win32.c

Go to the documentation of this file.

/****************************************************************************
*                                                                           *
*                         Win32 Randomness-Gathering Code                   *
*   Copyright Peter Gutmann, Matt Thomlinson and Blake Coverett 1996-2010   *
*                                                                           *
****************************************************************************/

/* This module is part of the cryptlib continuously seeded pseudorandom number
   generator.  For usage conditions, see random.c */

/* General includes */

#include "crypt.h"
#include "random/random.h"

/* OS-specific includes */

#include <tlhelp32.h>
#include <winperf.h>
#include <winioctl.h>
#include <process.h>

/* Some new CPU opcodes aren't supported by all compiler versions, if
   they're not available we define them here.  BC++ can only handle the
   emit directive outside an asm block, so we have to terminate the current
   block, emit the opcode, and then restart the asm block.  In addition
   while the 16-bit versions of BC++ had built-in asm support, the 32-bit
   versions removed it and generate a temporary asm file which is passed
   to Tasm32.  This is only included with some high-end versions of BC++
   and it's not possible to order it separately, so if you're building with
   BC++ and get error messages about a missing tasm32.exe, add NO_ASM to
   Project Options | Compiler | Defines */

#if defined( _MSC_VER )
  #if _MSC_VER <= 1100
    #define cpuid       __asm _emit 0x0F __asm _emit 0xA2
    #define rdtsc       __asm _emit 0x0F __asm _emit 0x31
  #endif /* VC++ 5.0 or earlier */
  #define xstore_rng    __asm _emit 0x0F __asm _emit 0xA7 __asm _emit 0xC0
  #define rdrand_eax    __asm _emit 0x0F __asm _emit 0xC7 __asm _emit 0xF0
#endif /* VC++ */
#if defined __BORLANDC__
  #define cpuid         } __emit__( 0x0F, 0xA2 ); __asm {
  #define rdtsc         } __emit__( 0x0F, 0x31 ); __asm {
  #define xstore_rng    } __emit__( 0x0F, 0xA7, 0xC0 ); __asm {
  #define rdrand_eax    } __emit__( 0x0F, 0xC7, 0xF0 ); __asm {
#endif /* BC++ */

/* Map a value that may be 32 or 64 bits depending on the platform to a 
   long */

#if defined( _MSC_VER ) && ( _MSC_VER >= 1400 )
  #define addRandomHandle( randomState, handle ) \
          addRandomLong( randomState, PtrToUlong( handle ) )
#else
  #define addRandomHandle   addRandomValue
#endif /* 32- vs. 64-bit VC++ */

/* The size of the intermediate buffer used to accumulate polled data */

#define RANDOM_BUFSIZE  4096
#if RANDOM_BUFSIZE > MAX_INTLENGTH_SHORT
  #error RANDOM_BUFSIZE exceeds randomness accumulator size
#endif /* RANDOM_BUFSIZE > MAX_INTLENGTH_SHORT */

/* Handles to various randomness objects */

static HANDLE hAdvAPI32;    /* Handle to misc.library */
static HANDLE hNetAPI32;    /* Handle to networking library */
static HANDLE hNTAPI;       /* Handle to NT kernel library */
static HANDLE hThread;      /* Background polling thread handle */
static unsigned int threadID;   /* Background polling thread ID */

/****************************************************************************
*                                                                           *
*                           System RNG Interface                            *
*                                                                           *
****************************************************************************/

/* The number of bytes to read from the system RNG on each slow poll */

#define SYSTEMRNG_BYTES     64

/* Intel Chipset CSP type and name */

#define PROV_INTEL_SEC  22
#define INTEL_DEF_PROV  "Intel Hardware Cryptographic Service Provider"

