db2/db_int.h

   1 /*-
   2  * See the file LICENSE for redistribution information.
   3  *
   4  * Copyright (c) 1996, 1997
   5  *      Sleepycat Software.  All rights reserved.
   6  *
   7  *      @(#)db_int.h.src        10.41 (Sleepycat) 1/8/98
   8  */
   9
  10 #ifndef _DB_INTERNAL_H_
  11 #define _DB_INTERNAL_H_
  12
  13 #include "db.h"                         /* Standard DB include file. */
  14 #include "queue.h"
  15 #include "os_func.h"
  16 #include "os_ext.h"
  17
  18 /*******************************************************
  19  * General purpose constants and macros.
  20  *******************************************************/
  21 #define UINT16_T_MAX        0xffff      /* Maximum 16 bit unsigned. */
  22 #define UINT32_T_MAX    0xffffffff      /* Maximum 32 bit unsigned. */
  23
  24 #define DB_MIN_PGSIZE   0x000200        /* Minimum page size. */
  25 #define DB_MAX_PGSIZE   0x010000        /* Maximum page size. */
  26
  27 #define DB_MINCACHE     10              /* Minimum cached pages */
  28
  29 #define MEGABYTE        1048576
  30
  31 /*
  32  * If we are unable to determine the underlying filesystem block size, use
  33  * 8K on the grounds that most OS's use less than 8K as their VM page size.
  34  */
  35 #define DB_DEF_IOSIZE   (8 * 1024)
  36
  37 /*
  38  * Aligning items to particular sizes or in pages or memory.  ALIGNP is a
  39  * separate macro, as we've had to cast the pointer to different integral
  40  * types on different architectures.
  41  *
  42  * We cast pointers into unsigned longs when manipulating them because C89
  43  * guarantees that u_long is the largest available integral type and further,
  44  * to never generate overflows.  However, neither C89 or C9X  requires that
  45  * any integer type be large enough to hold a pointer, although C9X created
  46  * the intptr_t type, which is guaranteed to hold a pointer but may or may
  47  * not exist.  At some point in the future, we should test for intptr_t and
  48  * use it where available.
  49  */
  50 #undef  ALIGNTYPE
  51 #define ALIGNTYPE               u_long
  52 #undef  ALIGNP
  53 #define ALIGNP(value, bound)    ALIGN((ALIGNTYPE)value, bound)
  54 #undef  ALIGN
  55 #define ALIGN(value, bound)     (((value) + (bound) - 1) & ~((bound) - 1))
  56
  57 /*
  58  * There are several on-page structures that are declared to have a number of
  59  * fields followed by a variable length array of items.  The structure size
  60  * without including the variable length array or the address of the first of
  61  * those elements can be found using SSZ.
  62  *
  63  * This macro can also be used to find the offset of a structure element in a
  64  * structure.  This is used in various places to copy structure elements from
  65  * unaligned memory references, e.g., pointers into a packed page.
  66  *
  67  * There are two versions because compilers object if you take the address of
  68  * an array.
  69  */
  70 #undef  SSZ
  71 #define SSZ(name, field)        ((int)&(((name *)0)->field))
  72
  73 #undef  SSZA
  74 #define SSZA(name, field)       ((int)&(((name *)0)->field[0]))
  75
  76 /* Macros to return per-process address, offsets based on shared regions. */
  77 #define R_ADDR(base, offset)    ((void *)((u_int8_t *)((base)->addr) + offset))
  78 #define R_OFFSET(base, p)       ((u_int8_t *)(p) - (u_int8_t *)(base)->addr)
  79
  80 /* Free and free-string macros that overwrite memory during debugging. */
  81 #ifdef DEBUG
  82 #undef  FREE
  83 #define FREE(p, len) {                                                  \
  84         memset(p, 0xff, len);                                           \
  85         __db_free(p);                                                   \
  86 }
  87 #undef  FREES
  88 #define FREES(p) {                                                      \
  89         FREE(p, strlen(p));                                             \
  90 }
  91 #else
  92 #undef  FREE
  93 #define FREE(p, len) {                                                  \
  94         __db_free(p);                                                   \
  95 }
  96 #undef  FREES
  97 #define FREES(p) {                                                      \
  98         __db_free(p);                                                   \
  99 }
 100 #endif
 101
 102 /* Structure used to print flag values. */
 103 typedef struct __fn {
 104         u_int32_t mask;                 /* Flag value. */
 105         const char *name;               /* Flag name. */
 106 } FN;
 107
 108 /* Set, clear and test flags. */
 109 #define F_SET(p, f)     (p)->flags |= (f)
 110 #define F_CLR(p, f)     (p)->flags &= ~(f)
 111 #define F_ISSET(p, f)   ((p)->flags & (f))
 112 #define LF_SET(f)       (flags |= (f))
 113 #define LF_CLR(f)       (flags &= ~(f))
 114 #define LF_ISSET(f)     (flags & (f))
 115
 116 /* Display separator string. */
 117 #undef  DB_LINE
 118 #define DB_LINE "=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-="
 119
 120 /* Unused, or not-used-yet variable.  "Shut that bloody compiler up!" */
 121 #define COMPQUIET(n, v) (n) = (v)
 122
 123 /*******************************************************
 124  * Files.
