dlls/krnl386.exe16/thunk.c

   1 /*
   2  * KERNEL32 thunks and other undocumented stuff
   3  *
   4  * Copyright 1996, 1997 Alexandre Julliard
   5  * Copyright 1997, 1998 Marcus Meissner
   6  * Copyright 1998       Ulrich Weigand
   7  *
   8  * This library is free software; you can redistribute it and/or
   9  * modify it under the terms of the GNU Lesser General Public
  10  * License as published by the Free Software Foundation; either
  11  * version 2.1 of the License, or (at your option) any later version.
  12  *
  13  * This library is distributed in the hope that it will be useful,
  14  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  15  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  16  * Lesser General Public License for more details.
  17  *
  18  * You should have received a copy of the GNU Lesser General Public
  19  * License along with this library; if not, write to the Free Software
  20  * Foundation, Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301, USA
  21  */
  22
  23 #include <string.h>
  24 #include <sys/types.h>
  25 #include <stdarg.h>
  26 #include <stdio.h>
  27
  28 #include "windef.h"
  29 #include "winbase.h"
  30 #include "winerror.h"
  31 #include "winternl.h"
  32 #include "wownt16.h"
  33 #include "wownt32.h"
  34 #include "wine/winbase16.h"
  35
  36 #include "wine/debug.h"
  37 #include "kernel16_private.h"
  38
  39 WINE_DEFAULT_DEBUG_CHANNEL(thunk);
  40
  41 struct ThunkDataCommon
  42 {
  43     char                   magic[4];         /* 00 */
  44     DWORD                  checksum;         /* 04 */
  45 };
  46
  47 struct ThunkDataLS16
  48 {
  49     struct ThunkDataCommon common;           /* 00 */
  50     SEGPTR                 targetTable;      /* 08 */
  51     DWORD                  firstTime;        /* 0C */
  52 };
  53
  54 struct ThunkDataLS32
  55 {
  56     struct ThunkDataCommon common;           /* 00 */
  57     DWORD *                targetTable;      /* 08 */
  58     char                   lateBinding[4];   /* 0C */
  59     DWORD                  flags;            /* 10 */
  60     DWORD                  reserved1;        /* 14 */
  61     DWORD                  reserved2;        /* 18 */
  62     DWORD                  offsetQTThunk;    /* 1C */
  63     DWORD                  offsetFTProlog;   /* 20 */
  64 };
  65
  66 struct ThunkDataSL16
  67 {
  68     struct ThunkDataCommon common;            /* 00 */
  69     DWORD                  flags1;            /* 08 */
  70     DWORD                  reserved1;         /* 0C */
  71     struct ThunkDataSL *   fpData;            /* 10 */
  72     SEGPTR                 spData;            /* 14 */
  73     DWORD                  reserved2;         /* 18 */
  74     char                   lateBinding[4];    /* 1C */
  75     DWORD                  flags2;            /* 20 */
  76     DWORD                  reserved3;         /* 20 */
  77     SEGPTR                 apiDatabase;       /* 28 */
  78 };
  79
  80 struct ThunkDataSL32
  81 {
  82     struct ThunkDataCommon common;            /* 00 */
  83     DWORD                  reserved1;         /* 08 */
  84     struct ThunkDataSL *   data;              /* 0C */
  85     char                   lateBinding[4];    /* 10 */
  86     DWORD                  flags;             /* 14 */
  87     DWORD                  reserved2;         /* 18 */
  88     DWORD                  reserved3;         /* 1C */
  89     DWORD                  offsetTargetTable; /* 20 */
  90 };
  91
  92 struct ThunkDataSL
  93 {
  94 #if 0
  95     This structure differs from the Win95 original,
  96     but this should not matter since it is strictly internal to
  97     the thunk handling routines in KRNL386 / KERNEL32.
  98
  99     For reference, here is the Win95 layout:
 100
 101     struct ThunkDataCommon common;            /* 00 */
 102     DWORD                  flags1;            /* 08 */
 103     SEGPTR                 apiDatabase;       /* 0C */
 104     WORD                   exePtr;            /* 10 */
 105     WORD                   segMBA;            /* 12 */
 106     DWORD                  lenMBATotal;       /* 14 */
 107     DWORD                  lenMBAUsed;        /* 18 */
 108     DWORD                  flags2;            /* 1C */
 109     char                   pszDll16[256];     /* 20 */
 110     char                   pszDll32[256];     /*120 */
 111
 112     We do it differently since all our thunk handling is done
 113     by 32-bit code. Therefore we do not need to provide
 114     easy access to this data, especially the process target
 115     table database, for 16-bit code.
 116 #endif
 117
 118     struct ThunkDataCommon common;
 119     DWORD                  flags1;
 120     struct SLApiDB *       apiDB;
 121     struct SLTargetDB *    targetDB;
 122     DWORD                  flags2;
 123     char                   pszDll16[256];
 124     char                   pszDll32[256];
 125 };
 126
 127 struct SLTargetDB
 128 {
 129      struct SLTargetDB *   next;
 130      DWORD                 process;
 131      DWORD *               targetTable;
 132 };
 133
 134 struct SLApiDB
 135 {
 136     DWORD                  nrArgBytes;
 137     DWORD                  errorReturnValue;
 138 };
 139
 140 SEGPTR CALL32_CBClient_RetAddr = 0;
 141 SEGPTR CALL32_CBClientEx_RetAddr = 0;
 142
 143 extern int call_entry_point( void *func, int nb_args, const DWORD *args );
 144 extern void __wine_call_from_16_thunk(void);
 145 extern void WINAPI FT_Prolog(void);
 146 extern void WINAPI FT_PrologPrime(void);
 147 extern void WINAPI QT_Thunk(void);
 148 extern void WINAPI QT_ThunkPrime(void);
 149
 150 /***********************************************************************
 151  *                                                                     *
 152  *                 Win95 internal thunks                               *
 153  *                                                                     *
 154  ***********************************************************************/
 155
 156 /***********************************************************************
 157  *           LogApiThk    (KERNEL.423)
 158  */
 159 void WINAPI LogApiThk( LPSTR func )
 160 {
 161     TRACE( "%s\n", debugstr_a(func) );
 162 }
 163
 164 /***********************************************************************
 165  *           LogApiThkLSF    (KERNEL32.42)
 166  *
 167  * NOTE: needs to preserve all registers!
 168  */
 169 __ASM_STDCALL_FUNC( LogApiThkLSF, 4, "ret $4" )
 170
 171 /***********************************************************************
 172  *           LogApiThkSL    (KERNEL32.44)
 173  *
 174  * NOTE: needs to preserve all registers!
 175  */
 176 __ASM_STDCALL_FUNC( LogApiThkSL, 4, "ret $4" )
 177
 178 /***********************************************************************
 179  *           LogCBThkSL    (KERNEL32.47)
 180  *
 181  * NOTE: needs to preserve all registers!
 182  */
 183 __ASM_STDCALL_FUNC( LogCBThkSL, 4, "ret $4" )
 184
 185 /***********************************************************************
 186  * Generates a FT_Prolog call.
 187  *
 188  *  0FB6D1                  movzbl edx,cl
 189  *  8B1495xxxxxxxx          mov edx,[4*edx + targetTable]
 190  *  68xxxxxxxx              push FT_Prolog
 191  *  C3                      lret
 192  */
 193 static void _write_ftprolog(LPBYTE relayCode ,DWORD *targetTable) {
 194         LPBYTE  x;
 195
 196         x       = relayCode;
 197         *x++    = 0x0f;*x++=0xb6;*x++=0xd1; /* movzbl edx,cl */
 198         *x++    = 0x8B;*x++=0x14;*x++=0x95;*(DWORD**)x= targetTable;
 199         x+=4;   /* mov edx, [4*edx + targetTable] */
 200         *x++    = 0x68; *(void **)x = FT_Prolog;
 201         x+=4;   /* push FT_Prolog */
 202         *x++    = 0xC3;         /* lret */
 203         /* fill rest with 0xCC / int 3 */
 204 }
 205
 206 /***********************************************************************
 207  *      _write_qtthunk                                  (internal)
 208  * Generates a QT_Thunk style call.
 209  *
 210  *  33C9                    xor ecx, ecx
 211  *  8A4DFC                  mov cl , [ebp-04]
 212  *  8B148Dxxxxxxxx          mov edx, [4*ecx + targetTable]
 213  *  B8yyyyyyyy              mov eax, QT_Thunk
 214  *  FFE0                    jmp eax
 215  */
 216 static void _write_qtthunk(
 217         LPBYTE relayCode,       /* [in] start of QT_Thunk stub */
 218         DWORD *targetTable      /* [in] start of thunk (for index lookup) */
 219 ) {
 220         LPBYTE  x;
 221
 222         x       = relayCode;
 223         *x++    = 0x33;*x++=0xC9; /* xor ecx,ecx */
 224         *x++    = 0x8A;*x++=0x4D;*x++=0xFC; /* movb cl,[ebp-04] */
 225         *x++    = 0x8B;*x++=0x14;*x++=0x8D;*(DWORD**)x= targetTable;
 226         x+=4;   /* mov edx, [4*ecx + targetTable */
 227         *x++    = 0xB8; *(void **)x = QT_Thunk;
 228         x+=4;   /* mov eax , QT_Thunk */
 229         *x++    = 0xFF; *x++ = 0xE0;    /* jmp eax */
 230         /* should fill the rest of the 32 bytes with 0xCC */
 231 }
 232
 233 /***********************************************************************
 234  *           _loadthunk
 235  */
 236 static LPVOID _loadthunk(LPCSTR module, LPCSTR func, LPCSTR module32,
 237                          struct ThunkDataCommon *TD32, DWORD checksum)
 238 {
 239     struct ThunkDataCommon *TD16;
 240     HMODULE16 hmod;
 241     int ordinal;
 242     static BOOL done;
 243
 244     if (!done)
 245     {
 246         LoadLibrary16( "gdi.exe" );
 247         LoadLibrary16( "user.exe" );
 248         done = TRUE;
 249     }
 250
 251     if ((hmod = LoadLibrary16(module)) <= 32)
 252     {
 253         ERR("(%s, %s, %s): Unable to load '%s', error %d\n",
 254                    module, func, module32, module, hmod);
 255         return 0;
 256     }
 257
 258     if (   !(ordinal = NE_GetOrdinal(hmod, func))
 259         || !(TD16 = MapSL((SEGPTR)NE_GetEntryPointEx(hmod, ordinal, FALSE))))
 260     {
 261         ERR("Unable to find thunk data '%s' in %s, required by %s (conflicting/incorrect DLL versions !?).\n",
 262                    func, module, module32);
 263         return 0;
 264     }
 265
 266     if (TD32 && memcmp(TD16->magic, TD32->magic, 4))
 267     {
 268         ERR("(%s, %s, %s): Bad magic %c%c%c%c (should be %c%c%c%c)\n",
 269                    module, func, module32,
 270                    TD16->magic[0], TD16->magic[1], TD16->magic[2], TD16->magic[3],
 271                    TD32->magic[0], TD32->magic[1], TD32->magic[2], TD32->magic[3]);
 272         return 0;
 273     }
 274
 275     if (TD32 && TD16->checksum != TD32->checksum)
 276     {
 277         ERR("(%s, %s, %s): Wrong checksum %08x (should be %08x)\n",
 278                    module, func, module32, TD16->checksum, TD32->checksum);
 279         return 0;
 280     }
 281
 282     if (!TD32 && checksum && checksum != *(LPDWORD)TD16)
 283     {
 284         ERR("(%s, %s, %s): Wrong checksum %08x (should be %08x)\n",
 285                    module, func, module32, *(LPDWORD)TD16, checksum);
 286         return 0;
 287     }
 288
 289     return TD16;
 290 }
 291
 292 /***********************************************************************
 293  *           GetThunkStuff    (KERNEL32.53)
 294  */
 295 LPVOID WINAPI GetThunkStuff(LPCSTR module, LPCSTR func)
 296 {
 297     return _loadthunk(module, func, "<kernel>", NULL, 0L);
 298 }
 299
 300 /***********************************************************************
 301  *           GetThunkBuff    (KERNEL32.52)
 302  * Returns a pointer to ThkBuf in the 16bit library SYSTHUNK.DLL.
 303  */
 304 LPVOID WINAPI GetThunkBuff(void)
 305 {
 306     return GetThunkStuff("SYSTHUNK.DLL", "ThkBuf");
 307 }
 308
 309 /***********************************************************************
 310  *              ThunkConnect32          (KERNEL32.@)
 311  * Connects a 32bit and a 16bit thunkbuffer.