/* A mapping from CryptoAPI to standard data types */

#define HCRYPTPROV          HANDLE

/* Type definitions for function pointers to call CryptoAPI functions */

typedef BOOL ( WINAPI *CRYPTACQUIRECONTEXT )( HCRYPTPROV *phProv,
                                              LPCTSTR pszContainer,
                                              LPCTSTR pszProvider, DWORD dwProvType,
                                              DWORD dwFlags );
typedef BOOL ( WINAPI *CRYPTGENRANDOM )( HCRYPTPROV hProv, DWORD dwLen,
                                         BYTE *pbBuffer );
typedef BOOL ( WINAPI *CRYPTRELEASECONTEXT )( HCRYPTPROV hProv, DWORD dwFlags );

/* Somewhat alternative functionality available as a direct call, for 
   Windows XP and newer.  This is the CryptoAPI RNG, which isn't anywhere
   near as good as the HW RNG, but we use it if it's present on the basis
   that at least it can't make things any worse.  This direct access version 
   is only available under Windows XP and newer, we don't go out of our way 
   to access the more general CryptoAPI one since the main purpose of using 
   it is to take advantage of any possible future hardware RNGs that may be 
   added, for example via TCPA devices */

typedef BOOL ( WINAPI *RTLGENRANDOM )( PVOID RandomBuffer, 
                                       ULONG RandomBufferLength );

/* Global function pointers. These are necessary because the functions need
   to be dynamically linked since older versions of Win95 and NT don't contain
   them */

static CRYPTACQUIRECONTEXT pCryptAcquireContext = NULL;
static CRYPTGENRANDOM pCryptGenRandom = NULL;
static CRYPTRELEASECONTEXT pCryptReleaseContext = NULL;
static RTLGENRANDOM pRtlGenRandom = NULL;

/* Handle to the RNG CSP */

static BOOLEAN systemRngAvailable;  /* Whether system RNG is available */
static HCRYPTPROV hProv;            /* Handle to Intel RNG CSP */

/* Try and connect to the system RNG if there's one present.  In theory we 
   could also try and get data from a TPM if there's one present, but the 
   TPM functions are only available under Vista (so they'd have to be 
   dynamically bound), the chances of a TPM being present and accessible 
   are pretty slim, and since we have to talk to the TPM at the raw APDU 
   level (not to mention all the horror stories in the MSDN forums about 
   getting any of this stuff to work) the amount of effort required versus 
   the potential gain really doesn't make it worthwhile:

    #include <tbs.h>

    TBS_HCONTEXT hTbsContext;
    TBS_CONTEXT_PARAMS tbsContextParams;
    HRESULT hResult;

    // Include dynamic-linking management for all functions

    // Get a handle to the TBS interface
    memset( &tspContextParams, 0, sizeof( TBS_CONTEXT_PARAMS ) );
    tspContextParams.version = TBS_CONTEXT_VERSION_ONE;
    result = Tbsi_Context_Create( &tbsContextParams, &hTbsContext );
    if( SUCCEEDED( hResult ) )
        {
        BYTE tpmCommand[] = {
            0x00, 0xC1,                 // WORD: TPM_TAG_RQU_COMMAND
            0x00, 0x00, 0x00, 0x0E,     // LONG: Total packet length
            0x00, 0x00, 0x00, 0x46,     // LONG: Command TPM_ORD_GetRandom
            0x00, 0x00, 0x00, 0x20      // Data: Get 0x20 bytes
        BYTE buffer[ 14 + 32 + 8 ];
        int length;

        hResult = Tbsip_Submit_Command( hTbsContext, 
                                        TBS_COMMAND_LOCALITY_ZERO,
                                        TBS_COMMAND_PRIORITY_NORMAL, 
                                        tpmCommand, 14, buffer, &length );
        if( SUCCEEDED( hResult ) )
            {
            // Start of buffer contains the 14 bytes of returned command 
            // followed by 32 bytes of random data
            }
        Tbsip_Context_Close( hTbsContext );
        } */

static void initSystemRNG( void )
    {
    systemRngAvailable = FALSE;
    hProv = NULL;
    if( ( hAdvAPI32 = GetModuleHandle( "AdvAPI32.dll" ) ) == NULL )
        return;