 125  *******************************************************/
 126 #ifndef MAXPATHLEN              /* Maximum path length. */
 127 #ifdef PATH_MAX
 128 #define MAXPATHLEN      PATH_MAX
 129 #else
 130 #define MAXPATHLEN      1024
 131 #endif
 132 #endif
 133
 134 #define PATH_DOT        "."     /* Current working directory. */
 135 #define PATH_SEPARATOR  "/"     /* Path separator character. */
 136
 137 #ifndef S_IRUSR                 /* UNIX specific file permissions. */
 138 #define S_IRUSR 0000400         /* R for owner */
 139 #define S_IWUSR 0000200         /* W for owner */
 140 #define S_IRGRP 0000040         /* R for group */
 141 #define S_IWGRP 0000020         /* W for group */
 142 #define S_IROTH 0000004         /* R for other */
 143 #define S_IWOTH 0000002         /* W for other */
 144 #endif
 145
 146 #ifndef S_ISDIR                 /* UNIX specific: directory test. */
 147 #define S_ISDIR(m)      ((m & 0170000) == 0040000)
 148 #endif
 149
 150 /*******************************************************
 151  * Mutex support.
 152  *******************************************************/
 153 typedef unsigned char tsl_t;
 154
 155
 156
 157 /*
 158  * !!!
 159  * Various systems require different alignments for mutexes (the worst we've
 160  * seen so far is 16-bytes on some HP architectures).  The mutex (tsl_t) must
 161  * be first in the db_mutex_t structure, which must itself be first in the
 162  * region.  This ensures the alignment is as returned by mmap(2), which should
 163  * be sufficient.  All other mutex users must ensure proper alignment locally.
 164  */
 165 #define MUTEX_ALIGNMENT 1
 166
 167 /*
 168  * The offset of a mutex in memory.
 169  *
 170  * !!!
 171  * Not an off_t, so backing file offsets MUST be less than 4Gb.  See the
 172  * off field of the db_mutex_t as well.
 173  */
 174 #define MUTEX_LOCK_OFFSET(a, b) ((u_int32_t)((u_int8_t *)b - (u_int8_t *)a))
 175
 176 typedef struct _db_mutex_t {
 177 #ifdef HAVE_SPINLOCKS
 178         tsl_t     tsl_resource;         /* Resource test and set. */
 179 #ifdef DEBUG
 180         u_long    pid;                  /* Lock holder: 0 or process pid. */
 181 #endif
 182 #else
 183         u_int32_t off;                  /* Backing file offset. */
 184         u_long    pid;                  /* Lock holder: 0 or process pid. */
 185 #endif
 186         u_int32_t spins;                /* Spins before block. */
 187         u_int32_t mutex_set_wait;       /* Granted after wait. */
 188         u_int32_t mutex_set_nowait;     /* Granted without waiting. */
 189 } db_mutex_t;
 190
 191 #include "mutex_ext.h"
 192
 193 /*******************************************************
 194  * Access methods.
 195  *******************************************************/
 196 /* Lock/unlock a DB thread. */
 197 #define DB_THREAD_LOCK(dbp)                                             \
 198         (F_ISSET(dbp, DB_AM_THREAD) ?                                   \
 199             __db_mutex_lock((db_mutex_t *)(dbp)->mutexp, -1) : 0)
 200 #define DB_THREAD_UNLOCK(dbp)                                           \
 201         (F_ISSET(dbp, DB_AM_THREAD) ?                                   \
 202             __db_mutex_unlock((db_mutex_t *)(dbp)->mutexp, -1) : 0)
 203
 204 /* Btree/recno local statistics structure. */
 205 struct __db_bt_lstat;   typedef struct __db_bt_lstat DB_BTREE_LSTAT;
 206 struct __db_bt_lstat {
 207         u_int32_t bt_freed;             /* Pages freed for reuse. */
 208         u_int32_t bt_pfxsaved;          /* Bytes saved by prefix compression. */
 209         u_int32_t bt_split;             /* Total number of splits. */
 210         u_int32_t bt_rootsplit;         /* Root page splits. */
 211         u_int32_t bt_fastsplit;         /* Fast splits. */
 212         u_int32_t bt_added;             /* Items added. */
 213         u_int32_t bt_deleted;           /* Items deleted. */
 214         u_int32_t bt_get;               /* Items retrieved. */
 215         u_int32_t bt_cache_hit;         /* Hits in fast-insert code. */
 216         u_int32_t bt_cache_miss;        /* Misses in fast-insert code. */
 217 };
 218
 219 /*******************************************************
 220  * Environment.
 221  *******************************************************/
 222 /* Type passed to __db_appname(). */
 223 typedef enum {
 224         DB_APP_NONE=0,                  /* No type (region). */
 225         DB_APP_DATA,                    /* Data file. */
 226         DB_APP_LOG,                     /* Log file. */
 227         DB_APP_TMP                      /* Temporary file. */
 228 } APPNAME;
 229
 230 /*******************************************************
 231  * Regions.
