synchronized to disk */
 301 #define HAMMER_INODE_RO         0x0200  /* read-only (because of as-of) */
 302 #define HAMMER_INODE_VHELD      0x0400  /* vnode held on sync */
 303 #define HAMMER_INODE_DONDISK    0x0800  /* data records may be on disk */
 304 #define HAMMER_INODE_BUFS       0x1000  /* dirty high level bps present */
 305 #define HAMMER_INODE_REFLUSH    0x2000  /* pipelined flush during flush */
 306 #define HAMMER_INODE_RECLAIM    0x4000  /* trying to reclaim */
 307 #define HAMMER_INODE_FLUSHW     0x8000  /* Someone waiting for flush */
 308
 309 #define HAMMER_INODE_TRUNCATED  0x00010000
 310 #define HAMMER_INODE_DELETING   0x00020000 /* inode delete request (frontend)*/
 311 #define HAMMER_INODE_RESIGNAL   0x00040000 /* re-signal on re-flush */
 312 #define HAMMER_INODE_ATIME      0x00100000 /* in-memory atime modified */
 313 #define HAMMER_INODE_MTIME      0x00200000 /* in-memory mtime modified */
 314 #define HAMMER_INODE_WOULDBLOCK 0x00400000 /* re-issue to new flush group */
 315
 316 #define HAMMER_INODE_MODMASK    (HAMMER_INODE_DDIRTY|                       \
 317                                  HAMMER_INODE_XDIRTY|HAMMER_INODE_BUFS|     \
 318                                  HAMMER_INODE_ATIME|HAMMER_INODE_MTIME|     \
 319                                  HAMMER_INODE_TRUNCATED|HAMMER_INODE_DELETING)
 320
 321 #define HAMMER_INODE_MODMASK_NOXDIRTY \
 322                                 (HAMMER_INODE_MODMASK & ~HAMMER_INODE_XDIRTY)
 323
 324 #define HAMMER_FLUSH_GROUP_SIZE 64
 325
 326 #define HAMMER_FLUSH_SIGNAL     0x0001
 327 #define HAMMER_FLUSH_RECURSION  0x0002
 328
 329 /*
 330  * Used by the inode reclaim code to pipeline reclaims and avoid
 331  * blowing out kernel memory or letting the flusher get too far
 332  * behind.
 333  */
 334 struct hammer_reclaim {
 335         TAILQ_ENTRY(hammer_reclaim) entry;
 336         int     okydoky;
 337 };
 338
 339 #define HAMMER_RECLAIM_FLUSH    2000
 340 #define HAMMER_RECLAIM_WAIT     4000
 341
 342 /*
 343  * Structure used to represent an unsynchronized record in-memory.  These
 344  * records typically represent directory entries.  Only non-historical
 345  * records are kept in-memory.
 346  *
 347  * Records are organized as a per-inode RB-Tree.  If the inode is not
 348  * on disk then neither are any records and the in-memory record tree
 349  * represents the entire contents of the inode.  If the inode is on disk
 350  * then the on-disk B-Tree is scanned in parallel with the in-memory
 351  * RB-Tree to synthesize the current state of the file.
 352  *
 353  * Records are also used to enforce the ordering of directory create/delete
 354  * operations.  A new inode will not be flushed to disk unless its related
 355  * directory entry is also being flushed at the same time.  A directory entry
 356  * will not be removed unless its related inode is also being removed at the
 357  * same time.
 358  */
 359 typedef enum hammer_record_type {
 360         HAMMER_MEM_RECORD_GENERAL,      /* misc record */
 361         HAMMER_MEM_RECORD_INODE,        /* inode record */
 362         HAMMER_MEM_RECORD_ADD,          /* positive memory cache record */
 363         HAMMER_MEM_RECORD_DEL,          /* negative delete-on-disk record */
 364         HAMMER_MEM_RECORD_DATA          /* bulk-data record w/on-disk ref */
 365 } hammer_record_type_t;
 366
 367 struct hammer_record {
 368         RB_ENTRY(hammer_record)         rb_node;
 369         TAILQ_ENTRY(hammer_record)      target_entry;
 370         hammer_inode_state_t            flush_state;
 371         int                             flush_group;
 372         hammer_record_type_t            type;
 373         struct hammer_lock              lock;
 374         struct hammer_reserve           *resv;
 375         struct hammer_inode             *ip;
 376         struct hammer_inode             *target_ip;
 377         struct hammer_btree_leaf_elm    leaf;
 378         union hammer_data_ondisk        *data;
 379         int                             flags;
 380 };
 381
 382 typedef struct hammer_record *hammer_record_t;
 383
 384 /*
 385  * Record flags.  Note that FE can only be set by the frontend if the
 386  * record has not been interlocked by the backend w/ BE.
 387  */
 388 #define HAMMER_RECF_ALLOCDATA           0x0001
 389 #define HAMMER_RECF_ONRBTREE            0x0002
 390 #define HAMMER_RECF_DELETED_FE          0x0004  /* deleted (frontend) */
 391 #define HAMMER_RECF_DELETED_BE          0x0008  /* deleted (backend) */
 392 #define HAMMER_RECF_UNUSED0010          0x0010
 393 #define HAMMER_RECF_INTERLOCK_BE        0x0020  /* backend interlock */
 394 #define HAMMER_RECF_WANTED              0x0040  /* wanted by the frontend */
 395 #define HAMMER_RECF_CONVERT_DELETE      0x0100 /* special case */
 396
 397 /*
 398  * hammer_delete_at_cursor() flags
 399  */
 400 #define HAMMER_DELETE_ADJUST            0x0001
 401 #define HAMMER_DELETE_DESTROY           0x0002
 402
 403 /*
 404  * In-memory structures representing on-disk structures.