    /* Get pointers to the CSP functions.  Although the acquire context
       function looks like a standard function, it's actually a macro which
       is mapped to (depending on the build type) CryptAcquireContextA or
       CryptAcquireContextW, so we access it under the straight-ASCII-
       function name */
    pCryptAcquireContext = ( CRYPTACQUIRECONTEXT ) GetProcAddress( hAdvAPI32,
                                                    "CryptAcquireContextA" );
    pCryptGenRandom = ( CRYPTGENRANDOM ) GetProcAddress( hAdvAPI32,
                                                    "CryptGenRandom" );
    pCryptReleaseContext = ( CRYPTRELEASECONTEXT ) GetProcAddress( hAdvAPI32,
                                                    "CryptReleaseContext" );

    /* Get a pointer to the native randomness function if it's available.  
       This isn't exported by name, so we have to get it by ordinal */
    pRtlGenRandom = ( RTLGENRANDOM ) GetProcAddress( hAdvAPI32,
                                                    "SystemFunction036" );

    /* Try and connect to the PIII RNG CSP.  The AMD 768 southbridge (from 
       the 760 MP chipset) also has a hardware RNG, but there doesn't appear 
       to be any driver support for this as there is for the Intel RNG so we 
       can't do much with it.  OTOH the Intel RNG is also effectively dead 
       as well, mostly due to virtually nonexistant support/marketing by 
       Intel, it's included here mostly for form's sake */
    if( ( pCryptAcquireContext == NULL || \
          pCryptGenRandom == NULL || pCryptReleaseContext == NULL || \
          pCryptAcquireContext( &hProv, NULL, INTEL_DEF_PROV,
                                PROV_INTEL_SEC, 0 ) == FALSE ) && \
        ( pRtlGenRandom == NULL ) )
        {
        hAdvAPI32 = NULL;
        hProv = NULL;
        }
    else
        {
        /* Remember that we have a system RNG available for use */
        systemRngAvailable = TRUE;
        }
    }

/* Read data from the system RNG, in theory the PIII hardware RNG but in
   practice more likely the CryptoAPI software RNG */

static void readSystemRNG( void )
    {
    BYTE buffer[ SYSTEMRNG_BYTES + 8 ];
    int quality = 0;

    if( !systemRngAvailable )
        return;

    /* Read SYSTEMRNG_BYTES bytes from the system RNG.  Rather presciently, 
       the code up until late 2007 stated that "We don't rely on this for 
       all our randomness requirements (particularly the software RNG) in 
       case it's broken in some way", and in November 2007 an analysis paper
       by Leo Dorrendorf, Zvi Gutterman, and Benny Pinkas showed that the
       CryptoAPI RNG is indeed broken, being neither forwards- nor 
       backwards-secure, being reseeded far too infrequently, and (as far as
       their reverse-engineering was able to tell) using far less entropy
       sources than cryptlib's built-in mechanisms even though the CryptoAPI
       one is deeply embedded in the OS.

       The way the CryptoAPI RNG works is that a system RNG is used to key 
       a set of eight RC4 ciphers, each of which is used in turn in a round-
       robin fashion to output 20 bytes of randomness which are then hashed 
       with SHA1, with the operation being approximately:

        output[  0...19 ] = SHA1( RC4[ i++ % 8 ] );
        output[ 20...39 ] = SHA1( RC4[ i++ % 8 ] );
        [...]