 312  */
 313 UINT WINAPI ThunkConnect32(
 314         struct ThunkDataCommon *TD,  /* [in/out] thunkbuffer */
 315         LPSTR thunkfun16,            /* [in] win16 thunkfunction */
 316         LPSTR module16,              /* [in] name of win16 dll */
 317         LPSTR module32,              /* [in] name of win32 dll */
 318         HMODULE hmod32,            /* [in] hmodule of win32 dll */
 319         DWORD dwReason               /* [in] initialisation argument */
 320 ) {
 321     BOOL directionSL;
 322
 323     if (!strncmp(TD->magic, "SL01", 4))
 324     {
 325         directionSL = TRUE;
 326
 327         TRACE("SL01 thunk %s (%p) <- %s (%s), Reason: %d\n",
 328               module32, TD, module16, thunkfun16, dwReason);
 329     }
 330     else if (!strncmp(TD->magic, "LS01", 4))
 331     {
 332         directionSL = FALSE;
 333
 334         TRACE("LS01 thunk %s (%p) -> %s (%s), Reason: %d\n",
 335               module32, TD, module16, thunkfun16, dwReason);
 336     }
 337     else
 338     {
 339         ERR("Invalid magic %c%c%c%c\n",
 340                    TD->magic[0], TD->magic[1], TD->magic[2], TD->magic[3]);
 341         return 0;
 342     }
 343
 344     switch (dwReason)
 345     {
 346         case DLL_PROCESS_ATTACH:
 347         {
 348             struct ThunkDataCommon *TD16;
 349             if (!(TD16 = _loadthunk(module16, thunkfun16, module32, TD, 0L)))
 350                 return 0;
 351
 352             if (directionSL)
 353             {
 354                 struct ThunkDataSL32 *SL32 = (struct ThunkDataSL32 *)TD;
 355                 struct ThunkDataSL16 *SL16 = (struct ThunkDataSL16 *)TD16;
 356                 struct SLTargetDB *tdb;
 357
 358                 if (SL16->fpData == NULL)
 359                 {
 360                     ERR("ThunkConnect16 was not called!\n");
 361                     return 0;
 362                 }
 363
 364                 SL32->data = SL16->fpData;
 365
 366                 tdb = HeapAlloc(GetProcessHeap(), 0, sizeof(*tdb));
 367                 tdb->process = GetCurrentProcessId();
 368                 tdb->targetTable = (DWORD *)(thunkfun16 + SL32->offsetTargetTable);
 369
 370                 tdb->next = SL32->data->targetDB;   /* FIXME: not thread-safe! */
 371                 SL32->data->targetDB = tdb;
 372
 373                 TRACE("Process %08x allocated TargetDB entry for ThunkDataSL %p\n",
 374                       GetCurrentProcessId(), SL32->data);
 375             }
 376             else
 377             {
 378                 struct ThunkDataLS32 *LS32 = (struct ThunkDataLS32 *)TD;
 379                 struct ThunkDataLS16 *LS16 = (struct ThunkDataLS16 *)TD16;
 380
 381                 LS32->targetTable = MapSL(LS16->targetTable);
 382
 383                 /* write QT_Thunk and FT_Prolog stubs */
 384                 _write_qtthunk ((LPBYTE)TD + LS32->offsetQTThunk,  LS32->targetTable);
 385                 _write_ftprolog((LPBYTE)TD + LS32->offsetFTProlog, LS32->targetTable);
 386             }
 387             break;
 388         }
 389
 390         case DLL_PROCESS_DETACH:
 391             /* FIXME: cleanup */
 392             break;
 393     }
 394
 395     return 1;
 396 }
 397
 398 /**********************************************************************
 399  *              QT_Thunk                        (KERNEL32.@)
 400  *
 401  * The target address is in EDX.
 402  * The 16bit arguments start at ESP.
 403  * The number of 16bit argument bytes is EBP-ESP-0x40 (64 Byte thunksetup).
 404  * So the stack layout is 16bit argument bytes and then the 64 byte
 405  * scratch buffer.
 406  * The scratch buffer is used as work space by Windows' QT_Thunk
 407  * function.
 408  * As the programs unfortunately don't always provide a fixed size
 409  * scratch buffer (danger, stack corruption ahead !!), we simply resort
 410  * to copying over the whole EBP-ESP range to the 16bit stack
 411  * (as there's no way to safely figure out the param count
 412  * due to this misbehaviour of some programs).
 413  * [ok]
 414  *
 415  * See DDJ article 9614c for a very good description of QT_Thunk (also
 416  * available online !).
 417  *
 418  * FIXME: DDJ talks of certain register usage rules; I'm not sure
 419  * whether we cover this 100%.
 420  */
 421 void WINAPI __regs_QT_Thunk( CONTEXT *context )
 422 {
 423     CONTEXT context16;
 424     DWORD argsize;
 425
 426     context16 = *context;
 427
 428     context16.SegCs = HIWORD(context->Edx);
 429     context16.Eip   = LOWORD(context->Edx);
 430     /* point EBP to the STACK16FRAME on the stack
 431      * for the call_to_16 to set up the register content on calling */
 432     context16.Ebp   = CURRENT_SP + FIELD_OFFSET(STACK16FRAME,bp);
 433
 434     /*
 435      * used to be (problematic):
 436      * argsize = context->Ebp - context->Esp - 0x40;
 437      * due to some programs abusing the API, we better assume the full
 438      * EBP - ESP range for copying instead: */
 439     argsize = context->Ebp - context->Esp;
 440
 441     /* ok, too much is insane; let's limit param count a bit again */
 442     if (argsize > 64)
 443         argsize = 64; /* 32 WORDs */
 444
 445     WOWCallback16Ex( 0, WCB16_REGS, argsize, (void *)context->Esp, (DWORD *)&context16 );
 446     context->Eax = context16.Eax;
 447     context->Edx = context16.Edx;
 448     context->Ecx = context16.Ecx;
 449
 450     /* make sure to update the Win32 ESP, too, in order to throw away
 451      * the number of parameters that the Win16 function
 452      * accepted (that it popped from the corresponding Win16 stack) */
 453     context->Esp +=   LOWORD(context16.Esp) - (CURRENT_SP - argsize);
 454 }
 455 DEFINE_REGS_ENTRYPOINT( QT_Thunk )
 456
 457
 458 /**********************************************************************
 459  *              FT_Prolog                       (KERNEL32.@)
 460  *
 461  * The set of FT_... thunk routines is used instead of QT_Thunk,
 462  * if structures have to be converted from 32-bit to 16-bit
 463  * (change of member alignment, conversion of members).
 464  *
 465  * The thunk function (as created by the thunk compiler) calls
 466  * FT_Prolog at the beginning, to set up a stack frame and
 467  * allocate a 64 byte buffer on the stack.
 468  * The input parameters (target address and some flags) are
 469  * saved for later use by FT_Thunk.
 470  *
 471  * Input:  EDX  16-bit target address (SEGPTR)
 472  *         CX   bits  0..7   target number (in target table)
 473  *              bits  8..9   some flags (unclear???)
 474  *              bits 10..15  number of DWORD arguments
 475  *
 476  * Output: A new stackframe is created, and a 64 byte buffer
 477  *         allocated on the stack. The layout of the stack
 478  *         on return is as follows:
 479  *
 480  *  (ebp+4)  return address to caller of thunk function
 481  *  (ebp)    old EBP
 482  *  (ebp-4)  saved EBX register of caller
 483  *  (ebp-8)  saved ESI register of caller
 484  *  (ebp-12) saved EDI register of caller
 485  *  (ebp-16) saved ECX register, containing flags
 486  *  (ebp-20) bitmap containing parameters that are to be converted
 487  *           by FT_Thunk; it is initialized to 0 by FT_Prolog and
 488  *           filled in by the thunk code before calling FT_Thunk
 489  *  (ebp-24)
 490  *    ...    (unclear)
 491  *  (ebp-44)
 492  *  (ebp-48) saved EAX register of caller (unclear, never restored???)
 493  *  (ebp-52) saved EDX register, containing 16-bit thunk target
 494  *  (ebp-56)
 495  *    ...    (unclear)
 496  *  (ebp-64)
 497  *
 498  *  ESP is EBP-64 after return.
 499  *
 500  */
 501 void WINAPI __regs_FT_Prolog( CONTEXT *context )
 502 {
 503     /* Build stack frame */
 504     stack32_push(context, context->Ebp);
 505     context->Ebp = context->Esp;
 506
 507     /* Allocate 64-byte Thunk Buffer */
 508     context->Esp -= 64;
 509     memset((char *)context->Esp, '\0', 64);
 510
 511     /* Store Flags (ECX) and Target Address (EDX) */
 512     /* Save other registers to be restored later */
 513     *(DWORD *)(context->Ebp -  4) = context->Ebx;
 514     *(DWORD *)(context->Ebp -  8) = context->Esi;
 515     *(DWORD *)(context->Ebp - 12) = context->Edi;
 516     *(DWORD *)(context->Ebp - 16) = context->Ecx;
 517
 518     *(DWORD *)(context->Ebp - 48) = context->Eax;
 519     *(DWORD *)(context->Ebp - 52) = context->Edx;
 520 }
 521 DEFINE_REGS_ENTRYPOINT( FT_Prolog )
 522
 523 /**********************************************************************
 524  *              FT_Thunk                        (KERNEL32.@)
 525  *
 526  * This routine performs the actual call to 16-bit code,
 527  * similar to QT_Thunk. The differences are:
 528  *  - The call target is taken from the buffer created by FT_Prolog
 529  *  - Those arguments requested by the thunk code (by setting the
 530  *    corresponding bit in the bitmap at EBP-20) are converted
 531  *    from 32-bit pointers to segmented pointers (those pointers
 532  *    are guaranteed to point to structures copied to the stack
 533  *    by the thunk code, so we always use the 16-bit stack selector
 534  *    for those addresses).
 535  *
 536  *    The bit #i of EBP-20 corresponds here to the DWORD starting at
 537  *    ESP+4 + 2*i.
 538  *
 539  * FIXME: It is unclear what happens if there are more than 32 WORDs
 540  *        of arguments, so that the single DWORD bitmap is no longer
 541  *        sufficient ...
 542  */
 543 void WINAPI __regs_FT_Thunk( CONTEXT *context )
 544 {
 545     DWORD mapESPrelative = *(DWORD *)(context->Ebp - 20);
 546     DWORD callTarget     = *(DWORD *)(context->Ebp - 52);
 547
 548     CONTEXT context16;
 549     DWORD i, argsize;
 550     DWORD newstack[32];
 551     LPBYTE oldstack;
 552
 553     context16 = *context;
 554
 555     context16.SegCs = HIWORD(callTarget);
 556     context16.Eip   = LOWORD(callTarget);
 557     context16.Ebp   = CURRENT_SP + FIELD_OFFSET(STACK16FRAME,bp);
 558
 559     argsize  = context->Ebp-context->Esp-0x40;
 560     if (argsize > sizeof(newstack)) argsize = sizeof(newstack);
 561     oldstack = (LPBYTE)context->Esp;
 562
 563     memcpy( newstack, oldstack, argsize );
 564
 565     for (i = 0; i < 32; i++)    /* NOTE: What about > 32 arguments? */
 566         if (mapESPrelative & (1 << i))
 567         {
 568             SEGPTR *arg = (SEGPTR *)newstack[i];
 569             *arg = MAKESEGPTR( CURRENT_SS, CURRENT_SP - argsize + (*(LPBYTE *)arg - oldstack));
 570         }
 571
 572     WOWCallback16Ex( 0, WCB16_REGS, argsize, newstack, (DWORD *)&context16 );
 573     context->Eax = context16.Eax;
 574     context->Edx = context16.Edx;
 575     context->Ecx = context16.Ecx;
 576
 577     context->Esp +=   LOWORD(context16.Esp) - (CURRENT_SP - argsize);
 578
 579     /* Copy modified buffers back to 32-bit stack */
 580     memcpy( oldstack, newstack, argsize );
 581 }
 582 DEFINE_REGS_ENTRYPOINT( FT_Thunk )
 583
 584 /***********************************************************************
 585  *              FT_Exit0 (KERNEL32.@)
 586  *              FT_Exit4 (KERNEL32.@)
 587  *              FT_Exit8 (KERNEL32.@)
 588  *              FT_Exit12 (KERNEL32.@)
 589  *              FT_Exit16 (KERNEL32.@)
 590  *              FT_Exit20 (KERNEL32.@)
 591  *              FT_Exit24 (KERNEL32.@)
 592  *              FT_Exit28 (KERNEL32.@)
 593  *              FT_Exit32 (KERNEL32.@)
 594  *              FT_Exit36 (KERNEL32.@)
 595  *              FT_Exit40 (KERNEL32.@)
 596  *              FT_Exit44 (KERNEL32.@)
 597  *              FT_Exit48 (KERNEL32.@)
 598  *              FT_Exit52 (KERNEL32.@)
 599  *              FT_Exit56 (KERNEL32.@)
 600  *
 601  * One of the FT_ExitNN functions is called at the end of the thunk code.
 602  * It removes the stack frame created by FT_Prolog, moves the function
 603  * return from EBX to EAX (yes, FT_Thunk did use EAX for the return
 604  * value, but the thunk code has moved it from EAX to EBX in the
 605  * meantime ... :-), restores the caller's EBX, ESI, and EDI registers,
 606  * and perform a return to the CALLER of the thunk code (while removing
 607  * the given number of arguments from the caller's stack).