 405  */
 406 struct hammer_volume;
 407 struct hammer_buffer;
 408 struct hammer_node;
 409 struct hammer_undo;
 410 struct hammer_reserve;
 411
 412 RB_HEAD(hammer_vol_rb_tree, hammer_volume);
 413 RB_HEAD(hammer_buf_rb_tree, hammer_buffer);
 414 RB_HEAD(hammer_nod_rb_tree, hammer_node);
 415 RB_HEAD(hammer_und_rb_tree, hammer_undo);
 416 RB_HEAD(hammer_res_rb_tree, hammer_reserve);
 417
 418 RB_PROTOTYPE2(hammer_vol_rb_tree, hammer_volume, rb_node,
 419               hammer_vol_rb_compare, int32_t);
 420 RB_PROTOTYPE2(hammer_buf_rb_tree, hammer_buffer, rb_node,
 421               hammer_buf_rb_compare, hammer_off_t);
 422 RB_PROTOTYPE2(hammer_nod_rb_tree, hammer_node, rb_node,
 423               hammer_nod_rb_compare, hammer_off_t);
 424 RB_PROTOTYPE2(hammer_und_rb_tree, hammer_undo, rb_node,
 425               hammer_und_rb_compare, hammer_off_t);
 426 RB_PROTOTYPE2(hammer_res_rb_tree, hammer_reserve, rb_node,
 427               hammer_res_rb_compare, hammer_off_t);
 428
 429 /*
 430  * IO management - embedded at the head of various in-memory structures
 431  *
 432  * VOLUME       - hammer_volume containing meta-data
 433  * META_BUFFER  - hammer_buffer containing meta-data
 434  * DATA_BUFFER  - hammer_buffer containing pure-data
 435  *
 436  * Dirty volume headers and dirty meta-data buffers are locked until the
 437  * flusher can sequence them out.  Dirty pure-data buffers can be written.
 438  * Clean buffers can be passively released.
 439  */
 440 typedef enum hammer_io_type {
 441         HAMMER_STRUCTURE_VOLUME,
 442         HAMMER_STRUCTURE_META_BUFFER,
 443         HAMMER_STRUCTURE_UNDO_BUFFER,
 444         HAMMER_STRUCTURE_DATA_BUFFER
 445 } hammer_io_type_t;
 446
 447 union hammer_io_structure;
 448 struct hammer_io;
 449
 450 struct worklist {
 451         LIST_ENTRY(worklist) node;
 452 };
 453
 454 TAILQ_HEAD(hammer_io_list, hammer_io);
 455 typedef struct hammer_io_list *hammer_io_list_t;
 456
 457 struct hammer_io {
 458         struct worklist         worklist;
 459         struct hammer_lock      lock;
 460         enum hammer_io_type     type;
 461         struct hammer_mount     *hmp;
 462         TAILQ_ENTRY(hammer_io)  mod_entry; /* list entry if modified */
 463         hammer_io_list_t        mod_list;
 464         struct buf              *bp;
 465         int64_t                 offset;    /* zone-2 offset */
 466         int                     bytes;     /* buffer cache buffer size */
 467         int                     loading;   /* loading/unloading interlock */
 468         int                     modify_refs;
 469
 470         u_int           modified : 1;   /* bp's data was modified */
 471         u_int           released : 1;   /* bp released (w/ B_LOCKED set) */
 472         u_int           running : 1;    /* bp write IO in progress */
 473         u_int           waiting : 1;    /* someone is waiting on us */
 474         u_int           validated : 1;  /* ondisk has been validated */
 475         u_int           waitdep : 1;    /* flush waits for dependancies */
 476         u_int           recovered : 1;  /* has recovery ref */
 477         u_int           waitmod : 1;    /* waiting for modify_refs */
 478         u_int           reclaim : 1;    /* reclaim requested */
 479         u_int           gencrc : 1;     /* crc needs to be generated */
 480 };
 481
 482 typedef struct hammer_io *hammer_io_t;
 483
 484 #define HAMMER_CLUSTER_SIZE     (64 * 1024)
 485 #if HAMMER_CLUSTER_SIZE > MAXBSIZE
 486 #undef  HAMMER_CLUSTER_SIZE
 487 #define HAMMER_CLUSTER_SIZE     MAXBSIZE
 488 #endif
 489 #define HAMMER_CLUSTER_BUFS     (HAMMER_CLUSTER_SIZE / HAMMER_BUFSIZE)
 490
 491 /*
 492  * In-memory volume representing on-disk buffer
 493  */
 494 struct hammer_volume {
 495         struct hammer_io io;
 496         RB_ENTRY(hammer_volume) rb_node;
 497         struct hammer_volume_ondisk *ondisk;
 498         int32_t vol_no;
 499         int64_t nblocks;        /* note: special calculation for statfs */
 500         int64_t buffer_base;    /* base offset of buffer 0 */
 501         hammer_off_t maxbuf_off; /* Maximum buffer offset (zone-2) */
 502         hammer_off_t maxraw_off; /* Maximum raw offset for device */
 503         char    *vol_name;
 504         struct vnode *devvp;
 505         int     vol_flags;
 506 };
 507
 508 typedef struct hammer_volume *hammer_volume_t;
 509
 510 /*
 511  * In-memory buffer (other then volume, super-cluster, or cluster),
 512  * representing an on-disk buffer.
 513  */
 514 struct hammer_buffer {
 515         struct hammer_io io;
 516         RB_ENTRY(hammer_buffer) rb_node;
 517         void *ondisk;
 518         struct hammer_volume *volume;
 519         hammer_off_t zoneX_offset;
 520         hammer_off_t zone2_offset;
 521         struct hammer_reserve *resv;
 522         struct hammer_node_list clist;
 523 };
 524
 525 typedef struct hammer_buffer *hammer_buffer_t;
 526
 527 /*
 528  * In-memory B-Tree node, representing an on-disk B-Tree node.
 529  *
 530  * This is a hang-on structure which is backed by a hammer_buffer,
 531  * indexed by a hammer_cluster, and used for fine-grained locking of
 532  * B-Tree nodes in order to properly control lock ordering.  A hammer_buffer
 533  * can contain multiple nodes representing wildly disassociated portions
 534  * of the B-Tree so locking cannot be done on a buffer-by-buffer basis.
 535  *
 536  * This structure uses a cluster-relative index to reduce the number
 537  * of layers required to access it, and also because all on-disk B-Tree
 538  * references are cluster-relative offsets.