       Each RC4 cipher is re-keyed every 16KB of output, so for 8 ciphers
       the rekey interval is every 128KB of output.  Furthermore, although
       the kernel RNG used to key the RC4 ciphers stores its state in-
       kernel, the RC4 cipher state is stored in user-space and per-process.
       This means that in most cases the RNG is never re-keyed.  Finally,
       the way the RNG works means that it's possible to recover earlier
       state in O( 2^23 ).  See "Cryptanalysis of the Random Number 
       Generator of the Windows Operating System" for details */
    if( hProv != NULL )
        {
        if( pCryptGenRandom( hProv, SYSTEMRNG_BYTES, buffer ) )
            quality = 70;
        }
    else
        {
        if( pRtlGenRandom( buffer, SYSTEMRNG_BYTES ) )
            quality = 30;
        }
    if( quality > 0 )
        {
        MESSAGE_DATA msgData;
        int status;

        setMessageData( &msgData, buffer, SYSTEMRNG_BYTES );
        status = krnlSendMessage( SYSTEM_OBJECT_HANDLE, 
                                  IMESSAGE_SETATTRIBUTE_S, &msgData, 
                                  CRYPT_IATTRIBUTE_ENTROPY );
        if( cryptStatusOK( status ) )
            {
            ( void ) krnlSendMessage( SYSTEM_OBJECT_HANDLE, 
                                      IMESSAGE_SETATTRIBUTE,
                                      ( MESSAGE_CAST ) &quality,
                                      CRYPT_IATTRIBUTE_ENTROPY_QUALITY );
            }
        zeroise( buffer, SYSTEMRNG_BYTES );
        }
    }

/* Clean up the system RNG access */

static void endSystemRNG( void )
    {
    if( hProv != NULL )
        {
        pCryptReleaseContext( hProv, 0 );
        hProv = NULL;
        }
    }

/****************************************************************************
*                                                                           *
*                           Hardware Monitoring Interface                   *
*                                                                           *
****************************************************************************/

/* These interfaces currently support data supplied by MBM, Everest, 
   SysTool, RivaTuner, HMonitor, ATI Tray Tools, CoreTemp, and GPU-Z.  Two 
   notable omissions are SVPro and HWMonitor, unfortunately the authors 
   haven't responded to any requests for interfacing information */

/* MBM data structures, originally by Alexander van Kaam, converted to C by
   [email protected], finally updated by Chris Zahrt <[email protected]>.
   The __int64 values below are actually (64-bit) doubles, but we overlay 
   them with __int64s since we don't care about the values */

#define BusType     char
#define SMBType     char
#define SensorType  char

typedef struct {
    SensorType iType;           /* Type of sensor */
    int Count;                  /* Number of sensor for that type */
    } SharedIndex;

typedef struct {
    SensorType ssType;          /* Type of sensor */
    unsigned char ssName[ 12 ]; /* Name of sensor */
    char sspadding1[ 3 ];       /* Padding of 3 bytes */
    __int64 /*double*/ ssCurrent;/* Current value */
    __int64 /*double*/ ssLow;   /* Lowest readout */
    __int64 /*double*/ ssHigh;  /* Highest readout */
    long ssCount;               /* Total number of readout */
    char sspadding2[ 4 ];       /* Padding of 4 bytes */
    BYTE /*long double*/ ssTotal[ 8 ];  /* Total amout of all readouts */
    char sspadding3[ 6 ];       /* Padding of 6 bytes */
    __int64 /*double*/ ssAlarm1;/* Temp & fan: high alarm; voltage: % off */
    __int64 /*double*/ ssAlarm2;/* Temp: low alarm */
    } SharedSensor;

typedef struct {
    short siSMB_Base;           /* SMBus base address */
    BusType siSMB_Type;         /* SMBus/Isa bus used to access chip */
    SMBType siSMB_Code;         /* SMBus sub type, Intel, AMD or ALi */
    char siSMB_Addr;            /* Address of sensor chip on SMBus */
    unsigned char siSMB_Name[ 41 ]; /* Nice name for SMBus */
    short siISA_Base;           /* ISA base address of sensor chip on ISA */
    int siChipType;             /* Chip nr, connects with Chipinfo.ini */
    char siVoltageSubType;      /* Subvoltage option selected */
    } SharedInfo;