 608  */
 609 #define FT_EXIT_RESTORE_REGS \
 610     "movl %ebx,%eax\n\t" \
 611     "movl -4(%ebp),%ebx\n\t" \
 612     "movl -8(%ebp),%esi\n\t" \
 613     "movl -12(%ebp),%edi\n\t" \
 614     "leave\n\t"
 615
 616 #define DEFINE_FT_Exit(n) \
 617     __ASM_STDCALL_FUNC( FT_Exit ## n, 0, FT_EXIT_RESTORE_REGS "ret $" #n )
 618
 619 DEFINE_FT_Exit(0)
 620 DEFINE_FT_Exit(4)
 621 DEFINE_FT_Exit(8)
 622 DEFINE_FT_Exit(12)
 623 DEFINE_FT_Exit(16)
 624 DEFINE_FT_Exit(20)
 625 DEFINE_FT_Exit(24)
 626 DEFINE_FT_Exit(28)
 627 DEFINE_FT_Exit(32)
 628 DEFINE_FT_Exit(36)
 629 DEFINE_FT_Exit(40)
 630 DEFINE_FT_Exit(44)
 631 DEFINE_FT_Exit(48)
 632 DEFINE_FT_Exit(52)
 633 DEFINE_FT_Exit(56)
 634
 635
 636 /***********************************************************************
 637  *              ThunkInitLS     (KERNEL32.43)
 638  * A thunkbuffer link routine
 639  * The thunkbuf looks like:
 640  *
 641  *      00: DWORD       length          ? don't know exactly
 642  *      04: SEGPTR      ptr             ? where does it point to?
 643  * The pointer ptr is written into the first DWORD of 'thunk'.
 644  * (probably correctly implemented)
 645  * [ok probably]
 646  * RETURNS
 647  *      segmented pointer to thunk?
 648  */
 649 DWORD WINAPI ThunkInitLS(
 650         LPDWORD thunk,  /* [in] win32 thunk */
 651         LPCSTR thkbuf,  /* [in] thkbuffer name in win16 dll */
 652         DWORD len,      /* [in] thkbuffer length */
 653         LPCSTR dll16,   /* [in] name of win16 dll */
 654         LPCSTR dll32    /* [in] name of win32 dll (FIXME: not used?) */
 655 ) {
 656         LPDWORD         addr;
 657
 658         if (!(addr = _loadthunk( dll16, thkbuf, dll32, NULL, len )))
 659                 return 0;
 660
 661         if (!addr[1])
 662                 return 0;
 663         *thunk = addr[1];
 664
 665         return addr[1];
 666 }
 667
 668 /***********************************************************************
 669  *              Common32ThkLS   (KERNEL32.45)
 670  *
 671  * This is another 32->16 thunk, independent of the QT_Thunk/FT_Thunk
 672  * style thunks. The basic difference is that the parameter conversion
 673  * is done completely on the *16-bit* side here. Thus we do not call
 674  * the 16-bit target directly, but call a common entry point instead.
 675  * This entry function then calls the target according to the target
 676  * number passed in the DI register.
 677  *
 678  * Input:  EAX    SEGPTR to the common 16-bit entry point
 679  *         CX     offset in thunk table (target number * 4)
 680  *         DX     error return value if execution fails (unclear???)
 681  *         EDX.HI number of DWORD parameters
 682  *
 683  * (Note that we need to move the thunk table offset from CX to DI !)
 684  *
 685  * The called 16-bit stub expects its stack to look like this:
 686  *     ...
 687  *   (esp+40)  32-bit arguments
 688  *     ...
 689  *   (esp+8)   32 byte of stack space available as buffer
 690  *   (esp)     8 byte return address for use with 0x66 lret
 691  *
 692  * The called 16-bit stub uses a 0x66 lret to return to 32-bit code,
 693  * and uses the EAX register to return a DWORD return value.
 694  * Thus we need to use a special assembly glue routine
 695  * (CallRegisterLongProc instead of CallRegisterShortProc).
 696  *
 697  * Finally, we return to the caller, popping the arguments off
 698  * the stack.  The number of arguments to be popped is returned
 699  * in the BL register by the called 16-bit routine.
 700  *
 701  */
 702 void WINAPI __regs_Common32ThkLS( CONTEXT *context )
 703 {
 704     CONTEXT context16;
 705     DWORD argsize;
 706
 707     context16 = *context;
 708
 709     context16.Edi   = LOWORD(context->Ecx);
 710     context16.SegCs = HIWORD(context->Eax);
 711     context16.Eip   = LOWORD(context->Eax);
 712     context16.Ebp   = CURRENT_SP + FIELD_OFFSET(STACK16FRAME,bp);
 713
 714     argsize = HIWORD(context->Edx) * 4;
 715
 716     /* FIXME: hack for stupid USER32 CallbackGlueLS routine */
 717     if (context->Edx == context->Eip)
 718         argsize = 6 * 4;
 719
 720     /* Note: the first 32 bytes we copy are just garbage from the 32-bit stack, in order to reserve
 721      *       the space. It is safe to do that since the register function prefix has reserved
 722      *       a lot more space than that below context->Esp.
 723      */
 724     WOWCallback16Ex( 0, WCB16_REGS, argsize + 32, (LPBYTE)context->Esp - 32, (DWORD *)&context16 );
 725     context->Eax = context16.Eax;
 726
 727     /* Clean up caller's stack frame */
 728     context->Esp += LOBYTE(context16.Ebx);
 729 }
 730 DEFINE_REGS_ENTRYPOINT( Common32ThkLS )
 731
 732 /***********************************************************************
 733  *              OT_32ThkLSF     (KERNEL32.40)
 734  *
 735  * YET Another 32->16 thunk. The difference to Common32ThkLS is that
 736  * argument processing is done on both the 32-bit and the 16-bit side:
 737  * The 32-bit side prepares arguments, copying them onto the stack.
 738  *
 739  * When this routine is called, the first word on the stack is the
 740  * number of argument bytes prepared by the 32-bit code, and EDX
 741  * contains the 16-bit target address.
 742  *
 743  * The called 16-bit routine is another relaycode, doing further
 744  * argument processing and then calling the real 16-bit target
 745  * whose address is stored at [bp-04].
 746  *
 747  * The call proceeds using a normal CallRegisterShortProc.
 748  * After return from the 16-bit relaycode, the arguments need
 749  * to be copied *back* to the 32-bit stack, since the 32-bit
 750  * relaycode processes output parameters.
 751  *
 752  * Note that we copy twice the number of arguments, since some of the
 753  * 16-bit relaycodes in SYSTHUNK.DLL directly access the original
 754  * arguments of the caller!
 755  *
 756  * (Note that this function seems only to be used for
 757  *  OLECLI32 -> OLECLI and OLESVR32 -> OLESVR thunking.)
 758  */
 759 void WINAPI __regs_OT_32ThkLSF( CONTEXT *context )
 760 {
 761     CONTEXT context16;
 762     DWORD argsize;
 763
 764     context16 = *context;
 765
 766     context16.SegCs = HIWORD(context->Edx);
 767     context16.Eip   = LOWORD(context->Edx);
 768     context16.Ebp   = CURRENT_SP + FIELD_OFFSET(STACK16FRAME,bp);
 769
 770     argsize = 2 * *(WORD *)context->Esp + 2;
 771
 772     WOWCallback16Ex( 0, WCB16_REGS, argsize, (void *)context->Esp, (DWORD *)&context16 );
 773     context->Eax = context16.Eax;
 774     context->Edx = context16.Edx;
 775
 776     /* Copy modified buffers back to 32-bit stack */
 777     memcpy( (LPBYTE)context->Esp,
 778             (LPBYTE)CURRENT_STACK16 - argsize, argsize );
 779
 780     context->Esp +=   LOWORD(context16.Esp) - (CURRENT_SP - argsize);
 781 }
 782 DEFINE_REGS_ENTRYPOINT( OT_32ThkLSF )
 783
 784 /***********************************************************************
 785  *              ThunkInitLSF            (KERNEL32.41)
 786  * A thunk setup routine.
 787  * Expects a pointer to a preinitialized thunkbuffer in the first argument
 788  * looking like:
 789  *|     00..03:         unknown (pointer, check _41, _43, _46)
 790  *|     04: EB1E                jmp +0x20
 791  *|
 792  *|     06..23:         unknown (space for replacement code, check .90)
 793  *|
 794  *|     24:>E800000000          call offset 29
 795  *|     29:>58                  pop eax            ( target of call )
 796  *|     2A: 2D25000000          sub eax,0x00000025 ( now points to offset 4 )
 797  *|     2F: BAxxxxxxxx          mov edx,xxxxxxxx
 798  *|     34: 68yyyyyyyy          push KERNEL32.90
 799  *|     39: C3                  ret
 800  *|
 801  *|     3A: EB1E                jmp +0x20
 802  *|     3E ... 59:      unknown (space for replacement code?)
 803  *|     5A: E8xxxxxxxx          call <32bitoffset xxxxxxxx>
 804  *|     5F: 5A                  pop edx
 805  *|     60: 81EA25xxxxxx        sub edx, 0x25xxxxxx
 806  *|     66: 52                  push edx
 807  *|     67: 68xxxxxxxx          push xxxxxxxx
 808  *|     6C: 68yyyyyyyy          push KERNEL32.89
 809  *|     71: C3                  ret
 810  *|     72: end?
 811  * This function checks if the code is there, and replaces the yyyyyyyy entries
 812  * by the functionpointers.
 813  * The thunkbuf looks like:
 814  *
 815  *|     00: DWORD       length          ? don't know exactly
 816  *|     04: SEGPTR      ptr             ? where does it point to?
 817  * The segpointer ptr is written into the first DWORD of 'thunk'.
 818  * [ok probably]
 819  * RETURNS
 820  *      unclear, pointer to win16 thkbuffer?
 821  */
 822 LPVOID WINAPI ThunkInitLSF(
 823         LPBYTE thunk,   /* [in] win32 thunk */
 824         LPCSTR thkbuf,  /* [in] thkbuffer name in win16 dll */
 825         DWORD len,      /* [in] length of thkbuffer */
 826         LPCSTR dll16,   /* [in] name of win16 dll */
 827         LPCSTR dll32    /* [in] name of win32 dll */
 828 ) {
 829         LPDWORD         addr,addr2;
 830
 831         /* FIXME: add checks for valid code ... */
 832         /* write pointers to kernel32.89 and kernel32.90 (+ordinal base of 1) */
 833         *(void **)(thunk+0x35) = QT_ThunkPrime;
 834         *(void **)(thunk+0x6D) = FT_PrologPrime;
 835
 836         if (!(addr = _loadthunk( dll16, thkbuf, dll32, NULL, len )))
 837                 return 0;
 838
 839         addr2 = MapSL(addr[1]);
 840         if (HIWORD(addr2))
 841                 *(DWORD*)thunk = (DWORD)addr2;
 842
 843         return addr2;
 844 }
 845
 846 /***********************************************************************
 847  *              FT_PrologPrime                  (KERNEL32.89)
 848  *
 849  * This function is called from the relay code installed by
 850  * ThunkInitLSF. It replaces the location from where it was
 851  * called by a standard FT_Prolog call stub (which is 'primed'
 852  * by inserting the correct target table pointer).
 853  * Finally, it calls that stub.
 854  *
 855  * Input:  ECX    target number + flags (passed through to FT_Prolog)
 856  *        (ESP)   offset of location where target table pointer
 857  *                is stored, relative to the start of the relay code
 858  *        (ESP+4) pointer to start of relay code
 859  *                (this is where the FT_Prolog call stub gets written to)
 860  *
 861  * Note: The two DWORD arguments get popped off the stack.
 862  *
 863  */
 864 void WINAPI __regs_FT_PrologPrime( CONTEXT *context )
 865 {
 866     DWORD  targetTableOffset;
 867     LPBYTE relayCode;
 868
 869     /* Compensate for the fact that the Wine register relay code thought
 870        we were being called, although we were in fact jumped to */
 871     context->Esp -= 4;
 872
 873     /* Write FT_Prolog call stub */
 874     targetTableOffset = stack32_pop(context);
 875     relayCode = (LPBYTE)stack32_pop(context);
 876     _write_ftprolog( relayCode, *(DWORD **)(relayCode+targetTableOffset) );
 877
 878     /* Jump to the call stub just created */
 879     context->Eip = (DWORD)relayCode;
 880 }
 881 DEFINE_REGS_ENTRYPOINT( FT_PrologPrime )
 882
 883 /***********************************************************************
 884  *              QT_ThunkPrime                   (KERNEL32.90)
 885  *
 886  * This function corresponds to FT_PrologPrime, but installs a
 887  * call stub for QT_Thunk instead.
 888  *
 889  * Input: (EBP-4) target number (passed through to QT_Thunk)
 890  *         EDX    target table pointer location offset
 891  *         EAX    start of relay code
 892  *
 893  */
 894 void WINAPI __regs_QT_ThunkPrime( CONTEXT *context )
 895 {
 896     DWORD  targetTableOffset;
 897     LPBYTE relayCode;
 898
 899     /* Compensate for the fact that the Wine register relay code thought
 900        we were being called, although we were in fact jumped to */
 901     context->Esp -= 4;
 902
 903     /* Write QT_Thunk call stub */
 904     targetTableOffset = context->Edx;
 905     relayCode = (LPBYTE)context->Eax;
 906     _write_qtthunk( relayCode, *(DWORD **)(relayCode+targetTableOffset) );
 907
 908     /* Jump to the call stub just created */
 909     context->Eip = (DWORD)relayCode;
 910 }
 911 DEFINE_REGS_ENTRYPOINT( QT_ThunkPrime )
 912
 913 /***********************************************************************
 914  *              ThunkInitSL (KERNEL32.46)
 915  * Another thunkbuf link routine.