 539  */
 540 struct hammer_node {
 541         struct hammer_lock      lock;           /* node-by-node lock */
 542         TAILQ_ENTRY(hammer_node) entry;         /* per-buffer linkage */
 543         RB_ENTRY(hammer_node)   rb_node;        /* per-cluster linkage */
 544         hammer_off_t            node_offset;    /* full offset spec */
 545         struct hammer_mount     *hmp;
 546         struct hammer_buffer    *buffer;        /* backing buffer */
 547         hammer_node_ondisk_t    ondisk;         /* ptr to on-disk structure */
 548         TAILQ_HEAD(, hammer_cursor) cursor_list;  /* deadlock recovery */
 549         struct hammer_node_cache_list cache_list; /* passive caches */
 550         int                     flags;
 551         int                     loading;        /* load interlock */
 552 };
 553
 554 #define HAMMER_NODE_DELETED     0x0001
 555 #define HAMMER_NODE_FLUSH       0x0002
 556 #define HAMMER_NODE_CRCGOOD     0x0004
 557 #define HAMMER_NODE_NEEDSCRC    0x0008
 558 #define HAMMER_NODE_NEEDSMIRROR 0x0010
 559
 560 typedef struct hammer_node      *hammer_node_t;
 561
 562 /*
 563  * List of locked nodes.
 564  */
 565 struct hammer_node_locklist {
 566         struct hammer_node_locklist *next;
 567         hammer_node_t   node;
 568 };
 569
 570 typedef struct hammer_node_locklist *hammer_node_locklist_t;
 571
 572
 573 /*
 574  * Common I/O management structure - embedded in in-memory structures
 575  * which are backed by filesystem buffers.
 576  */
 577 union hammer_io_structure {
 578         struct hammer_io        io;
 579         struct hammer_volume    volume;
 580         struct hammer_buffer    buffer;
 581 };
 582
 583 typedef union hammer_io_structure *hammer_io_structure_t;
 584
 585 /*
 586  * The reserve structure prevents the blockmap from allocating
 587  * out of a reserved bigblock.  Such reservations are used by
 588  * the direct-write mechanism.
 589  *
 590  * The structure is also used to hold off on reallocations of
 591  * big blocks from the freemap until flush dependancies have
 592  * been dealt with.
 593  */
 594 struct hammer_reserve {
 595         RB_ENTRY(hammer_reserve) rb_node;
 596         TAILQ_ENTRY(hammer_reserve) delay_entry;
 597         int             flush_group;
 598         int             flags;
 599         int             refs;
 600         int             zone;
 601         int             append_off;
 602         hammer_off_t    zone_offset;
 603 };
 604
 605 typedef struct hammer_reserve *hammer_reserve_t;
 606
 607 #define HAMMER_RESF_ONDELAY     0x0001
 608
 609 #include "hammer_cursor.h"
 610
 611 /*
 612  * The undo structure tracks recent undos to avoid laying down duplicate
 613  * undos within a flush group, saving us a significant amount of overhead.
 614  *
 615  * This is strictly a heuristic.
 616  */
 617 #define HAMMER_MAX_UNDOS        1024
 618 #define HAMMER_MAX_FLUSHERS     4
 619
 620 struct hammer_undo {
 621         RB_ENTRY(hammer_undo)   rb_node;
 622         TAILQ_ENTRY(hammer_undo) lru_entry;
 623         hammer_off_t            offset;
 624         int                     bytes;
 625 };
 626
 627 typedef struct hammer_undo *hammer_undo_t;
 628
 629 struct hammer_flusher_info;
 630
 631 struct hammer_flusher {
 632         int             signal;         /* flusher thread sequencer */
 633         int             act;            /* currently active flush group */
 634         int             done;           /* set to act when complete */
 635         int             next;           /* next flush group */
 636         int             group_lock;     /* lock sequencing of the next flush */
 637         int             exiting;        /* request master exit */
 638         int             count;          /* number of slave flushers */
 639         int             running;        /* number of slave flushers running */
 640         thread_t        td;             /* master flusher thread */
 641         hammer_tid_t    tid;            /* last flushed transaction id */
 642         int             finalize_want;          /* serialize finalization */
 643         struct hammer_lock finalize_lock;       /* serialize finalization */
 644         struct hammer_transaction trans;        /* shared transaction */
 645         struct hammer_flusher_info *info[HAMMER_MAX_FLUSHERS];
 646 };
 647
 648 /*
 649  * Internal hammer mount data structure
 650  */
 651 struct hammer_mount {
 652         struct mount *mp;
 653         /*struct vnode *rootvp;*/
 654         struct hammer_ino_rb_tree rb_inos_root;
 655         struct hammer_vol_rb_tree rb_vols_root;
 656         struct hammer_nod_rb_tree rb_nods_root;
 657         struct hammer_und_rb_tree rb_undo_root;
 658         struct hammer_res_rb_tree rb_resv_root;
 659         struct hammer_buf_rb_tree rb_bufs_root;
 660         struct hammer_pfs_rb_tree rb_pfsm_root;
 661         struct hammer_volume *rootvol;
 662         struct hammer_base_elm root_btree_beg;
 663         struct hammer_base_elm root_btree_end;
 664         int     flags;
 665         int     hflags;
 666         int     ronly;
 667         int     nvolumes;
 668         int     volume_iterator;
 669         int     masterid;       /* -1 or 0-15 - clustering and mirroring */
 670         int     rsv_inodes;     /* reserved space due to dirty inodes */
 671         int64_t rsv_databytes;  /* reserved space due to record data */