typedef struct {
    __int64 /*double*/ sdVersion;/* Version number (example: 51090) */
    SharedIndex sdIndex[ 10 ];  /* Sensor index */
    SharedSensor sdSensor[ 100 ];   /* Sensor info */
    SharedInfo sdInfo;          /* Misc.info */
    unsigned char sdStart[ 41 ];    /* Start time */
    /* We don't use the next two fields both because they're not random and
       because it provides a nice safety margin in case of data size mis-
       estimates (we always under-estimate the buffer size) */
/*  unsigned char sdCurrent[ 41 ];  /* Current time */
/*  unsigned char sdPath[ 256 ];    /* MBM path */
    } SharedData;

/* Read data from MBM via the shared-memory interface */

static void readMBMData( void )
    {
    HANDLE hMBMData;
    const SharedData *mbmDataPtr;

    if( ( hMBMData = OpenFileMapping( FILE_MAP_READ, FALSE,
                                      "$M$B$M$5$S$D$" ) ) == NULL )
        return;
    if( ( mbmDataPtr = ( SharedData * ) \
            MapViewOfFile( hMBMData, FILE_MAP_READ, 0, 0, 0 ) ) != NULL )
        {
        MESSAGE_DATA msgData;
        static const int quality = 20;

        setMessageData( &msgData, ( MESSAGE_CAST ) mbmDataPtr, 
                        sizeof( SharedData ) );
        krnlSendMessage( SYSTEM_OBJECT_HANDLE, IMESSAGE_SETATTRIBUTE_S,
                         &msgData, CRYPT_IATTRIBUTE_ENTROPY );
        krnlSendMessage( SYSTEM_OBJECT_HANDLE, IMESSAGE_SETATTRIBUTE,
                         ( MESSAGE_CAST ) &quality,
                         CRYPT_IATTRIBUTE_ENTROPY_QUALITY );
        UnmapViewOfFile( mbmDataPtr );
        }
    CloseHandle( hMBMData );
    }

/* Read data from Everest via the shared-memory interface.  Everest returns 
   information as an enormous XML text string so we have to be careful about 
   handling of lengths.  In general the returned length is 1-3K, so we 
   hard-limit it at 2K to ensure there are no problems if the trailing null 
   gets lost */

static void readEverestData( void )
    {
    HANDLE hEverestData;
    const void *everestDataPtr;

    if( ( hEverestData = OpenFileMapping( FILE_MAP_READ, FALSE,
                                          "EVEREST_SensorValues" ) ) == NULL )
        return;
    if( ( everestDataPtr = MapViewOfFile( hEverestData, 
                                          FILE_MAP_READ, 0, 0, 0 ) ) != NULL )
        {
        const int length = strlen( everestDataPtr );

        if( length > 128 )
            {
            MESSAGE_DATA msgData;
            static const int quality = 40;

            setMessageData( &msgData, ( MESSAGE_CAST ) everestDataPtr, 
                            min( length, 2048 ) );
            krnlSendMessage( SYSTEM_OBJECT_HANDLE, IMESSAGE_SETATTRIBUTE_S,
                             &msgData, CRYPT_IATTRIBUTE_ENTROPY );
            krnlSendMessage( SYSTEM_OBJECT_HANDLE, IMESSAGE_SETATTRIBUTE,
                             ( MESSAGE_CAST ) &quality,
                             CRYPT_IATTRIBUTE_ENTROPY_QUALITY );
            }
        UnmapViewOfFile( everestDataPtr );
        }
    CloseHandle( hEverestData );
    }

/* SysTool data structures, from forums.techpowerup.com.  This uses the
   standard shared memory interface, but it gets a bit tricky because it 
   uses the OLE 'VARIANT' data type which we can't access because it's 
   included via a complex nested inclusion framework in windows.h, and the 
   use of '#pragma once' when we included it for use in crypt.h means that 
   we can't re-include any of the necessary files.  Since we don't actually 
   care what's inside a VARIANT, and it has a fixed size of 16 bytes, we 
   just use it as an opaque blob */

typedef BYTE VARIANT[ 16 ]; /* Kludge for OLE data type */
typedef BYTE UINT8;
typedef WORD UINT16;