 916  * The start of the thunkbuf looks like this:
 917  *      00: DWORD       length
 918  *      04: SEGPTR      address for thunkbuffer pointer
 919  * [ok probably]
 920  *
 921  * RETURNS
 922  *  Nothing.
 923  */
 924 VOID WINAPI ThunkInitSL(
 925         LPBYTE thunk,           /* [in] start of thunkbuffer */
 926         LPCSTR thkbuf,          /* [in] name/ordinal of thunkbuffer in win16 dll */
 927         DWORD len,              /* [in] length of thunkbuffer */
 928         LPCSTR dll16,           /* [in] name of win16 dll containing the thkbuf */
 929         LPCSTR dll32            /* [in] win32 dll. FIXME: strange, unused */
 930 ) {
 931         LPDWORD         addr;
 932
 933         if (!(addr = _loadthunk( dll16, thkbuf, dll32, NULL, len )))
 934                 return;
 935
 936         *(DWORD*)MapSL(addr[1]) = (DWORD)thunk;
 937 }
 938
 939 /**********************************************************************
 940  *           SSInit             (KERNEL.700)
 941  * RETURNS
 942  *      TRUE for success.
 943  */
 944 BOOL WINAPI SSInit16(void)
 945 {
 946     return TRUE;
 947 }
 948
 949 /**********************************************************************
 950  *           SSOnBigStack       (KERNEL32.87)
 951  * Check if thunking is initialized (ss selector set up etc.)
 952  * We do that differently, so just return TRUE.
 953  * [ok]
 954  * RETURNS
 955  *      TRUE for success.
 956  */
 957 BOOL WINAPI SSOnBigStack(void)
 958 {
 959     TRACE("Yes, thunking is initialized\n");
 960     return TRUE;
 961 }
 962
 963 /**********************************************************************
 964  *           SSConfirmSmallStack     (KERNEL.704)
 965  *
 966  * Abort if not on small stack.
 967  *
 968  * This must be a register routine as it has to preserve *all* registers.
 969  */
 970 void WINAPI SSConfirmSmallStack( CONTEXT *context )
 971 {
 972     /* We are always on the small stack while in 16-bit code ... */
 973 }
 974
 975 /**********************************************************************
 976  *           SSCall (KERNEL32.88)
 977  * One of the real thunking functions. This one seems to be for 32<->32
 978  * thunks. It should probably be capable of crossing processboundaries.
 979  *
 980  * And YES, I've seen nr=48 (somewhere in the Win95 32<->16 OLE coupling)
 981  * [ok]
 982  *
 983  * RETURNS
 984  *  Thunked function result.
 985  */
 986 DWORD WINAPIV SSCall(
 987         DWORD nr,       /* [in] number of argument bytes */
 988         DWORD flags,    /* [in] FIXME: flags ? */
 989         FARPROC fun,    /* [in] function to call */
 990         ...             /* [in/out] arguments */
 991 ) {
 992     DWORD i,ret;
 993     DWORD *args = ((DWORD *)&fun) + 1;
 994
 995     TRACE("(%d,0x%08x,%p,[",nr,flags,fun);
 996     for (i = 0; i < nr/4; i++) TRACE("0x%08x,",args[i]);
 997     TRACE("])\n");
 998     ret = call_entry_point( fun, nr / sizeof(DWORD), args );
 999     TRACE(" returning %d ...\n",ret);
1000     return ret;
1001 }
1002
1003 /**********************************************************************
1004  *           W32S_BackTo32                      (KERNEL32.51)
1005  */
1006 void WINAPI __regs_W32S_BackTo32( CONTEXT *context )
1007 {
1008     LPDWORD stack = (LPDWORD)context->Esp;
1009     FARPROC proc = (FARPROC)context->Eip;
1010
1011     context->Eax = call_entry_point( proc, 10, stack + 1 );
1012     context->Eip = stack32_pop(context);
1013 }
1014 DEFINE_REGS_ENTRYPOINT( W32S_BackTo32 )
1015
1016 /**********************************************************************
1017  *                      AllocSLCallback         (KERNEL32.@)
1018  *
1019  * Allocate a 16->32 callback.
1020  *
1021  * NOTES
1022  * Win95 uses some structchains for callbacks. It allocates them
1023  * in blocks of 100 entries, size 32 bytes each, layout:
1024  * blockstart:
1025  *|     0:      PTR     nextblockstart
1026  *|     4:      entry   *first;
1027  *|     8:      WORD    sel ( start points to blockstart)
1028  *|     A:      WORD    unknown
1029  * 100xentry:
1030  *|     00..17:         Code
1031  *|     18:     PDB     *owning_process;
1032  *|     1C:     PTR     blockstart
1033  *
1034  * We ignore this for now. (Just a note for further developers)
1035  * FIXME: use this method, so we don't waste selectors...
1036  *
1037  * Following code is then generated by AllocSLCallback. The code is 16 bit, so
1038  * the 0x66 prefix switches from word->long registers.
1039  *
1040  *|     665A            pop     edx
1041  *|     6668x arg2 x    pushl   <arg2>
1042  *|     6652            push    edx
1043  *|     EAx arg1 x      jmpf    <arg1>
1044  *
1045  * returns the startaddress of this thunk.
1046  *
1047  * Note, that they look very similar to the ones allocates by THUNK_Alloc.
1048  * RETURNS
1049  *      A segmented pointer to the start of the thunk
1050  */
1051 DWORD WINAPI
1052 AllocSLCallback(
1053         DWORD finalizer,        /* [in] Finalizer function */
1054         DWORD callback          /* [in] Callback function */
1055 ) {
1056         LPBYTE  x,thunk = HeapAlloc( GetProcessHeap(), 0, 32 );
1057         WORD    sel;
1058
1059         x=thunk;
1060         *x++=0x66;*x++=0x5a;                            /* popl edx */
1061         *x++=0x66;*x++=0x68;*(DWORD*)x=finalizer;x+=4;  /* pushl finalizer */
1062         *x++=0x66;*x++=0x52;                            /* pushl edx */
1063         *x++=0xea;*(DWORD*)x=callback;x+=4;             /* jmpf callback */
1064
1065         *(DWORD*)(thunk+18) = GetCurrentProcessId();
1066
1067         sel = SELECTOR_AllocBlock( thunk, 32, LDT_FLAGS_CODE );
1068         return (sel<<16)|0;
1069 }
1070
1071 /**********************************************************************
1072  *              FreeSLCallback          (KERNEL32.@)
1073  * Frees the specified 16->32 callback
1074  *
1075  * RETURNS
1076  *  Nothing.
1077  */
1078 void WINAPI
1079 FreeSLCallback(
1080         DWORD x /* [in] 16 bit callback (segmented pointer?) */
1081 ) {
1082         FIXME("(0x%08x): stub\n",x);
1083 }
1084
1085 /**********************************************************************
1086  *              AllocMappedBuffer       (KERNEL32.38)
1087  *
1088  * This is an undocumented KERNEL32 function that
1089  * SMapLS's a GlobalAlloc'ed buffer.
1090  *
1091  * RETURNS
1092  *       EDI register: pointer to buffer
1093  *
1094  * NOTES
1095  *       The buffer is preceded by 8 bytes:
1096  *        ...
1097  *       edi+0   buffer
1098  *       edi-4   SEGPTR to buffer
1099  *       edi-8   some magic Win95 needs for SUnMapLS
1100  *               (we use it for the memory handle)
1101  *
1102  *       The SEGPTR is used by the caller!
1103  */
1104 void WINAPI __regs_AllocMappedBuffer(
1105               CONTEXT *context /* [in] EDI register: size of buffer to allocate */
1106 ) {
1107     HGLOBAL handle = GlobalAlloc(0, context->Edi + 8);
1108     DWORD *buffer = GlobalLock(handle);
1109     DWORD ptr = 0;
1110
1111     if (buffer)
1112         if (!(ptr = MapLS(buffer + 2)))
1113         {
1114             GlobalUnlock(handle);
1115             GlobalFree(handle);
1116         }
1117
1118     if (!ptr)
1119         context->Eax = context->Edi = 0;
1120     else
1121     {
1122         buffer[0] = (DWORD)handle;
1123         buffer[1] = ptr;
1124
1125         context->Eax = ptr;
1126         context->Edi = (DWORD)(buffer + 2);
1127     }
1128 }
1129 DEFINE_REGS_ENTRYPOINT( AllocMappedBuffer )
1130
1131 /**********************************************************************
1132  *              FreeMappedBuffer        (KERNEL32.39)
1133  *
1134  * Free a buffer allocated by AllocMappedBuffer
1135  *
1136  * RETURNS
1137  *  Nothing.
1138  */
1139 void WINAPI __regs_FreeMappedBuffer(
1140               CONTEXT *context /* [in] EDI register: pointer to buffer */
1141 ) {
1142     if (context->Edi)
1143     {
1144         DWORD *buffer = (DWORD *)context->Edi - 2;
1145
1146         UnMapLS(buffer[1]);
1147
1148         GlobalUnlock((HGLOBAL)buffer[0]);
1149         GlobalFree((HGLOBAL)buffer[0]);
1150     }
1151 }
1152 DEFINE_REGS_ENTRYPOINT( FreeMappedBuffer )
1153
1154 /**********************************************************************
1155  *              GetTEBSelectorFS        (KERNEL.475)
1156  *      Set the 16-bit %fs to the 32-bit %fs (current TEB selector)
1157  */
1158 void WINAPI GetTEBSelectorFS16(void)
1159 {
1160     CURRENT_STACK16->fs = get_fs();
1161 }
1162
1163 /**********************************************************************
1164  *              IsPeFormat              (KERNEL.431)
1165  *
1166  * Determine if a file is a PE format executable.
1167  *
1168  * RETURNS
1169  *  TRUE, if it is.
1170  *  FALSE if the file could not be opened or is not a PE file.
1171  *
1172  * NOTES
1173  *  If fn is given as NULL then the function expects hf16 to be valid.
1174  */
1175 BOOL16 WINAPI IsPeFormat16(
1176         LPSTR   fn,     /* [in] Filename to the executable */
1177         HFILE16 hf16)   /* [in] An open file handle */
1178 {
1179     BOOL ret = FALSE;
1180     IMAGE_DOS_HEADER mzh;
1181     OFSTRUCT ofs;
1182     DWORD xmagic;
1183
1184     if (fn) hf16 = OpenFile16(fn,&ofs,OF_READ);
1185     if (hf16 == HFILE_ERROR16) return FALSE;
1186     _llseek16(hf16,0,SEEK_SET);
1187     if (sizeof(mzh)!=_lread16(hf16,&mzh,sizeof(mzh))) goto done;
1188     if (mzh.e_magic!=IMAGE_DOS_SIGNATURE) goto done;
1189     _llseek16(hf16,mzh.e_lfanew,SEEK_SET);
1190     if (sizeof(DWORD)!=_lread16(hf16,&xmagic,sizeof(DWORD))) goto done;
1191     ret = (xmagic == IMAGE_NT_SIGNATURE);
1192  done:
1193     _lclose16(hf16);
1194     return ret;
1195 }
1196
1197
1198 /***********************************************************************
1199  *           K32Thk1632Prolog                   (KERNEL32.@)
1200  */
1201 void WINAPI __regs_K32Thk1632Prolog( CONTEXT *context )
1202 {
1203    LPBYTE code = (LPBYTE)context->Eip - 5;
1204
1205    /* Arrrgh! SYSTHUNK.DLL just has to re-implement another method
1206       of 16->32 thunks instead of using one of the standard methods!
1207       This means that SYSTHUNK.DLL itself switches to a 32-bit stack,
1208       and does a far call to the 32-bit code segment of OLECLI32/OLESVR32.
1209       Unfortunately, our CallTo/CallFrom mechanism is therefore completely
1210       bypassed, which means it will crash the next time the 32-bit OLE
1211       code thunks down again to 16-bit (this *will* happen!).