 672         int     rsv_recs;       /* reserved space due to dirty records */
 673         int     rsv_fromdelay;  /* bigblocks reserved due to flush delay */
 674         int     last_newrecords;
 675         int     count_newrecords;
 676
 677         int     inode_reclaims; /* inodes pending reclaim by flusher */
 678         int     count_inodes;   /* total number of inodes */
 679         int     count_iqueued;  /* inodes queued to flusher */
 680
 681         struct hammer_flusher flusher;
 682
 683         u_int   check_interrupt;
 684         uuid_t  fsid;
 685         struct hammer_io_list volu_list;        /* dirty undo buffers */
 686         struct hammer_io_list undo_list;        /* dirty undo buffers */
 687         struct hammer_io_list data_list;        /* dirty data buffers */
 688         struct hammer_io_list alt_data_list;    /* dirty data buffers */
 689         struct hammer_io_list meta_list;        /* dirty meta bufs    */
 690         struct hammer_io_list lose_list;        /* loose buffers      */
 691         int     locked_dirty_space;             /* meta/volu count    */
 692         int     io_running_space;
 693         int     objid_cache_count;
 694         hammer_tid_t    asof;                   /* snapshot mount */
 695         hammer_off_t    next_tid;
 696         int64_t copy_stat_freebigblocks;        /* number of free bigblocks */
 697
 698         u_int32_t namekey_iterator;
 699         struct netexport export;
 700         struct hammer_lock sync_lock;
 701         struct hammer_lock free_lock;
 702         struct hammer_lock undo_lock;
 703         struct hammer_lock blkmap_lock;
 704         struct hammer_blockmap  blockmap[HAMMER_MAX_ZONES];
 705         struct hammer_undo      undos[HAMMER_MAX_UNDOS];
 706         int                     undo_alloc;
 707         TAILQ_HEAD(, hammer_undo)  undo_lru_list;
 708         TAILQ_HEAD(, hammer_inode) flush_list;
 709         TAILQ_HEAD(, hammer_reserve) delay_list;
 710         TAILQ_HEAD(, hammer_objid_cache) objid_cache_list;
 711         TAILQ_HEAD(, hammer_reclaim) reclaim_list;
 712 };
 713
 714 typedef struct hammer_mount     *hammer_mount_t;
 715
 716 #define HAMMER_MOUNT_UNUSED0001 0x0001
 717
 718 struct hammer_sync_info {
 719         int error;
 720         int waitfor;
 721 };
 722
 723 #endif
 724
 725 /*
 726  * checkspace slop (8MB chunks), higher numbers are more conservative.
 727  */
 728 #define HAMMER_CHKSPC_REBLOCK   25
 729 #define HAMMER_CHKSPC_MIRROR    20
 730 #define HAMMER_CHKSPC_WRITE     20
 731 #define HAMMER_CHKSPC_CREATE    20
 732 #define HAMMER_CHKSPC_REMOVE    10
 733
 734 #if defined(_KERNEL)
 735
 736 extern struct vop_ops hammer_vnode_vops;
 737 extern struct vop_ops hammer_spec_vops;
 738 extern struct vop_ops hammer_fifo_vops;
 739 extern struct bio_ops hammer_bioops;
 740
 741 extern int hammer_debug_io;
 742 extern int hammer_debug_general;
 743 extern int hammer_debug_debug;
 744 extern int hammer_debug_inode;
 745 extern int hammer_debug_locks;
 746 extern int hammer_debug_btree;
 747 extern int hammer_debug_tid;
 748 extern int hammer_debug_recover;
 749 extern int hammer_debug_recover_faults;
 750 extern int hammer_debug_cluster_enable;
 751 extern int hammer_count_inodes;
 752 extern int hammer_count_iqueued;
 753 extern int hammer_count_reclaiming;
 754 extern int hammer_count_records;
 755 extern int hammer_count_record_datas;
 756 extern int hammer_count_volumes;
 757 extern int hammer_count_buffers;
 758 extern int hammer_count_nodes;
 759 extern int64_t hammer_count_extra_space_used;
 760 extern int64_t hammer_stats_btree_lookups;
 761 extern int64_t hammer_stats_btree_searches;
 762 extern int64_t hammer_stats_btree_inserts;
 763 extern int64_t hammer_stats_btree_deletes;
 764 extern int64_t hammer_stats_btree_elements;
 765 extern int64_t hammer_stats_btree_splits;
 766 extern int64_t hammer_stats_btree_iterations;
 767 extern int64_t hammer_stats_record_iterations;
 768 extern int hammer_count_dirtybufspace;
 769 extern int hammer_count_refedbufs;
 770 extern int hammer_count_reservations;
 771 extern int hammer_count_io_running_read;
 772 extern int hammer_count_io_running_write;
 773 extern int hammer_count_io_locked;
 774 extern int hammer_limit_dirtybufspace;
 775 extern int hammer_limit_iqueued;
 776 extern int hammer_limit_recs;
 777 extern int hammer_bio_count;
 778 extern int hammer_verify_zone;
 779 extern int hammer_write_mode;
 780 extern int64_t hammer_contention_count;
 781
 782 int     hammer_vop_inactive(struct vop_inactive_args *);
 783 int     hammer_vop_reclaim(struct vop_reclaim_args *);
 784 int     hammer_get_vnode(struct hammer_inode *ip, struct vnode **vpp);
 785 struct hammer_inode *hammer_get_inode(hammer_transaction_t trans,
 786                         hammer_inode_t dip, int64_t obj_id,
 787                         hammer_tid_t asof, u_int32_t localization,
 788                         int flags, int *errorp);
 789 void    hammer_scan_inode_snapshots(hammer_mount_t hmp,
 790                         hammer_inode_info_t iinfo,
 791                         int (*callback)(hammer_inode_t ip, void *data),
 792                         void *data);
 793 void    hammer_put_inode(struct hammer_inode *ip);
 794 void    hammer_put_inode_ref(struct hammer_inode *ip);
 795 void    hammer_inode_waitreclaims(hammer_mount_t hmp);
 796
 797 int     hammer_unload_volume(hammer_volume_t volume, void *data __unused);