#define SH_MEM_MAX_SENSORS 128

typedef enum { sUnknown, sNumber, sTemperature, sVoltage, sRPM, sBytes, 
               sBytesPerSecond, sMhz, sPercentage, sString, sPWM 
             } SYSTOOL_SENSOR_TYPE;
 
typedef struct {
    WCHAR m_name[ 255 ];    /* Sensor name */
    WCHAR m_section[ 64 ];  /* Section in which this sensor appears */
    SYSTOOL_SENSOR_TYPE m_sensorType;
    LONG m_updateInProgress;/* Nonzero when sensor is being updated */
    UINT32 m_timestamp;     /* GetTickCount() of last update */
    VARIANT m_value;        /* Sensor data */
    WCHAR m_unit[ 8 ];      /* Unit for text output */
    UINT8 m_nDecimals;      /* Default number of decimals for formatted output */
    } SYSTOOL_SHMEM_SENSOR;
 
typedef struct {
    UINT32 m_version;       /* Version of shared memory structure */
    UINT16 m_nSensors;      /* Number of records with data in m_sensors */
    SYSTOOL_SHMEM_SENSOR m_sensors[ SH_MEM_MAX_SENSORS ];
    } SYSTOOL_SHMEM;

static void readSysToolData( void )
    {
    HANDLE hSysToolData;
    const SYSTOOL_SHMEM *sysToolDataPtr;

    if( ( hSysToolData = OpenFileMapping( FILE_MAP_READ, FALSE,
                                          "SysToolSensors" ) ) == NULL )
        return;
    if( ( sysToolDataPtr = ( SYSTOOL_SHMEM * ) \
            MapViewOfFile( hSysToolData, FILE_MAP_READ, 0, 0, 0 ) ) != NULL )
        {
        MESSAGE_DATA msgData;
        static const int quality = 40;

        setMessageData( &msgData, ( MESSAGE_CAST ) sysToolDataPtr, 
                        sizeof( SYSTOOL_SHMEM ) );
        krnlSendMessage( SYSTEM_OBJECT_HANDLE, IMESSAGE_SETATTRIBUTE_S,
                         &msgData, CRYPT_IATTRIBUTE_ENTROPY );
        krnlSendMessage( SYSTEM_OBJECT_HANDLE, IMESSAGE_SETATTRIBUTE,
                         ( MESSAGE_CAST ) &quality,
                         CRYPT_IATTRIBUTE_ENTROPY_QUALITY );
        UnmapViewOfFile( sysToolDataPtr );
        }
    CloseHandle( hSysToolData );
    }

/* RivaTuner data structures via the shared-memory interface, from the 
   RivaTuner sample source.  The DWORD values below are actually (32-bit) 
   floats, but we overlay them with DWORDs since we don't care about the 
   values.  The information is only accessible when the RivaTuner hardware 
   monitoring window is open.  This one is the easiest of the shared-mem 
   interfaces to monitor because for fowards-compatibility reasons it 
   stores a Windows-style indicator of the size of each struct entry in the 
   data header, so we can get the total size simply by multiplying out the 
   number of entries by the indicated size.  As a safety measure we limit 
   the maximum data size to 2K in case a value gets corrupted */

typedef struct {
    DWORD dwSignature;  /* 'RTHM' if active */
    DWORD dwVersion;    /* Must be 0x10001 or above */
    DWORD dwNumEntries; /* No.of RTHM_SHARED_MEMORY_ENTRY entries */
    time_t time;        /* Last polling time */
    DWORD dwEntrySize;  /* Size of entries in RTHM_SHARED_MEMORY_ENTRY array */
    } RTHM_SHARED_MEMORY_HEADER;