1212
1213       The following hack tries to recognize this situation.
1214       This is possible since the called stubs in OLECLI32/OLESVR32 all
1215       look exactly the same:
1216         00   E8xxxxxxxx    call K32Thk1632Prolog
1217         05   FF55FC        call [ebp-04]
1218         08   E8xxxxxxxx    call K32Thk1632Epilog
1219         0D   66CB          retf
1220
1221       If we recognize this situation, we try to simulate the actions
1222       of our CallTo/CallFrom mechanism by copying the 16-bit stack
1223       to our 32-bit stack, creating a proper STACK16FRAME and
1224       updating cur_stack. */
1225
1226    if (   code[5] == 0xFF && code[6] == 0x55 && code[7] == 0xFC
1227        && code[13] == 0x66 && code[14] == 0xCB)
1228    {
1229       DWORD argSize = context->Ebp - context->Esp;
1230       char *stack16 = (char *)context->Esp - 4;
1231       STACK16FRAME *frame16 = (STACK16FRAME *)stack16 - 1;
1232       STACK32FRAME *frame32 = (STACK32FRAME *)kernel_get_thread_data()->stack;
1233       char *stack32 = (char *)frame32 - argSize;
1234       WORD  stackSel  = SELECTOROF(frame32->frame16);
1235       DWORD stackBase = GetSelectorBase(stackSel);
1236
1237       TRACE("before SYSTHUNK hack: EBP: %08x ESP: %08x cur_stack: %04x:%04x\n",
1238             context->Ebp, context->Esp, CURRENT_SS, CURRENT_SP);
1239
1240       memset(frame16, '\0', sizeof(STACK16FRAME));
1241       frame16->frame32 = frame32;
1242       frame16->ebp = context->Ebp;
1243
1244       memcpy(stack32, stack16, argSize);
1245       CURRENT_SS = stackSel;
1246       CURRENT_SP = (DWORD)frame16 - stackBase;
1247
1248       context->Esp = (DWORD)stack32 + 4;
1249       context->Ebp = context->Esp + argSize;
1250
1251       TRACE("after  SYSTHUNK hack: EBP: %08x ESP: %08x cur_stack: %04x:%04x\n",
1252             context->Ebp, context->Esp, CURRENT_SS, CURRENT_SP);
1253    }
1254
1255     /* entry_point is never used again once the entry point has
1256        been called.  Thus we re-use it to hold the Win16Lock count */
1257    ReleaseThunkLock(&CURRENT_STACK16->entry_point);
1258 }
1259 DEFINE_REGS_ENTRYPOINT( K32Thk1632Prolog )
1260
1261 /***********************************************************************
1262  *           K32Thk1632Epilog                   (KERNEL32.@)
1263  */
1264 void WINAPI __regs_K32Thk1632Epilog( CONTEXT *context )
1265 {
1266    LPBYTE code = (LPBYTE)context->Eip - 13;
1267
1268    RestoreThunkLock(CURRENT_STACK16->entry_point);
1269
1270    /* We undo the SYSTHUNK hack if necessary. See K32Thk1632Prolog. */
1271
1272    if (   code[5] == 0xFF && code[6] == 0x55 && code[7] == 0xFC
1273        && code[13] == 0x66 && code[14] == 0xCB)
1274    {
1275       STACK16FRAME *frame16 = CURRENT_STACK16;
1276       char *stack16 = (char *)(frame16 + 1);
1277       DWORD argSize = frame16->ebp - (DWORD)stack16;
1278       char *stack32 = (char *)frame16->frame32 - argSize;
1279
1280       DWORD nArgsPopped = context->Esp - (DWORD)stack32;
1281
1282       TRACE("before SYSTHUNK hack: EBP: %08x ESP: %08x cur_stack: %04x:%04x\n",
1283             context->Ebp, context->Esp, CURRENT_SS, CURRENT_SP);
1284
1285       kernel_get_thread_data()->stack = (SEGPTR)frame16->frame32;
1286
1287       context->Esp = (DWORD)stack16 + nArgsPopped;
1288       context->Ebp = frame16->ebp;
1289
1290       TRACE("after  SYSTHUNK hack: EBP: %08x ESP: %08x cur_stack: %04x:%04x\n",
1291             context->Ebp, context->Esp, CURRENT_SS, CURRENT_SP);
1292    }
1293 }
1294 DEFINE_REGS_ENTRYPOINT( K32Thk1632Epilog )
1295
1296 /*********************************************************************
1297  *                   PK16FNF [KERNEL32.91]
1298  *
1299  *  This routine fills in the supplied 13-byte (8.3 plus terminator)
1300  *  string buffer with the 8.3 filename of a recently loaded 16-bit
1301  *  module.  It is unknown exactly what modules trigger this
1302  *  mechanism or what purpose this serves.  Win98 Explorer (and
1303  *  probably also Win95 with IE 4 shell integration) calls this
1304  *  several times during initialization.
1305  *
1306  *  FIXME: find out what this really does and make it work.
1307  */
1308 void WINAPI PK16FNF(LPSTR strPtr)
1309 {
1310        FIXME("(%p): stub\n", strPtr);
1311
1312        /* fill in a fake filename that'll be easy to recognize */
1313        strcpy(strPtr, "WINESTUB.FIX");
1314 }
1315
1316 /***********************************************************************
1317  * 16->32 Flat Thunk routines:
1318  */
1319
1320 /***********************************************************************
1321  *              ThunkConnect16          (KERNEL.651)
1322  * Connects a 32bit and a 16bit thunkbuffer.
1323  */
1324 UINT WINAPI ThunkConnect16(
1325         LPSTR module16,              /* [in] name of win16 dll */
1326         LPSTR module32,              /* [in] name of win32 dll */
1327         HINSTANCE16 hInst16,         /* [in] hInst of win16 dll */
1328         DWORD dwReason,              /* [in] initialisation argument */
1329         struct ThunkDataCommon *TD,  /* [in/out] thunkbuffer */
1330         LPSTR thunkfun32,            /* [in] win32 thunkfunction */
1331         WORD cs                      /* [in] CS of win16 dll */
1332 ) {
1333     BOOL directionSL;
1334
1335     if (!strncmp(TD->magic, "SL01", 4))
1336     {
1337         directionSL = TRUE;
1338
1339         TRACE("SL01 thunk %s (%p) -> %s (%s), Reason: %d\n",
1340               module16, TD, module32, thunkfun32, dwReason);
1341     }
1342     else if (!strncmp(TD->magic, "LS01", 4))
1343     {
1344         directionSL = FALSE;
1345
1346         TRACE("LS01 thunk %s (%p) <- %s (%s), Reason: %d\n",
1347               module16, TD, module32, thunkfun32, dwReason);
1348     }
1349     else
1350     {
1351         ERR("Invalid magic %c%c%c%c\n",
1352             TD->magic[0], TD->magic[1], TD->magic[2], TD->magic[3]);
1353         return 0;
1354     }
1355
1356     switch (dwReason)
1357     {
1358         case DLL_PROCESS_ATTACH:
1359             if (directionSL)
1360             {
1361                 struct ThunkDataSL16 *SL16 = (struct ThunkDataSL16 *)TD;
1362                 struct ThunkDataSL   *SL   = SL16->fpData;
1363
1364                 if (SL == NULL)
1365                 {
1366                     SL = HeapAlloc(GetProcessHeap(), 0, sizeof(*SL));
1367
1368                     SL->common   = SL16->common;
1369                     SL->flags1   = SL16->flags1;
1370                     SL->flags2   = SL16->flags2;
1371
1372                     SL->apiDB    = MapSL(SL16->apiDatabase);
1373                     SL->targetDB = NULL;
1374
1375                     lstrcpynA(SL->pszDll16, module16, 255);
1376                     lstrcpynA(SL->pszDll32, module32, 255);
1377
1378                     /* We should create a SEGPTR to the ThunkDataSL,
1379                        but since the contents are not in the original format,
1380                        any access to this by 16-bit code would crash anyway. */
1381                     SL16->spData = 0;
1382                     SL16->fpData = SL;
1383                 }
1384
1385
1386                 if (SL->flags2 & 0x80000000)
1387                 {
1388                     TRACE("Preloading 32-bit library\n");
1389                     LoadLibraryA(module32);
1390                 }
1391             }
1392             else
1393             {
1394                 /* nothing to do */
1395             }
1396             break;
1397
1398         case DLL_PROCESS_DETACH:
1399             /* FIXME: cleanup */
1400             break;
1401     }
1402
1403     return 1;
1404 }
1405
1406
1407 /***********************************************************************
1408  *           C16ThkSL                           (KERNEL.630)
1409  */
1410
1411 void WINAPI C16ThkSL(CONTEXT *context)
1412 {
1413     LPBYTE stub = MapSL(context->Eax), x = stub;
1414
1415     /* We produce the following code:
1416      *
1417      *   mov ax, __FLATDS
1418      *   mov es, ax
1419      *   movzx ecx, cx
1420      *   mov edx, es:[ecx + $EDX]
1421      *   push bp
1422      *   push edx
1423      *   push dx
1424      *   push edx
1425      *   call __FLATCS:__wine_call_from_16_thunk
1426      */
1427
1428     *x++ = 0xB8; *(WORD *)x = get_ds(); x += sizeof(WORD);
1429     *x++ = 0x8E; *x++ = 0xC0;
1430     *x++ = 0x66; *x++ = 0x0F; *x++ = 0xB7; *x++ = 0xC9;
1431     *x++ = 0x67; *x++ = 0x66; *x++ = 0x26; *x++ = 0x8B;
1432                  *x++ = 0x91; *(DWORD *)x = context->Edx; x += sizeof(DWORD);
1433
1434     *x++ = 0x55;
1435     *x++ = 0x66; *x++ = 0x52;
1436     *x++ = 0x52;
1437     *x++ = 0x66; *x++ = 0x52;
1438     *x++ = 0x66; *x++ = 0x9A;
1439     *(void **)x = __wine_call_from_16_thunk; x += sizeof(void *);
1440     *(WORD *)x = get_cs(); x += sizeof(WORD);
1441
1442     /* Jump to the stub code just created */
1443     context->Eip = LOWORD(context->Eax);
1444     context->SegCs  = HIWORD(context->Eax);
1445
1446     /* Since C16ThkSL got called by a jmp, we need to leave the
1447        original return address on the stack */
1448     context->Esp -= 4;
1449 }
1450
1451 /***********************************************************************
1452  *           C16ThkSL01                         (KERNEL.631)
1453  */
1454
1455 void WINAPI C16ThkSL01(CONTEXT *context)
1456 {
1457     LPBYTE stub = MapSL(context->Eax), x = stub;
1458
1459     if (stub)
1460     {
1461         struct ThunkDataSL16 *SL16 = MapSL(context->Edx);
1462         struct ThunkDataSL *td = SL16->fpData;
1463
1464         DWORD procAddress = (DWORD)GetProcAddress16(GetModuleHandle16("KERNEL"), (LPCSTR)631);
1465
1466         if (!td)
1467         {
1468             ERR("ThunkConnect16 was not called!\n");
1469             return;
1470         }
1471
1472         TRACE("Creating stub for ThunkDataSL %p\n", td);
1473
1474
1475         /* We produce the following code:
1476          *
1477          *   xor eax, eax
1478          *   mov edx, $td
1479          *   call C16ThkSL01
1480          *   push bp
1481          *   push edx
1482          *   push dx
1483          *   push edx
1484          *   call __FLATCS:__wine_call_from_16_thunk
1485          */
1486
1487         *x++ = 0x66; *x++ = 0x33; *x++ = 0xC0;
1488         *x++ = 0x66; *x++ = 0xBA; *(void **)x = td; x += sizeof(void *);
1489         *x++ = 0x9A; *(DWORD *)x = procAddress; x += sizeof(DWORD);
1490
1491         *x++ = 0x55;
1492         *x++ = 0x66; *x++ = 0x52;
1493         *x++ = 0x52;
1494         *x++ = 0x66; *x++ = 0x52;
1495         *x++ = 0x66; *x++ = 0x9A;
1496         *(void **)x = __wine_call_from_16_thunk; x += sizeof(void *);
1497         *(WORD *)x = get_cs(); x += sizeof(WORD);
1498
1499         /* Jump to the stub code just created */
1500         context->Eip = LOWORD(context->Eax);
1501         context->SegCs  = HIWORD(context->Eax);
1502
1503         /* Since C16ThkSL01 got called by a jmp, we need to leave the
1504            original return address on the stack */
1505         context->Esp -= 4;
1506     }
1507     else
1508     {
1509         struct ThunkDataSL *td = (struct ThunkDataSL *)context->Edx;
1510         DWORD targetNr = LOWORD(context->Ecx) / 4;
1511         struct SLTargetDB *tdb;
1512
1513         TRACE("Process %08x calling target %d of ThunkDataSL %p\n",
1514               GetCurrentProcessId(), targetNr, td);
1515
1516         for (tdb = td->targetDB; tdb; tdb = tdb->next)
1517             if (tdb->process == GetCurrentProcessId())
1518                 break;
1519
1520         if (!tdb)
1521         {
1522             TRACE("Loading 32-bit library %s\n", td->pszDll32);
1523             LoadLibraryA(td->pszDll32);
1524
1525             for (tdb = td->targetDB; tdb; tdb = tdb->next)
1526                 if (tdb->process == GetCurrentProcessId())
1527                     break;
1528         }
1529
1530         if (tdb)
1531         {
1532             context->Edx = tdb->targetTable[targetNr];
1533
1534             TRACE("Call target is %08x\n", context->Edx);
1535         }
1536         else
1537         {
1538             WORD *stack = MapSL( MAKESEGPTR(context->SegSs, LOWORD(context->Esp)) );
1539             context->Edx = (context->Edx & ~0xffff) | HIWORD(td->apiDB[targetNr].errorReturnValue);
1540             context->Eax = (context->Eax & ~0xffff) | LOWORD(td->apiDB[targetNr].errorReturnValue);
1541             context->Eip = stack[2];
1542             context->SegCs  = stack[3];
1543             context->Esp += td->apiDB[targetNr].nrArgBytes + 4;
1544
1545             ERR("Process %08x did not ThunkConnect32 %s to %s\n",
1546                 GetCurrentProcessId(), td->pszDll32, td->pszDll16);
1547         }
1548     }
1549 }
1550
1551
1552 /***********************************************************************
1553  * 16<->32 Thunklet/Callback API:
1554  */
1555
1556 #include "pshpack1.h"
1557 typedef struct _THUNKLET
1558 {
1559     BYTE        prefix_target;
1560     BYTE        pushl_target;
1561     DWORD       target;
1562
1563     BYTE        prefix_relay;
1564     BYTE        pushl_relay;
1565     DWORD       relay;
1566
1567     BYTE        jmp_glue;
1568     DWORD       glue;
1569
1570     BYTE        type;
1571     HINSTANCE16 owner;
1572     struct _THUNKLET *next;
1573 } THUNKLET;
1574 #include "poppack.h"
1575
1576 #define THUNKLET_TYPE_LS  1
1577 #define THUNKLET_TYPE_SL  2
1578
1579 static HANDLE  ThunkletHeap = 0;
1580 static WORD ThunkletCodeSel;
1581 static THUNKLET *ThunkletAnchor = NULL;
1582
1583 static FARPROC ThunkletSysthunkGlueLS = 0;
1584 static SEGPTR    ThunkletSysthunkGlueSL = 0;
1585
1586 static FARPROC ThunkletCallbackGlueLS = 0;
1587 static SEGPTR    ThunkletCallbackGlueSL = 0;
1588
1589
1590 /* map a thunk allocated on ThunkletHeap to a 16-bit pointer */
1591 static inline SEGPTR get_segptr( void *thunk )
1592 {
1593     if (!thunk) return 0;
1594     return MAKESEGPTR( ThunkletCodeSel, (char *)thunk - (char *)ThunkletHeap );
1595 }
1596
1597 /***********************************************************************
1598  *           THUNK_Init
1599  */
1600 static BOOL THUNK_Init(void)
1601 {
1602     LPBYTE thunk;
1603
1604     ThunkletHeap = HeapCreate( HEAP_CREATE_ENABLE_EXECUTE, 0x10000, 0x10000 );
1605     if (!ThunkletHeap) return FALSE;
1606
1607     ThunkletCodeSel = SELECTOR_AllocBlock( ThunkletHeap, 0x10000, LDT_FLAGS_CODE );
1608
1609     thunk = HeapAlloc( ThunkletHeap, 0, 5 );
1610     if (!thunk) return FALSE;
1611
1612     ThunkletSysthunkGlueLS = (FARPROC)thunk;
1613     *thunk++ = 0x58;                             /* popl eax */
1614     *thunk++ = 0xC3;                             /* ret      */
1615
1616     ThunkletSysthunkGlueSL = get_segptr( thunk );
1617     *thunk++ = 0x66; *thunk++ = 0x58;            /* popl eax */
1618     *thunk++ = 0xCB;                             /* lret     */
1619
1620     return TRUE;
1621 }
1622
1623 /***********************************************************************
1624  *     SetThunkletCallbackGlue             (KERNEL.560)
1625  */
1626 void WINAPI SetThunkletCallbackGlue16( FARPROC glueLS, SEGPTR glueSL )
1627 {
1628     ThunkletCallbackGlueLS = glueLS;
1629     ThunkletCallbackGlueSL = glueSL;
1630 }
1631
1632
1633 /***********************************************************************
1634  *     THUNK_FindThunklet
1635  */
1636 static THUNKLET *THUNK_FindThunklet( DWORD target, DWORD relay,
1637                                      DWORD glue, BYTE type )
1638 {
1639     THUNKLET *thunk;
1640
1641     for (thunk = ThunkletAnchor; thunk; thunk = thunk->next)
1642         if (    thunk->type   == type
1643              && thunk->target == target
1644              && thunk->relay  == relay
1645              && ( type == THUNKLET_TYPE_LS ?