 798 int     hammer_adjust_volume_mode(hammer_volume_t volume, void *data __unused);
 799
 800 int     hammer_unload_buffer(hammer_buffer_t buffer, void *data __unused);
 801 int     hammer_install_volume(hammer_mount_t hmp, const char *volname);
 802
 803 int     hammer_ip_lookup(hammer_cursor_t cursor);
 804 int     hammer_ip_first(hammer_cursor_t cursor);
 805 int     hammer_ip_next(hammer_cursor_t cursor);
 806 int     hammer_ip_resolve_data(hammer_cursor_t cursor);
 807 int     hammer_ip_delete_record(hammer_cursor_t cursor, hammer_inode_t ip,
 808                         hammer_tid_t tid);
 809 int     hammer_delete_at_cursor(hammer_cursor_t cursor, int delete_flags,
 810                         hammer_tid_t delete_tid, u_int32_t delete_ts,
 811                         int track, int64_t *stat_bytes);
 812 int     hammer_ip_check_directory_empty(hammer_transaction_t trans,
 813                         hammer_inode_t ip);
 814 int     hammer_sync_hmp(hammer_mount_t hmp, int waitfor);
 815 int     hammer_queue_inodes_flusher(hammer_mount_t hmp, int waitfor);
 816
 817 hammer_record_t
 818         hammer_alloc_mem_record(hammer_inode_t ip, int data_len);
 819 void    hammer_flush_record_done(hammer_record_t record, int error);
 820 void    hammer_wait_mem_record_ident(hammer_record_t record, const char *ident);
 821 void    hammer_rel_mem_record(hammer_record_t record);
 822
 823 int     hammer_cursor_up(hammer_cursor_t cursor);
 824 int     hammer_cursor_up_locked(hammer_cursor_t cursor);
 825 int     hammer_cursor_down(hammer_cursor_t cursor);
 826 int     hammer_cursor_upgrade(hammer_cursor_t cursor);
 827 int     hammer_cursor_upgrade_node(hammer_cursor_t cursor);
 828 void    hammer_cursor_downgrade(hammer_cursor_t cursor);
 829 int     hammer_cursor_seek(hammer_cursor_t cursor, hammer_node_t node,
 830                         int index);
 831 void    hammer_lock_ex_ident(struct hammer_lock *lock, const char *ident);
 832 int     hammer_lock_ex_try(struct hammer_lock *lock);
 833 void    hammer_lock_sh(struct hammer_lock *lock);
 834 int     hammer_lock_sh_try(struct hammer_lock *lock);
 835 int     hammer_lock_upgrade(struct hammer_lock *lock);
 836 void    hammer_lock_downgrade(struct hammer_lock *lock);
 837 int     hammer_lock_status(struct hammer_lock *lock);
 838 void    hammer_unlock(struct hammer_lock *lock);
 839 void    hammer_ref(struct hammer_lock *lock);
 840 void    hammer_unref(struct hammer_lock *lock);
 841
 842 void    hammer_sync_lock_ex(hammer_transaction_t trans);
 843 void    hammer_sync_lock_sh(hammer_transaction_t trans);
 844 int     hammer_sync_lock_sh_try(hammer_transaction_t trans);
 845 void    hammer_sync_unlock(hammer_transaction_t trans);
 846
 847 u_int32_t hammer_to_unix_xid(uuid_t *uuid);
 848 void hammer_guid_to_uuid(uuid_t *uuid, u_int32_t guid);
 849 void    hammer_time_to_timespec(u_int64_t xtime, struct timespec *ts);
 850 u_int64_t hammer_timespec_to_time(struct timespec *ts);
 851 hammer_tid_t hammer_str_to_tid(const char *str, int *ispfs,
 852                         u_int32_t *localizationp);
 853 hammer_tid_t hammer_alloc_objid(hammer_mount_t hmp, hammer_inode_t dip);
 854 void hammer_clear_objid(hammer_inode_t dip);
 855 void hammer_destroy_objid_cache(hammer_mount_t hmp);
 856
 857 int hammer_enter_undo_history(hammer_mount_t hmp, hammer_off_t offset,
 858                         int bytes);
 859 void hammer_clear_undo_history(hammer_mount_t hmp);
 860 enum vtype hammer_get_vnode_type(u_int8_t obj_type);
 861 int hammer_get_dtype(u_int8_t obj_type);
 862 u_int8_t hammer_get_obj_type(enum vtype vtype);
 863 int64_t hammer_directory_namekey(const void *name, int len);
 864 int     hammer_nohistory(hammer_inode_t ip);
 865
 866 int     hammer_init_cursor(hammer_transaction_t trans, hammer_cursor_t cursor,
 867                         hammer_node_cache_t cache, hammer_inode_t ip);
 868 void    hammer_normalize_cursor(hammer_cursor_t cursor);
 869 void    hammer_done_cursor(hammer_cursor_t cursor);
 870 int     hammer_recover_cursor(hammer_cursor_t cursor);
 871 void    hammer_unlock_cursor(hammer_cursor_t cursor, int also_ip);
 872 int     hammer_lock_cursor(hammer_cursor_t cursor, int also_ip);
 873 hammer_cursor_t hammer_push_cursor(hammer_cursor_t ocursor);
 874 void    hammer_pop_cursor(hammer_cursor_t ocursor, hammer_cursor_t ncursor);
 875
 876 void    hammer_cursor_replaced_node(hammer_node_t onode, hammer_node_t nnode);
 877 void    hammer_cursor_removed_node(hammer_node_t onode, hammer_node_t parent,
 878                         int index);
 879 void    hammer_cursor_split_node(hammer_node_t onode, hammer_node_t nnode,
 880                         int index);
 881 void    hammer_cursor_inserted_element(hammer_node_t node, int index);
 882 void    hammer_cursor_deleted_element(hammer_node_t node, int index);
 883
 884 int     hammer_btree_lookup(hammer_cursor_t cursor);
 885 int     hammer_btree_first(hammer_cursor_t cursor);
 886 int     hammer_btree_last(hammer_cursor_t cursor);
 887 int     hammer_btree_extract(hammer_cursor_t cursor, int flags);
 888 int     hammer_btree_iterate(hammer_cursor_t cursor);
 889 int     hammer_btree_iterate_reverse(hammer_cursor_t cursor);
 890 int     hammer_btree_insert(hammer_cursor_t cursor,
 891                             hammer_btree_leaf_elm_t elm, int *doprop);
 892 int     hammer_btree_delete(hammer_cursor_t cursor);
 893 void    hammer_btree_do_propagation(hammer_cursor_t cursor,