typedef struct {
    char czSrc[ 32 ];   /* Source description */
    char czDim[ 16 ];   /* Source measurement units */
    DWORD /*float*/ data;   /* Source data */
    DWORD /*float*/ offset; /* Source offset, e.g. temp.compensation */
    DWORD /*float*/ dataTransformed;/* Source data in transformed(?) form */
    DWORD flags;        /* Misc.flags */
    } RTHM_SHARED_MEMORY_ENTRY;

static void readRivaTunerData( void )
    {
    HANDLE hRivaTunerData;
    RTHM_SHARED_MEMORY_HEADER *rivaTunerHeaderPtr;

    if( ( hRivaTunerData = OpenFileMapping( FILE_MAP_READ, FALSE,
                                            "RTHMSharedMemory" ) ) == NULL )
        return;
    if( ( rivaTunerHeaderPtr = ( RTHM_SHARED_MEMORY_HEADER * ) \
            MapViewOfFile( hRivaTunerData, FILE_MAP_READ, 0, 0, 0 ) ) != NULL )
        {
        if( rivaTunerHeaderPtr->dwSignature == 'RTHM' && \
            rivaTunerHeaderPtr->dwVersion >= 0x10001 )
            {
            MESSAGE_DATA msgData;
            const BYTE *entryPtr = ( ( BYTE * ) rivaTunerHeaderPtr ) + \
                                  sizeof( RTHM_SHARED_MEMORY_HEADER );
            const int entryTotalSize = rivaTunerHeaderPtr->dwNumEntries * \
                                       rivaTunerHeaderPtr->dwEntrySize;
            static const int quality = 10;

            setMessageData( &msgData, ( MESSAGE_CAST ) entryPtr, 
                            min( entryTotalSize, 2048 ) );
            krnlSendMessage( SYSTEM_OBJECT_HANDLE, IMESSAGE_SETATTRIBUTE_S,
                             &msgData, CRYPT_IATTRIBUTE_ENTROPY );
            krnlSendMessage( SYSTEM_OBJECT_HANDLE, IMESSAGE_SETATTRIBUTE,
                             ( MESSAGE_CAST ) &quality,
                             CRYPT_IATTRIBUTE_ENTROPY_QUALITY );
            }
        UnmapViewOfFile( rivaTunerHeaderPtr );
        }
    CloseHandle( hRivaTunerData );
    }

/* Read data from HMonitor via the shared-memory interface.  The DWORD 
   values below are actually (32-bit) floats, but we overlay them with 
   DWORDs since we don't care about the values.  Like RivaTuner's data the 
   info structure contains a length value at the start, so it's fairly easy 
   to work with */

typedef struct {
    WORD length;
    WORD version;   /* Single 'DWORD length' before version 4.1 */
    DWORD /*float*/ temp[ 3 ];
    DWORD /*float*/ voltage[ 7 ];
    int fan[ 3 ];
    } HMONITOR_DATA;

static void readHMonitorData( void )
    {
    HANDLE hHMonitorData;
    HMONITOR_DATA *hMonitorDataPtr;

    if( ( hHMonitorData = OpenFileMapping( FILE_MAP_READ, FALSE,
                                           "Hmonitor_Counters_Block" ) ) == NULL )
        return;
    if( ( hMonitorDataPtr = ( HMONITOR_DATA * ) \
            MapViewOfFile( hHMonitorData, FILE_MAP_READ, 0, 0, 0 ) ) != NULL )
        {
        if( hMonitorDataPtr->version >= 0x4100 && \
            ( hMonitorDataPtr->length >= 48 && hMonitorDataPtr->length <= 1024 ) )
            {
            MESSAGE_DATA msgData;
            static const int quality = 40;

            setMessageData( &msgData, hMonitorDataPtr, hMonitorDataPtr->length );
            krnlSendMessage( SYSTEM_OBJECT_HANDLE, IMESSAGE_SETATTRIBUTE_S,
                             &msgData, CRYPT_IATTRIBUTE_ENTROPY );
            krnlSendMessage( SYSTEM_OBJECT_HANDLE, IMESSAGE_SETATTRIBUTE,
                             ( MESSAGE_CAST ) &quality, 
                             CRYPT_IATTRIBUTE_ENTROPY_QUALITY );
            }
        UnmapViewOfFile( hMonitorDataPtr );
        }
    CloseHandle( hHMonitorData );
    }