1646                     ( thunk->glue == glue - (DWORD)&thunk->type )
1647                   : ( thunk->glue == glue ) ) )
1648             return thunk;
1649
1650      return NULL;
1651 }
1652
1653 /***********************************************************************
1654  *     THUNK_AllocLSThunklet
1655  */
1656 static FARPROC THUNK_AllocLSThunklet( SEGPTR target, DWORD relay,
1657                                       FARPROC glue, HTASK16 owner )
1658 {
1659     THUNKLET *thunk = THUNK_FindThunklet( (DWORD)target, relay, (DWORD)glue,
1660                                           THUNKLET_TYPE_LS );
1661     if (!thunk)
1662     {
1663         TDB *pTask = GlobalLock16( owner );
1664
1665         if (!ThunkletHeap) THUNK_Init();
1666         if ( !(thunk = HeapAlloc( ThunkletHeap, 0, sizeof(THUNKLET) )) )
1667             return 0;
1668
1669         thunk->prefix_target = thunk->prefix_relay = 0x90;
1670         thunk->pushl_target  = thunk->pushl_relay  = 0x68;
1671         thunk->jmp_glue = 0xE9;
1672
1673         thunk->target  = (DWORD)target;
1674         thunk->relay   = relay;
1675         thunk->glue    = (DWORD)glue - (DWORD)&thunk->type;
1676
1677         thunk->type    = THUNKLET_TYPE_LS;
1678         thunk->owner   = pTask? pTask->hInstance : 0;
1679
1680         thunk->next    = ThunkletAnchor;
1681         ThunkletAnchor = thunk;
1682     }
1683
1684     return (FARPROC)thunk;
1685 }
1686
1687 /***********************************************************************
1688  *     THUNK_AllocSLThunklet
1689  */
1690 static SEGPTR THUNK_AllocSLThunklet( FARPROC target, DWORD relay,
1691                                      SEGPTR glue, HTASK16 owner )
1692 {
1693     THUNKLET *thunk = THUNK_FindThunklet( (DWORD)target, relay, (DWORD)glue,
1694                                           THUNKLET_TYPE_SL );
1695     if (!thunk)
1696     {
1697         TDB *pTask = GlobalLock16( owner );
1698
1699         if (!ThunkletHeap) THUNK_Init();
1700         if ( !(thunk = HeapAlloc( ThunkletHeap, 0, sizeof(THUNKLET) )) )
1701             return 0;
1702
1703         thunk->prefix_target = thunk->prefix_relay = 0x66;
1704         thunk->pushl_target  = thunk->pushl_relay  = 0x68;
1705         thunk->jmp_glue = 0xEA;
1706
1707         thunk->target  = (DWORD)target;
1708         thunk->relay   = relay;
1709         thunk->glue    = (DWORD)glue;
1710
1711         thunk->type    = THUNKLET_TYPE_SL;
1712         thunk->owner   = pTask? pTask->hInstance : 0;
1713
1714         thunk->next    = ThunkletAnchor;
1715         ThunkletAnchor = thunk;
1716     }
1717
1718     return get_segptr( thunk );
1719 }
1720
1721 /**********************************************************************
1722  *     IsLSThunklet
1723  */
1724 static BOOL16 IsLSThunklet( THUNKLET *thunk )
1725 {
1726     return    thunk->prefix_target == 0x90 && thunk->pushl_target == 0x68
1727            && thunk->prefix_relay  == 0x90 && thunk->pushl_relay  == 0x68
1728            && thunk->jmp_glue == 0xE9 && thunk->type == THUNKLET_TYPE_LS;
1729 }
1730
1731 /**********************************************************************
1732  *     IsSLThunklet                        (KERNEL.612)
1733  */
1734 BOOL16 WINAPI IsSLThunklet16( THUNKLET *thunk )
1735 {
1736     return    thunk->prefix_target == 0x66 && thunk->pushl_target == 0x68
1737            && thunk->prefix_relay  == 0x66 && thunk->pushl_relay  == 0x68
1738            && thunk->jmp_glue == 0xEA && thunk->type == THUNKLET_TYPE_SL;
1739 }
1740
1741
1742
1743 /***********************************************************************
1744  *     AllocLSThunkletSysthunk             (KERNEL.607)
1745  */
1746 FARPROC WINAPI AllocLSThunkletSysthunk16( SEGPTR target,
1747                                           FARPROC relay, DWORD dummy )
1748 {
1749     if (!ThunkletSysthunkGlueLS) THUNK_Init();
1750     return THUNK_AllocLSThunklet( (SEGPTR)relay, (DWORD)target,
1751                                   ThunkletSysthunkGlueLS, GetCurrentTask() );
1752 }
1753
1754 /***********************************************************************
1755  *     AllocSLThunkletSysthunk             (KERNEL.608)
1756  */
1757 SEGPTR WINAPI AllocSLThunkletSysthunk16( FARPROC target,
1758                                        SEGPTR relay, DWORD dummy )
1759 {
1760     if (!ThunkletSysthunkGlueSL) THUNK_Init();
1761     return THUNK_AllocSLThunklet( (FARPROC)relay, (DWORD)target,
1762                                   ThunkletSysthunkGlueSL, GetCurrentTask() );
1763 }
1764
1765
1766 /***********************************************************************
1767  *     AllocLSThunkletCallbackEx           (KERNEL.567)
1768  */
1769 FARPROC WINAPI AllocLSThunkletCallbackEx16( SEGPTR target,
1770                                             DWORD relay, HTASK16 task )
1771 {
1772     THUNKLET *thunk = MapSL( target );
1773     if ( !thunk ) return NULL;
1774
1775     if (   IsSLThunklet16( thunk ) && thunk->relay == relay
1776         && thunk->glue == (DWORD)ThunkletCallbackGlueSL )
1777         return (FARPROC)thunk->target;
1778
1779     return THUNK_AllocLSThunklet( target, relay,
1780                                   ThunkletCallbackGlueLS, task );
1781 }
1782
1783 /***********************************************************************
1784  *     AllocSLThunkletCallbackEx           (KERNEL.568)
1785  */
1786 SEGPTR WINAPI AllocSLThunkletCallbackEx16( FARPROC target,
1787                                          DWORD relay, HTASK16 task )
1788 {
1789     THUNKLET *thunk = (THUNKLET *)target;
1790     if ( !thunk ) return 0;
1791
1792     if (   IsLSThunklet( thunk ) && thunk->relay == relay
1793         && thunk->glue == (DWORD)ThunkletCallbackGlueLS - (DWORD)&thunk->type )
1794         return (SEGPTR)thunk->target;
1795
1796     return THUNK_AllocSLThunklet( target, relay,
1797                                   ThunkletCallbackGlueSL, task );
1798 }
1799
1800 /***********************************************************************
1801  *     AllocLSThunkletCallback             (KERNEL.561)
1802  *     AllocLSThunkletCallback_dup         (KERNEL.606)
1803  */
1804 FARPROC WINAPI AllocLSThunkletCallback16( SEGPTR target, DWORD relay )
1805 {
1806     return AllocLSThunkletCallbackEx16( target, relay, GetCurrentTask() );
1807 }
1808
1809 /***********************************************************************
1810  *     AllocSLThunkletCallback             (KERNEL.562)
1811  *     AllocSLThunkletCallback_dup         (KERNEL.605)
1812  */
1813 SEGPTR WINAPI AllocSLThunkletCallback16( FARPROC target, DWORD relay )
1814 {
1815     return AllocSLThunkletCallbackEx16( target, relay, GetCurrentTask() );
1816 }
1817
1818 /***********************************************************************
1819  *     FindLSThunkletCallback              (KERNEL.563)
1820  *     FindLSThunkletCallback_dup          (KERNEL.609)
1821  */
1822 FARPROC WINAPI FindLSThunkletCallback( SEGPTR target, DWORD relay )
1823 {
1824     THUNKLET *thunk = MapSL( target );
1825     if (   thunk && IsSLThunklet16( thunk ) && thunk->relay == relay
1826         && thunk->glue == (DWORD)ThunkletCallbackGlueSL )
1827         return (FARPROC)thunk->target;
1828
1829     thunk = THUNK_FindThunklet( (DWORD)target, relay,
1830                                 (DWORD)ThunkletCallbackGlueLS,
1831                                 THUNKLET_TYPE_LS );
1832     return (FARPROC)thunk;
1833 }
1834
1835 /***********************************************************************
1836  *     FindSLThunkletCallback              (KERNEL.564)
1837  *     FindSLThunkletCallback_dup          (KERNEL.610)
1838  */
1839 SEGPTR WINAPI FindSLThunkletCallback( FARPROC target, DWORD relay )
1840 {
1841     THUNKLET *thunk = (THUNKLET *)target;
1842     if (   thunk && IsLSThunklet( thunk ) && thunk->relay == relay
1843         && thunk->glue == (DWORD)ThunkletCallbackGlueLS - (DWORD)&thunk->type )
1844         return (SEGPTR)thunk->target;
1845
1846     thunk = THUNK_FindThunklet( (DWORD)target, relay,
1847                                 (DWORD)ThunkletCallbackGlueSL,
1848                                 THUNKLET_TYPE_SL );
1849     return get_segptr( thunk );
1850 }
1851
1852
1853 /***********************************************************************
1854  *     FreeThunklet            (KERNEL.611)
1855  */
1856 BOOL16 WINAPI FreeThunklet16( DWORD unused1, DWORD unused2 )
1857 {
1858     return FALSE;
1859 }
1860
1861
1862 /***********************************************************************
1863  * Callback Client API
1864  */
1865
1866 #define N_CBC_FIXED    20
1867 #define N_CBC_VARIABLE 10
1868 #define N_CBC_TOTAL    (N_CBC_FIXED + N_CBC_VARIABLE)
1869
1870 static SEGPTR CBClientRelay16[ N_CBC_TOTAL ];
1871 static FARPROC *CBClientRelay32[ N_CBC_TOTAL ];
1872
1873 /***********************************************************************
1874  *     RegisterCBClient                    (KERNEL.619)
1875  */
1876 INT16 WINAPI RegisterCBClient16( INT16 wCBCId,
1877                                  SEGPTR relay16, FARPROC *relay32 )
1878 {
1879     /* Search for free Callback ID */
1880     if ( wCBCId == -1 )
1881         for ( wCBCId = N_CBC_FIXED; wCBCId < N_CBC_TOTAL; wCBCId++ )
1882             if ( !CBClientRelay16[ wCBCId ] )
1883                 break;
1884
1885     /* Register Callback ID */
1886     if ( wCBCId > 0 && wCBCId < N_CBC_TOTAL )
1887     {