 894                             hammer_pseudofs_inmem_t pfsm,
 895                             hammer_btree_leaf_elm_t leaf);
 896 int     hammer_btree_cmp(hammer_base_elm_t key1, hammer_base_elm_t key2);
 897 int     hammer_btree_chkts(hammer_tid_t ts, hammer_base_elm_t key);
 898 int     hammer_btree_correct_rhb(hammer_cursor_t cursor, hammer_tid_t tid);
 899 int     hammer_btree_correct_lhb(hammer_cursor_t cursor, hammer_tid_t tid);
 900
 901 int     btree_set_parent(hammer_transaction_t trans, hammer_node_t node,
 902                         hammer_btree_elm_t elm);
 903 int     hammer_btree_lock_children(hammer_cursor_t cursor,
 904                         struct hammer_node_locklist **locklistp);
 905 void    hammer_btree_unlock_children(struct hammer_node_locklist **locklistp);
 906 int     hammer_btree_search_node(hammer_base_elm_t elm, hammer_node_ondisk_t node);
 907 hammer_node_t hammer_btree_get_parent(hammer_node_t node, int *parent_indexp,
 908                         int *errorp, int try_exclusive);
 909
 910 void    hammer_print_btree_node(hammer_node_ondisk_t ondisk);
 911 void    hammer_print_btree_elm(hammer_btree_elm_t elm, u_int8_t type, int i);
 912
 913 void    *hammer_bread(struct hammer_mount *hmp, hammer_off_t off,
 914                         int *errorp, struct hammer_buffer **bufferp);
 915 void    *hammer_bnew(struct hammer_mount *hmp, hammer_off_t off,
 916                         int *errorp, struct hammer_buffer **bufferp);
 917 void    *hammer_bread_ext(struct hammer_mount *hmp, hammer_off_t off, int bytes,
 918                         int *errorp, struct hammer_buffer **bufferp);
 919 void    *hammer_bnew_ext(struct hammer_mount *hmp, hammer_off_t off, int bytes,
 920                         int *errorp, struct hammer_buffer **bufferp);
 921
 922 hammer_volume_t hammer_get_root_volume(hammer_mount_t hmp, int *errorp);
 923
 924 hammer_volume_t hammer_get_volume(hammer_mount_t hmp,
 925                         int32_t vol_no, int *errorp);
 926 hammer_buffer_t hammer_get_buffer(hammer_mount_t hmp, hammer_off_t buf_offset,
 927                         int bytes, int isnew, int *errorp);
 928 void            hammer_del_buffers(hammer_mount_t hmp, hammer_off_t base_offset,
 929                         hammer_off_t zone2_offset, int bytes);
 930
 931 int             hammer_ref_volume(hammer_volume_t volume);
 932 int             hammer_ref_buffer(hammer_buffer_t buffer);
 933 void            hammer_flush_buffer_nodes(hammer_buffer_t buffer);
 934
 935 void            hammer_rel_volume(hammer_volume_t volume, int flush);
 936 void            hammer_rel_buffer(hammer_buffer_t buffer, int flush);
 937
 938 int             hammer_vfs_export(struct mount *mp, int op,
 939                         const struct export_args *export);
 940 hammer_node_t   hammer_get_node(hammer_mount_t hmp, hammer_off_t node_offset,
 941                         int isnew, int *errorp);
 942 void            hammer_ref_node(hammer_node_t node);
 943 hammer_node_t   hammer_ref_node_safe(struct hammer_mount *hmp,
 944                         hammer_node_cache_t cache, int *errorp);
 945 void            hammer_rel_node(hammer_node_t node);
 946 void            hammer_delete_node(hammer_transaction_t trans,
 947                         hammer_node_t node);
 948 void            hammer_cache_node(hammer_node_cache_t cache,
 949                         hammer_node_t node);
 950 void            hammer_uncache_node(hammer_node_cache_t cache);
 951 void            hammer_flush_node(hammer_node_t node);
 952
 953 void hammer_dup_buffer(struct hammer_buffer **bufferp,
 954                         struct hammer_buffer *buffer);
 955 hammer_node_t hammer_alloc_btree(hammer_transaction_t trans, int *errorp);
 956 void *hammer_alloc_data(hammer_transaction_t trans, int32_t data_len,
 957                         u_int16_t rec_type, hammer_off_t *data_offsetp,
 958                         struct hammer_buffer **data_bufferp, int *errorp);
 959
 960 int hammer_generate_undo(hammer_transaction_t trans, hammer_io_t io,
 961                         hammer_off_t zone1_offset, void *base, int len);
 962
 963 void hammer_put_volume(struct hammer_volume *volume, int flush);
 964 void hammer_put_buffer(struct hammer_buffer *buffer, int flush);
 965
 966 hammer_off_t hammer_freemap_alloc(hammer_transaction_t trans,
 967                         hammer_off_t owner, int *errorp);
 968 void hammer_freemap_free(hammer_transaction_t trans, hammer_off_t phys_offset,
 969                         hammer_off_t owner, int *errorp);
 970 int hammer_checkspace(hammer_mount_t hmp, int slop);
 971 hammer_off_t hammer_blockmap_alloc(hammer_transaction_t trans, int zone,
 972                         int bytes, int *errorp);
 973 hammer_reserve_t hammer_blockmap_reserve(hammer_mount_t hmp, int zone,
 974                         int bytes, hammer_off_t *zone_offp, int *errorp);
 975 void hammer_blockmap_reserve_complete(hammer_mount_t hmp,
 976                         hammer_reserve_t resv);
 977 void hammer_reserve_clrdelay(hammer_mount_t hmp, hammer_reserve_t resv);
 978 void hammer_blockmap_free(hammer_transaction_t trans,
 979                         hammer_off_t bmap_off, int bytes);
 980 void hammer_blockmap_finalize(hammer_transaction_t trans,
 981                         hammer_off_t bmap_off, int bytes);
 982 int hammer_blockmap_getfree(hammer_mount_t hmp, hammer_off_t bmap_off,