 232  *******************************************************/
 233 /*
 234  * The shared memory regions share an initial structure so that the general
 235  * region code can handle races between the region being deleted and other
 236  * processes waiting on the region mutex.
 237  *
 238  * !!!
 239  * Note, the mutex must be the first entry in the region; see comment above.
 240  */
 241 typedef struct _rlayout {
 242         db_mutex_t lock;                /* Region mutex. */
 243         u_int32_t  refcnt;              /* Region reference count. */
 244         size_t     size;                /* Region length. */
 245         int        majver;              /* Major version number. */
 246         int        minver;              /* Minor version number. */
 247         int        patch;               /* Patch version number. */
 248
 249 #define DB_R_DELETED    0x01            /* Region was deleted. */
 250         u_int32_t  flags;
 251 } RLAYOUT;
 252
 253 /*******************************************************
 254  * Mpool.
 255  *******************************************************/
 256 /*
 257  * File types for DB access methods.  Negative numbers are reserved to DB.
 258  */
 259 #define DB_FTYPE_BTREE          -1      /* Btree. */
 260 #define DB_FTYPE_HASH           -2      /* Hash. */
 261
 262 /* Structure used as the DB pgin/pgout pgcookie. */
 263 typedef struct __dbpginfo {
 264         size_t  db_pagesize;            /* Underlying page size. */
 265         int     needswap;               /* If swapping required. */
 266 } DB_PGINFO;
 267
 268 /*******************************************************
 269  * Log.
 270  *******************************************************/
 271 /* Initialize an LSN to 'zero'. */
 272 #define ZERO_LSN(LSN) {                                                 \
 273         (LSN).file = 0;                                                 \
 274         (LSN).offset = 0;                                               \
 275 }
 276
 277 /* Return 1 if LSN is a 'zero' lsn, otherwise return 0. */
 278 #define IS_ZERO_LSN(LSN)        ((LSN).file == 0)
 279
 280 /* Test if we need to log a change. */
 281 #define DB_LOGGING(dbp)                                                 \
 282         (F_ISSET(dbp, DB_AM_LOGGING) && !F_ISSET(dbp, DB_AM_RECOVER))
 283
 284 #ifdef DEBUG
 285 /*
 286  * Debugging macro to log operations.
 287  *      If DEBUG_WOP is defined, log operations that modify the database.
 288  *      If DEBUG_ROP is defined, log operations that read the database.
 289  *
 290  * D dbp
 291  * T txn
 292  * O operation (string)
 293  * K key
 294  * A data
 295  * F flags
 296  */
 297 #define LOG_OP(D, T, O, K, A, F) {                                      \
 298         DB_LSN _lsn;                                                    \
 299         DBT _op;                                                        \
 300         if (DB_LOGGING((D))) {                                          \
 301                 memset(&_op, 0, sizeof(_op));                           \
 302                 _op.data = O;                                           \
 303                 _op.size = strlen(O) + 1;                               \
 304                 (void)__db_debug_log((D)->dbenv->lg_info,               \
 305                     T, &_lsn, 0, &_op, (D)->log_fileid, K, A, F);       \
 306         }                                                               \
 307 }
 308 #ifdef DEBUG_ROP
 309 #define DEBUG_LREAD(D, T, O, K, A, F)   LOG_OP(D, T, O, K, A, F)
 310 #else
 311 #define DEBUG_LREAD(D, T, O, K, A, F)
 312 #endif
 313 #ifdef DEBUG_WOP
 314 #define DEBUG_LWRITE(D, T, O, K, A, F)  LOG_OP(D, T, O, K, A, F)
 315 #else
 316 #define DEBUG_LWRITE(D, T, O, K, A, F)
 317 #endif
 318 #else
 319 #define DEBUG_LREAD(D, T, O, K, A, F)
 320 #define DEBUG_LWRITE(D, T, O, K, A, F)
 321 #endif /* DEBUG */
 322
 323 /*******************************************************
 324  * Transactions and recovery.
 325  *******************************************************/
 326 /*
 327  * Out of band value for a lock.  The locks are returned to callers as offsets
 328  * into the lock regions.  Since the RLAYOUT structure begins all regions, an
 329  * offset of 0 is guaranteed not to be a valid lock.
 330  */
 331 #define LOCK_INVALID    0
 332
 333 /* The structure allocated for every transaction. */
 334 struct __db_txn {
 335         DB_TXNMGR       *mgrp;          /* Pointer to transaction manager. */
 336         DB_TXN          *parent;        /* Pointer to transaction's parent. */
 337         DB_LSN          last_lsn;       /* Lsn of last log write. */
 338         u_int32_t       txnid;          /* Unique transaction id. */
 339         size_t          off;            /* Detail structure within region. */
 340         TAILQ_ENTRY(__db_txn) links;
 341 };
 342 #endif /* !_DB_INTERNAL_H_ */