1888         CBClientRelay16[ wCBCId ] = relay16;
1889         CBClientRelay32[ wCBCId ] = relay32;
1890     }
1891     else
1892         wCBCId = 0;
1893
1894     return wCBCId;
1895 }
1896
1897 /***********************************************************************
1898  *     UnRegisterCBClient                  (KERNEL.622)
1899  */
1900 INT16 WINAPI UnRegisterCBClient16( INT16 wCBCId,
1901                                    SEGPTR relay16, FARPROC *relay32 )
1902 {
1903     if (    wCBCId >= N_CBC_FIXED && wCBCId < N_CBC_TOTAL
1904          && CBClientRelay16[ wCBCId ] == relay16
1905          && CBClientRelay32[ wCBCId ] == relay32 )
1906     {
1907         CBClientRelay16[ wCBCId ] = 0;
1908         CBClientRelay32[ wCBCId ] = 0;
1909     }
1910     else
1911         wCBCId = 0;
1912
1913     return wCBCId;
1914 }
1915
1916
1917 /***********************************************************************
1918  *     InitCBClient                        (KERNEL.623)
1919  */
1920 void WINAPI InitCBClient16( FARPROC glueLS )
1921 {
1922     HMODULE16 kernel = GetModuleHandle16( "KERNEL" );
1923     SEGPTR glueSL = (SEGPTR)GetProcAddress16( kernel, (LPCSTR)604 );
1924
1925     SetThunkletCallbackGlue16( glueLS, glueSL );
1926 }
1927
1928 /***********************************************************************
1929  *     CBClientGlueSL                      (KERNEL.604)
1930  */
1931 void WINAPI CBClientGlueSL( CONTEXT *context )
1932 {
1933     /* Create stack frame */
1934     SEGPTR stackSeg = stack16_push( 12 );
1935     LPWORD stackLin = MapSL( stackSeg );
1936     SEGPTR glue, *glueTab;
1937
1938     stackLin[3] = (WORD)context->Ebp;
1939     stackLin[2] = (WORD)context->Esi;
1940     stackLin[1] = (WORD)context->Edi;
1941     stackLin[0] = (WORD)context->SegDs;
1942
1943     context->Ebp = OFFSETOF( stackSeg ) + 6;
1944     context->Esp = OFFSETOF( stackSeg ) - 4;
1945     context->SegGs = 0;
1946
1947     /* Jump to 16-bit relay code */
1948     glueTab = MapSL( CBClientRelay16[ stackLin[5] ] );
1949     glue = glueTab[ stackLin[4] ];
1950     context->SegCs = SELECTOROF( glue );
1951     context->Eip   = OFFSETOF  ( glue );
1952 }
1953
1954 /***********************************************************************
1955  *     CBClientThunkSL                      (KERNEL.620)
1956  */
1957 extern DWORD CALL32_CBClient( FARPROC proc, LPWORD args, WORD *stackLin, DWORD *esi );
1958 void WINAPI CBClientThunkSL( CONTEXT *context )
1959 {
1960     /* Call 32-bit relay code */
1961
1962     LPWORD args = MapSL( MAKESEGPTR( context->SegSs, LOWORD(context->Ebp) ) );
1963     FARPROC proc = CBClientRelay32[ args[2] ][ args[1] ];
1964
1965     /* fill temporary area for the asm code (see comments in winebuild) */
1966     SEGPTR stack = stack16_push( 12 );
1967     LPWORD stackLin = MapSL(stack);
1968     /* stackLin[0] and stackLin[1] reserved for the 32-bit stack ptr */
1969     stackLin[2] = get_ds();
1970     stackLin[3] = 0;
1971     stackLin[4] = OFFSETOF(stack) + 12;
1972     stackLin[5] = SELECTOROF(stack);
1973     stackLin[6] = OFFSETOF(CALL32_CBClientEx_RetAddr);  /* overwrite return address */
1974     stackLin[7] = SELECTOROF(CALL32_CBClientEx_RetAddr);
1975     context->Eax = CALL32_CBClient( proc, args, stackLin + 4, &context->Esi );
1976     stack16_pop( 12 );
1977 }
1978
1979 /***********************************************************************
1980  *     CBClientThunkSLEx                    (KERNEL.621)
1981  */
1982 extern DWORD CALL32_CBClientEx( FARPROC proc, LPWORD args, WORD *stackLin, DWORD *esi, INT *nArgs );
1983 void WINAPI CBClientThunkSLEx( CONTEXT *context )
1984 {
1985     /* Call 32-bit relay code */
1986
1987     LPWORD args = MapSL( MAKESEGPTR( context->SegSs, LOWORD(context->Ebp) ) );
1988     FARPROC proc = CBClientRelay32[ args[2] ][ args[1] ];
1989     INT nArgs;
1990     LPWORD stackLin;
1991
1992     /* fill temporary area for the asm code (see comments in winebuild) */
1993     SEGPTR stack = stack16_push( 24 );
1994     stackLin = MapSL(stack);
1995     stackLin[0] = OFFSETOF(stack) + 4;
1996     stackLin[1] = SELECTOROF(stack);
1997     stackLin[2] = get_ds();
1998     stackLin[5] = OFFSETOF(stack) + 24;
1999     /* stackLin[6] and stackLin[7] reserved for the 32-bit stack ptr */
2000     stackLin[8] = get_ds();
2001     stackLin[9] = 0;
2002     stackLin[10] = OFFSETOF(CALL32_CBClientEx_RetAddr);
2003     stackLin[11] = SELECTOROF(CALL32_CBClientEx_RetAddr);
2004
2005     context->Eax = CALL32_CBClientEx( proc, args, stackLin, &context->Esi, &nArgs );
2006     stack16_pop( 24 );
2007
2008     /* Restore registers saved by CBClientGlueSL */
2009     stackLin = (LPWORD)((LPBYTE)CURRENT_STACK16 + sizeof(STACK16FRAME) - 4);
2010     context->Ebp = (context->Ebp & ~0xffff) | stackLin[3];
2011     context->Esi = (context->Esi & ~0xffff) | stackLin[2];
2012     context->Edi = (context->Edi & ~0xffff) | stackLin[1];
2013     context->SegDs = stackLin[0];
2014     context->Esp += 16+nArgs;
2015
2016     /* Return to caller of CBClient thunklet */
2017     context->SegCs = stackLin[9];
2018     context->Eip   = stackLin[8];
2019 }
2020
2021
2022 /***********************************************************************
2023  *           Get16DLLAddress       (KERNEL32.@)
2024  *
2025  * This function is used by a Win32s DLL if it wants to call a Win16 function.
2026  * A 16:16 segmented pointer to the function is returned.
2027  * Written without any docu.
2028  */
2029 SEGPTR WINAPI Get16DLLAddress(HMODULE16 handle, LPSTR func_name)
2030 {
2031     static WORD code_sel32;
2032     FARPROC16 proc_16;
2033     LPBYTE thunk;
2034
2035     if (!code_sel32)
2036     {
2037         if (!ThunkletHeap) THUNK_Init();
2038         code_sel32 = SELECTOR_AllocBlock( ThunkletHeap, 0x10000, LDT_FLAGS_CODE | LDT_FLAGS_32BIT );
2039         if (!code_sel32) return 0;
2040     }
2041     if (!(thunk = HeapAlloc( ThunkletHeap, 0, 32 ))) return 0;
2042
2043     if (!handle) handle = GetModuleHandle16("WIN32S16");
2044     proc_16 = GetProcAddress16(handle, func_name);
2045
2046     /* movl proc_16, $edx */
2047     *thunk++ = 0xba;
2048     *(FARPROC16 *)thunk = proc_16;
2049     thunk += sizeof(FARPROC16);
2050
2051      /* jmpl QT_Thunk */
2052     *thunk++ = 0xea;
2053     *(void **)thunk = QT_Thunk;
2054     thunk += sizeof(FARPROC16);
2055     *(WORD *)thunk = get_cs();
2056
2057     return MAKESEGPTR( code_sel32, (char *)thunk - (char *)ThunkletHeap );
2058 }
2059
2060
2061 /***********************************************************************
2062  *              GetWin16DOSEnv                  (KERNEL32.34)
2063  * Returns some internal value.... probably the default environment database?
2064  */
2065 DWORD WINAPI GetWin16DOSEnv(void)
2066 {
2067         FIXME("stub, returning 0\n");
2068         return 0;
2069 }
2070
2071 /**********************************************************************
2072  *           GetPK16SysVar    (KERNEL32.92)
2073  */
2074 LPVOID WINAPI GetPK16SysVar(void)
2075 {
2076     static BYTE PK16SysVar[128];
2077
2078     FIXME("()\n");
2079     return PK16SysVar;
2080 }
2081
2082 /**********************************************************************
2083  *           CommonUnimpStub    (KERNEL32.17)
2084  */
2085 int WINAPI DECLSPEC_HIDDEN __regs_CommonUnimpStub( const char *name, int type )
2086 {
2087     FIXME("generic stub %s\n", debugstr_a(name));
2088
2089     switch (type)
2090     {
2091     case 15:  return -1;
2092     case 14:  return ERROR_CALL_NOT_IMPLEMENTED;
2093     case 13:  return ERROR_NOT_SUPPORTED;
2094     case 1:   return 1;
2095     default:  return 0;
2096     }
2097 }
2098 __ASM_STDCALL_FUNC( CommonUnimpStub, 0,
2099                     "pushl %ecx\n\t"
2100                     __ASM_CFI(".cfi_adjust_cfa_offset 4\n\t")
2101                     "shrl $4,%ecx\n\t"
2102                     "andl $0xf,%ecx\n\t"
2103                     "pushl %ecx\n\t"
2104                     __ASM_CFI(".cfi_adjust_cfa_offset 4\n\t")
2105                     "pushl %eax\n\t"
2106                     __ASM_CFI(".cfi_adjust_cfa_offset 4\n\t")
2107                     "call " __ASM_STDCALL("__regs_CommonUnimpStub",8) "\n\t"
2108                     __ASM_CFI(".cfi_adjust_cfa_offset -8\n\t")
2109                     "popl %ecx\n\t"
2110                     __ASM_CFI(".cfi_adjust_cfa_offset -4\n\t")
2111                     "andl $0xf,%ecx\n\t"
2112                     "movl (%esp),%edx\n\t"
2113                     "leal (%esp,%ecx,4),%esp\n\t"
2114                     "movl %edx,(%esp)\n\t"
2115                     "ret" )
2116
2117 /**********************************************************************
2118  *           HouseCleanLogicallyDeadHandles    (KERNEL32.33)
2119  */
2120 void WINAPI HouseCleanLogicallyDeadHandles(void)
2121 {
2122     /* Whatever this is supposed to do, our handles probably
2123        don't need it :-) */
2124 }
2125
2126 /**********************************************************************
2127  *              @ (KERNEL32.100)
2128  */
2129 BOOL WINAPI _KERNEL32_100(HANDLE threadid,DWORD exitcode,DWORD x)
2130 {
2131         FIXME("(%p,%d,0x%08x): stub\n",threadid,exitcode,x);
2132         return TRUE;
2133 }
2134
2135 /**********************************************************************
2136  *              @ (KERNEL32.99)
2137  *
2138  * Checks whether the clock has to be switched from daylight
2139  * savings time to standard time or vice versa.