 983                         int *curp, int *errorp);
 984 hammer_off_t hammer_blockmap_lookup(hammer_mount_t hmp, hammer_off_t bmap_off,
 985                         int *errorp);
 986 hammer_off_t hammer_undo_lookup(hammer_mount_t hmp, hammer_off_t bmap_off,
 987                         int *errorp);
 988 int64_t hammer_undo_used(hammer_transaction_t trans);
 989 int64_t hammer_undo_space(hammer_transaction_t trans);
 990 int64_t hammer_undo_max(hammer_mount_t hmp);
 991
 992 void hammer_start_transaction(struct hammer_transaction *trans,
 993                               struct hammer_mount *hmp);
 994 void hammer_simple_transaction(struct hammer_transaction *trans,
 995                               struct hammer_mount *hmp);
 996 void hammer_start_transaction_fls(struct hammer_transaction *trans,
 997                                   struct hammer_mount *hmp);
 998 void hammer_done_transaction(struct hammer_transaction *trans);
 999
1000 void hammer_modify_inode(hammer_inode_t ip, int flags);
1001 void hammer_flush_inode(hammer_inode_t ip, int flags);
1002 void hammer_flush_inode_done(hammer_inode_t ip);
1003 void hammer_wait_inode(hammer_inode_t ip);
1004
1005 int  hammer_create_inode(struct hammer_transaction *trans, struct vattr *vap,
1006                         struct ucred *cred, struct hammer_inode *dip,
1007                         hammer_pseudofs_inmem_t pfsm,
1008                         struct hammer_inode **ipp);
1009 void hammer_rel_inode(hammer_inode_t ip, int flush);
1010 int hammer_reload_inode(hammer_inode_t ip, void *arg __unused);
1011 int hammer_ino_rb_compare(hammer_inode_t ip1, hammer_inode_t ip2);
1012
1013 int hammer_sync_inode(hammer_inode_t ip);
1014 void hammer_test_inode(hammer_inode_t ip);
1015 void hammer_inode_unloadable_check(hammer_inode_t ip, int getvp);
1016
1017 int  hammer_ip_add_directory(struct hammer_transaction *trans,
1018                         hammer_inode_t dip, const char *name, int bytes,
1019                         hammer_inode_t nip);
1020 int  hammer_ip_del_directory(struct hammer_transaction *trans,
1021                         hammer_cursor_t cursor, hammer_inode_t dip,
1022                         hammer_inode_t ip);
1023 hammer_record_t hammer_ip_add_bulk(hammer_inode_t ip, off_t file_offset,
1024                         void *data, int bytes, int *errorp);
1025 int  hammer_ip_frontend_trunc(struct hammer_inode *ip, off_t file_size);
1026 int  hammer_ip_add_record(struct hammer_transaction *trans,
1027                         hammer_record_t record);
1028 int  hammer_ip_delete_range(hammer_cursor_t cursor, hammer_inode_t ip,
1029                         int64_t ran_beg, int64_t ran_end, int truncating);
1030 int  hammer_ip_delete_clean(hammer_cursor_t cursor, hammer_inode_t ip,
1031                         int *countp);
1032 int  hammer_ip_sync_data(hammer_cursor_t cursor, hammer_inode_t ip,
1033                         int64_t offset, void *data, int bytes);
1034 int  hammer_ip_sync_record(hammer_transaction_t trans, hammer_record_t rec);
1035 int  hammer_ip_sync_record_cursor(hammer_cursor_t cursor, hammer_record_t rec);
1036 hammer_pseudofs_inmem_t  hammer_load_pseudofs(hammer_transaction_t trans,
1037                         u_int32_t localization, int *errorp);
1038 int  hammer_mkroot_pseudofs(hammer_transaction_t trans, struct ucred *cred,
1039                         hammer_pseudofs_inmem_t pfsm);
1040 int  hammer_save_pseudofs(hammer_transaction_t trans,
1041                         hammer_pseudofs_inmem_t pfsm);
1042 int  hammer_unload_pseudofs(hammer_transaction_t trans, u_int32_t localization);
1043 void hammer_rel_pseudofs(hammer_mount_t hmp, hammer_pseudofs_inmem_t pfsm);
1044 int hammer_ioctl(hammer_inode_t ip, u_long com, caddr_t data, int fflag,
1045                         struct ucred *cred);
1046
1047 void hammer_io_init(hammer_io_t io, hammer_mount_t hmp,
1048                         enum hammer_io_type type);
1049 int hammer_io_read(struct vnode *devvp, struct hammer_io *io,
1050                         hammer_off_t limit);
1051 int hammer_io_new(struct vnode *devvp, struct hammer_io *io);
1052 void hammer_io_inval(hammer_volume_t volume, hammer_off_t zone2_offset);
1053 struct buf *hammer_io_release(struct hammer_io *io, int flush);
1054 void hammer_io_flush(struct hammer_io *io);
1055 void hammer_io_waitdep(struct hammer_io *io);
1056 void hammer_io_wait_all(hammer_mount_t hmp, const char *ident);
1057 int hammer_io_direct_read(hammer_mount_t hmp, struct bio *bio);
1058 int hammer_io_direct_write(hammer_mount_t hmp, hammer_btree_leaf_elm_t leaf,
1059                           struct bio *bio);
1060 void hammer_io_direct_uncache(hammer_mount_t hmp, hammer_btree_leaf_elm_t leaf);
1061 void hammer_io_write_interlock(hammer_io_t io);
1062 void hammer_io_done_interlock(hammer_io_t io);
1063 void hammer_io_clear_modify(struct hammer_io *io, int inval);
1064 void hammer_io_clear_modlist(struct hammer_io *io);
1065 void hammer_modify_volume(hammer_transaction_t trans, hammer_volume_t volume,
1066                         void *base, int len);
1067 void hammer_modify_buffer(hammer_transaction_t trans, hammer_buffer_t buffer,
1068                         void *base, int len);
1069 void hammer_modify_volume_done(hammer_volume_t volume);
1070 void hammer_modify_buffer_done(hammer_buffer_t buffer);
1071
1072 int hammer_ioc_reblock(hammer_transaction_t trans, hammer_inode_t ip,