2140  *
2141  */
2142 DWORD WINAPI _KERNEL32_99(DWORD x)
2143 {
2144         FIXME("(0x%08x): stub\n",x);
2145         return 1;
2146 }
2147
2148
2149 /***********************************************************************
2150  * Helper for k32 family functions
2151  */
2152 static void *user32_proc_address(const char *proc_name)
2153 {
2154     static HMODULE hUser32;
2155
2156     if(!hUser32) hUser32 = LoadLibraryA("user32.dll");
2157     return GetProcAddress(hUser32, proc_name);
2158 }
2159
2160 /***********************************************************************
2161  *              k32CharToOemBuffA   (KERNEL32.11)
2162  */
2163 BOOL WINAPI k32CharToOemBuffA(LPCSTR s, LPSTR d, DWORD len)
2164 {
2165     WCHAR *bufW;
2166
2167     if ((bufW = HeapAlloc( GetProcessHeap(), 0, len * sizeof(WCHAR) )))
2168     {
2169         MultiByteToWideChar( CP_ACP, 0, s, len, bufW, len );
2170         WideCharToMultiByte( CP_OEMCP, 0, bufW, len, d, len, NULL, NULL );
2171         HeapFree( GetProcessHeap(), 0, bufW );
2172     }
2173     return TRUE;
2174 }
2175
2176 /***********************************************************************
2177  *              k32CharToOemA   (KERNEL32.10)
2178  */
2179 BOOL WINAPI k32CharToOemA(LPCSTR s, LPSTR d)
2180 {
2181     if (!s || !d) return TRUE;
2182     return k32CharToOemBuffA( s, d, strlen(s) + 1 );
2183 }
2184
2185 /***********************************************************************
2186  *              k32OemToCharBuffA   (KERNEL32.13)
2187  */
2188 BOOL WINAPI k32OemToCharBuffA(LPCSTR s, LPSTR d, DWORD len)
2189 {
2190     WCHAR *bufW;
2191
2192     if ((bufW = HeapAlloc( GetProcessHeap(), 0, len * sizeof(WCHAR) )))
2193     {
2194         MultiByteToWideChar( CP_OEMCP, 0, s, len, bufW, len );
2195         WideCharToMultiByte( CP_ACP, 0, bufW, len, d, len, NULL, NULL );
2196         HeapFree( GetProcessHeap(), 0, bufW );
2197     }
2198     return TRUE;
2199 }
2200
2201 /***********************************************************************
2202  *              k32OemToCharA   (KERNEL32.12)
2203  */
2204 BOOL WINAPI k32OemToCharA(LPCSTR s, LPSTR d)
2205 {
2206     return k32OemToCharBuffA( s, d, strlen(s) + 1 );
2207 }
2208
2209 /**********************************************************************
2210  *              k32LoadStringA   (KERNEL32.14)
2211  */
2212 INT WINAPI k32LoadStringA(HINSTANCE instance, UINT resource_id,
2213                           LPSTR buffer, INT buflen)
2214 {
2215     static INT (WINAPI *pLoadStringA)(HINSTANCE, UINT, LPSTR, INT);
2216
2217     if(!pLoadStringA) pLoadStringA = user32_proc_address("LoadStringA");
2218     return pLoadStringA(instance, resource_id, buffer, buflen);
2219 }
2220
2221 /***********************************************************************
2222  *              k32wvsprintfA   (KERNEL32.16)
2223  */
2224 INT WINAPI k32wvsprintfA(LPSTR buffer, LPCSTR spec, __ms_va_list args)
2225 {
2226     static INT (WINAPI *pwvsprintfA)(LPSTR, LPCSTR, __ms_va_list);
2227
2228     if(!pwvsprintfA) pwvsprintfA = user32_proc_address("wvsprintfA");
2229     return (*pwvsprintfA)(buffer, spec, args);
2230 }
2231
2232 /***********************************************************************
2233  *              k32wsprintfA   (KERNEL32.15)
2234  */
2235 INT WINAPIV k32wsprintfA(LPSTR buffer, LPCSTR spec, ...)
2236 {
2237     __ms_va_list args;
2238     INT res;
2239
2240     __ms_va_start(args, spec);
2241     res = k32wvsprintfA(buffer, spec, args);
2242     __ms_va_end(args);
2243     return res;
2244 }
2245
2246 /**********************************************************************
2247  *           Catch    (KERNEL.55)
2248  *
2249  * Real prototype is:
2250  *   INT16 WINAPI Catch( LPCATCHBUF lpbuf );
2251  */
2252 void WINAPI Catch16( LPCATCHBUF lpbuf, CONTEXT *context )
2253 {
2254     /* Note: we don't save the current ss, as the catch buffer is */
2255     /* only 9 words long. Hopefully no one will have the silly    */
2256     /* idea to change the current stack before calling Throw()... */
2257
2258     /* Windows uses:
2259      * lpbuf[0] = ip
2260      * lpbuf[1] = cs
2261      * lpbuf[2] = sp
2262      * lpbuf[3] = bp
2263      * lpbuf[4] = si
2264      * lpbuf[5] = di
2265      * lpbuf[6] = ds
2266      * lpbuf[7] = unused
2267      * lpbuf[8] = ss
2268      */
2269
2270     lpbuf[0] = LOWORD(context->Eip);
2271     lpbuf[1] = context->SegCs;
2272     /* Windows pushes 4 more words before saving sp */
2273     lpbuf[2] = LOWORD(context->Esp) - 4 * sizeof(WORD);
2274     lpbuf[3] = LOWORD(context->Ebp);
2275     lpbuf[4] = LOWORD(context->Esi);
2276     lpbuf[5] = LOWORD(context->Edi);
2277     lpbuf[6] = context->SegDs;
2278     lpbuf[7] = 0;
2279     lpbuf[8] = context->SegSs;
2280     context->Eax &= ~0xffff;  /* Return 0 */
2281 }
2282
2283
2284 /**********************************************************************
2285  *           Throw    (KERNEL.56)
2286  *
2287  * Real prototype is:
2288  *   INT16 WINAPI Throw( LPCATCHBUF lpbuf, INT16 retval );
2289  */
2290 void WINAPI Throw16( LPCATCHBUF lpbuf, INT16 retval, CONTEXT *context )
2291 {
2292     STACK16FRAME *pFrame;
2293     STACK32FRAME *frame32;
2294
2295     context->Eax = (context->Eax & ~0xffff) | (WORD)retval;
2296
2297     /* Find the frame32 corresponding to the frame16 we are jumping to */
2298     pFrame = CURRENT_STACK16;
2299     frame32 = pFrame->frame32;
2300     while (frame32 && frame32->frame16)
2301     {
2302         if (OFFSETOF(frame32->frame16) < CURRENT_SP)
2303             break;  /* Something strange is going on */
2304         if (OFFSETOF(frame32->frame16) > lpbuf[2])
2305         {
2306             /* We found the right frame */
2307             pFrame->frame32 = frame32;
2308             break;
2309         }
2310         frame32 = ((STACK16FRAME *)MapSL(frame32->frame16))->frame32;
2311     }
2312     RtlUnwind( &pFrame->frame32->frame, NULL, NULL, 0 );
2313
2314     context->Eip = lpbuf[0];
2315     context->SegCs  = lpbuf[1];
2316     context->Esp = lpbuf[2] + 4 * sizeof(WORD) - sizeof(WORD) /*extra arg*/;
2317     context->Ebp = lpbuf[3];
2318     context->Esi = lpbuf[4];
2319     context->Edi = lpbuf[5];
2320     context->SegDs  = lpbuf[6];
2321
2322     if (lpbuf[8] != context->SegSs)
2323         ERR("Switching stack segment with Throw() not supported; expect crash now\n" );
2324 }
2325
2326
2327 /*
2328  *  16-bit WOW routines (in KERNEL)
2329  */
2330
2331 /**********************************************************************
2332  *           GetVDMPointer32W      (KERNEL.516)
2333  */
2334 DWORD WINAPI GetVDMPointer32W16( SEGPTR vp, UINT16 fMode )
2335 {
2336     GlobalPageLock16(GlobalHandle16(SELECTOROF(vp)));
2337     return (DWORD)K32WOWGetVDMPointer( vp, 0, (DWORD)fMode );
2338 }
2339
2340 /***********************************************************************
2341  *           LoadLibraryEx32W      (KERNEL.513)
2342  */
2343 DWORD WINAPI LoadLibraryEx32W16( LPCSTR lpszLibFile, DWORD hFile, DWORD dwFlags )
2344 {
2345     HMODULE hModule;
2346     DWORD mutex_count;
2347     OFSTRUCT ofs;
2348     const char *p;
2349
2350     if (!lpszLibFile)
2351     {
2352         SetLastError(ERROR_INVALID_PARAMETER);
2353         return 0;
2354     }
2355
2356     /* if the file cannot be found, call LoadLibraryExA anyway, since it might be
2357        a builtin module. This case is handled in MODULE_LoadLibraryExA */
2358
2359     if ((p = strrchr( lpszLibFile, '.' )) && !strchr( p, '\\' ))  /* got an extension */
2360     {
2361         if (OpenFile16( lpszLibFile, &ofs, OF_EXIST ) != HFILE_ERROR16)
2362             lpszLibFile = ofs.szPathName;
2363     }
2364     else
2365     {
2366         char buffer[MAX_PATH+4];
2367         strcpy( buffer, lpszLibFile );
2368         strcat( buffer, ".dll" );
2369         if (OpenFile16( buffer, &ofs, OF_EXIST ) != HFILE_ERROR16)
2370             lpszLibFile = ofs.szPathName;
2371     }
2372
2373     ReleaseThunkLock( &mutex_count );
2374     hModule = LoadLibraryExA( lpszLibFile, (HANDLE)hFile, dwFlags );
2375     RestoreThunkLock( mutex_count );
2376
2377     return (DWORD)hModule;
2378 }
2379
2380 /***********************************************************************
2381  *           GetProcAddress32W     (KERNEL.515)
2382  */
2383 DWORD WINAPI GetProcAddress32W16( DWORD hModule, LPCSTR lpszProc )
2384 {
2385     return (DWORD)GetProcAddress( (HMODULE)hModule, lpszProc );
2386 }
2387
2388 /***********************************************************************
2389  *           FreeLibrary32W        (KERNEL.514)
2390  */
2391 DWORD WINAPI FreeLibrary32W16( DWORD hLibModule )
2392 {
2393     BOOL retv;
2394     DWORD mutex_count;
2395
2396     ReleaseThunkLock( &mutex_count );
2397     retv = FreeLibrary( (HMODULE)hLibModule );
2398     RestoreThunkLock( mutex_count );
2399     return (DWORD)retv;
2400 }
2401
2402 /**********************************************************************
2403  *           WOW_CallProc32W
2404  */
2405 static DWORD WOW_CallProc32W16( FARPROC proc32, DWORD nrofargs, DWORD *args )
2406 {
2407     DWORD ret;
2408     DWORD mutex_count;
2409
2410     ReleaseThunkLock( &mutex_count );
2411     if (!proc32) ret = 0;
2412     else ret = call_entry_point( proc32, nrofargs & ~CPEX_DEST_CDECL, args );
2413     RestoreThunkLock( mutex_count );
2414
2415     TRACE("returns %08x\n",ret);
2416     return ret;
2417 }
2418
2419 /**********************************************************************
2420  *           CallProc32W           (KERNEL.517)
2421  */
2422 DWORD WINAPIV CallProc32W16( DWORD nrofargs, DWORD argconvmask, FARPROC proc32, VA_LIST16 valist )
2423 {
2424     DWORD args[32];
2425     unsigned int i;
2426
2427     TRACE("(%d,%d,%p args[",nrofargs,argconvmask,proc32);
2428
2429     for (i=0;i<nrofargs;i++)
2430     {
2431         if (argconvmask & (1<<i))
2432         {
2433             SEGPTR ptr = VA_ARG16( valist, SEGPTR );
2434             /* pascal convention, have to reverse the arguments order */
2435             args[nrofargs - i - 1] = (DWORD)MapSL(ptr);
2436             TRACE("%08x(%p),",ptr,MapSL(ptr));
2437         }
2438         else
2439         {
2440             DWORD arg = VA_ARG16( valist, DWORD );
2441             /* pascal convention, have to reverse the arguments order */
2442             args[nrofargs - i - 1] = arg;
2443             TRACE("%d,", arg);
2444         }
2445     }
2446     TRACE("])\n");
2447
2448     /* POP nrofargs DWORD arguments and 3 DWORD parameters */
2449     stack16_pop( (3 + nrofargs) * sizeof(DWORD) );
2450
2451     return WOW_CallProc32W16( proc32, nrofargs, args );
2452 }
2453
2454 /**********************************************************************
2455  *           _CallProcEx32W         (KERNEL.518)
2456  */
2457 DWORD WINAPIV CallProcEx32W16( DWORD nrofargs, DWORD argconvmask, FARPROC proc32, VA_LIST16 valist )
2458 {
2459     DWORD args[32];
2460     unsigned int i, count = min( 32, nrofargs & ~CPEX_DEST_CDECL );
2461
2462     TRACE("(%s,%d,%d,%p args[", nrofargs & CPEX_DEST_CDECL ? "cdecl": "stdcall",
2463           nrofargs & ~CPEX_DEST_CDECL, argconvmask, proc32);
2464
2465     for (i = 0; i < count; i++)
2466     {
2467         if (argconvmask & (1<<i))
2468         {
2469             SEGPTR ptr = VA_ARG16( valist, SEGPTR );
2470             args[i] = (DWORD)MapSL(ptr);
2471             TRACE("%08x(%p),",ptr,MapSL(ptr));
2472         }
2473         else
2474         {
2475             DWORD arg = VA_ARG16( valist, DWORD );
2476             args[i] = arg;
2477             TRACE("%d,", arg);
2478         }
2479     }
2480     TRACE("])\n");
2481     return WOW_CallProc32W16( proc32, nrofargs, args );
2482 }
2483
2484
2485 /**********************************************************************
2486  *           WOW16Call               (KERNEL.500)
2487  *
2488  * FIXME!!!
2489  *
2490  */
2491 DWORD WINAPIV WOW16Call(WORD x, WORD y, WORD z, VA_LIST16 args)
2492 {
2493         int     i;
2494         DWORD   calladdr;
2495         FIXME("(0x%04x,0x%04x,%d),calling (",x,y,z);
2496
2497         for (i=0;i<x/2;i++) {
2498                 WORD    a = VA_ARG16(args,WORD);
2499                 FIXME("%04x ",a);
2500         }
2501         calladdr = VA_ARG16(args,DWORD);
2502         stack16_pop( 3*sizeof(WORD) + x + sizeof(DWORD) );
2503         FIXME(") calling address was 0x%08x\n",calladdr);
2504         return 0;
2505 }