1073                         struct hammer_ioc_reblock *reblock);
1074 int hammer_ioc_prune(hammer_transaction_t trans, hammer_inode_t ip,
1075                         struct hammer_ioc_prune *prune);
1076 int hammer_ioc_mirror_read(hammer_transaction_t trans, hammer_inode_t ip,
1077                         struct hammer_ioc_mirror_rw *mirror);
1078 int hammer_ioc_mirror_write(hammer_transaction_t trans, hammer_inode_t ip,
1079                         struct hammer_ioc_mirror_rw *mirror);
1080 int hammer_ioc_set_pseudofs(hammer_transaction_t trans, hammer_inode_t ip,
1081                         struct ucred *cred, struct hammer_ioc_pseudofs_rw *pfs);
1082 int hammer_ioc_get_pseudofs(hammer_transaction_t trans, hammer_inode_t ip,
1083                         struct hammer_ioc_pseudofs_rw *pfs);
1084 int hammer_ioc_destroy_pseudofs(hammer_transaction_t trans, hammer_inode_t ip,
1085                         struct hammer_ioc_pseudofs_rw *pfs);
1086 int hammer_ioc_downgrade_pseudofs(hammer_transaction_t trans, hammer_inode_t ip,
1087                         struct hammer_ioc_pseudofs_rw *pfs);
1088 int hammer_ioc_upgrade_pseudofs(hammer_transaction_t trans, hammer_inode_t ip,
1089                         struct hammer_ioc_pseudofs_rw *pfs);
1090
1091 int hammer_signal_check(hammer_mount_t hmp);
1092
1093 void hammer_flusher_create(hammer_mount_t hmp);
1094 void hammer_flusher_destroy(hammer_mount_t hmp);
1095 void hammer_flusher_sync(hammer_mount_t hmp);
1096 int  hammer_flusher_async(hammer_mount_t hmp);
1097 void hammer_flusher_wait(hammer_mount_t hmp, int seq);
1098 int  hammer_flusher_meta_limit(hammer_mount_t hmp);
1099 int  hammer_flusher_meta_halflimit(hammer_mount_t hmp);
1100 int  hammer_flusher_undo_exhausted(hammer_transaction_t trans, int quarter);
1101 void hammer_flusher_clean_loose_ios(hammer_mount_t hmp);
1102 void hammer_flusher_finalize(hammer_transaction_t trans, int final);
1103
1104 int hammer_recover(hammer_mount_t hmp, hammer_volume_t rootvol);
1105 void hammer_recover_flush_buffers(hammer_mount_t hmp,
1106                         hammer_volume_t root_volume, int final);
1107
1108 void hammer_crc_set_blockmap(hammer_blockmap_t blockmap);
1109 void hammer_crc_set_volume(hammer_volume_ondisk_t ondisk);
1110 void hammer_crc_set_leaf(void *data, hammer_btree_leaf_elm_t leaf);
1111
1112 int hammer_crc_test_blockmap(hammer_blockmap_t blockmap);
1113 int hammer_crc_test_volume(hammer_volume_ondisk_t ondisk);
1114 int hammer_crc_test_btree(hammer_node_ondisk_t ondisk);
1115 int hammer_crc_test_leaf(void *data, hammer_btree_leaf_elm_t leaf);
1116 void hkprintf(const char *ctl, ...);
1117 udev_t hammer_fsid_to_udev(uuid_t *uuid);
1118
1119
1120 int hammer_blocksize(int64_t file_offset);
1121 int64_t hammer_blockdemarc(int64_t file_offset1, int64_t file_offset2);
1122
1123 #endif
1124
1125 static __inline void
1126 hammer_wait_mem_record(hammer_record_t record)
1127 {
1128         hammer_wait_mem_record_ident(record, "hmmwai");
1129 }
1130
1131 static __inline void
1132 hammer_lock_ex(struct hammer_lock *lock)
1133 {
1134         hammer_lock_ex_ident(lock, "hmrlck");
1135 }
1136
1137 /*
1138  * Indicate that a B-Tree node is being modified.
1139  */
1140 static __inline void
1141 hammer_modify_node_noundo(hammer_transaction_t trans, hammer_node_t node)
1142 {
1143         hammer_modify_buffer(trans, node->buffer, NULL, 0);
1144 }
1145
1146 static __inline void
1147 hammer_modify_node_all(hammer_transaction_t trans, struct hammer_node *node)
1148 {
1149         hammer_modify_buffer(trans, node->buffer,
1150                              node->ondisk, sizeof(*node->ondisk));
1151 }
1152
1153 static __inline void
1154 hammer_modify_node(hammer_transaction_t trans, hammer_node_t node,
1155                    void *base, int len)
1156 {
1157         hammer_crc_t *crcptr;
1158
1159         KKASSERT((char *)base >= (char *)node->ondisk &&
1160                  (char *)base + len <=
1161                     (char *)node->ondisk + sizeof(*node->ondisk));
1162         hammer_modify_buffer(trans, node->buffer, base, len);
1163         crcptr = &node->ondisk->crc;
1164         hammer_modify_buffer(trans, node->buffer, crcptr, sizeof(hammer_crc_t));
1165         --node->buffer->io.modify_refs; /* only want one ref */
1166 }
1167
1168 /*
1169  * Indicate that the specified modifications have been completed.
1170  *
1171  * Do not try to generate the crc here, it's very expensive to do and a
1172  * sequence of insertions or deletions can result in many calls to this
1173  * function on the same node.
1174  */
1175 static __inline void
1176 hammer_modify_node_done(hammer_node_t node)
1177 {
1178         node->flags |= HAMMER_NODE_CRCGOOD;
1179         if ((node->flags & HAMMER_NODE_NEEDSCRC) == 0) {
1180                 node->flags |= HAMMER_NODE_NEEDSCRC;
1181                 node->buffer->io.gencrc = 1;
1182                 hammer_ref_node(node);
1183         }
1184         hammer_modify_buffer_done(node->buffer);
1185 }
1186
1187 #define hammer_modify_volume_field(trans, vol, field)           \
1188         hammer_modify_volume(trans, vol, &(vol)->ondisk->field, \
1189                              sizeof((vol)->ondisk->field))
1190
1191 #define hammer_modify_node_field(trans, node, field)            \
1192         hammer_modify_node(trans, node, &(node)->ondisk->field, \
1193                              sizeof((node)->ondisk->field))
1194