/* Read data from ATI Tray Tools via the shared-memory interface.  This has
   an added twist in that just because it's available to be read doesn't 
   mean that it contains any data, so we check that at least the GPU speed 
   and temp-monitoring-supported flag have nonzero values */

typedef struct {
    DWORD CurGPU;       /* GPU speed */
    DWORD CurMEM;       /* Video memory speed */
    DWORD isGameActive;
    DWORD is3DActive;   /* Boolean: 3D mode active */
    DWORD isTempMonSupported;   /* Boolean: Temp monitoring available */
    DWORD GPUTemp;      /* GPU temp */
    DWORD ENVTemp;      /* Video card temp */
    DWORD FanDuty;      /* Fan duty cycle */
    DWORD MAXGpuTemp, MINGpuTemp;   /* Min/max GPU temp */
    DWORD MAXEnvTemp, MINEnvTemp;   /* Min/max video card temp */
    DWORD CurD3DAA, CurD3DAF;   /* Direct3D info */
    DWORD CurOGLAA, CurOGLAF;   /* OpenGL info */
    DWORD IsActive;     /* Another 3D boolean */
    DWORD CurFPS;       /* FPS rate */
    DWORD FreeVideo;    /* Available video memory */
    DWORD FreeTexture;  /* Available texture memory */
    DWORD Cur3DApi;     /* API used (D3D, OpenGL, etc) */
    DWORD MemUsed;      /* ? */
    } TRAY_TOOLS_DATA;

static void readATITrayToolsData( void )
    {
    HANDLE hTrayToolsData;
    TRAY_TOOLS_DATA *trayToolsDataPtr;

    if( ( hTrayToolsData = OpenFileMapping( FILE_MAP_READ, FALSE,
                                            "ATITRAY_SMEM" ) ) == NULL )
        return;
    if( ( trayToolsDataPtr = ( TRAY_TOOLS_DATA * ) \
            MapViewOfFile( hTrayToolsData, FILE_MAP_READ, 0, 0, 0 ) ) != NULL )
        {
        if( trayToolsDataPtr->CurGPU >= 100 && \
            trayToolsDataPtr->isTempMonSupported != 0 )
            {
            MESSAGE_DATA msgData;
            static const int quality = 10;

            setMessageData( &msgData, trayToolsDataPtr,
                            sizeof( TRAY_TOOLS_DATA ) );
            krnlSendMessage( SYSTEM_OBJECT_HANDLE, IMESSAGE_SETATTRIBUTE_S,
                             &msgData, CRYPT_IATTRIBUTE_ENTROPY );
            krnlSendMessage( SYSTEM_OBJECT_HANDLE, IMESSAGE_SETATTRIBUTE,
                             ( MESSAGE_CAST ) &quality,
                             CRYPT_IATTRIBUTE_ENTROPY_QUALITY );
            }
        UnmapViewOfFile( trayToolsDataPtr );
        }
    CloseHandle( hTrayToolsData );
    }

/* CoreTemp data structure, from www.alcpu.com/CoreTemp.   The DWORD values 
   below are actually (32-bit) floats but we overlay them with DWORDs since 
   we don't care about the values */

typedef struct 
    {
    unsigned int uiLoad[ 256 ];
    unsigned int uiTjMax[ 128 ];
    unsigned int uiCoreCnt;
    unsigned int uiCPUCnt;
    DWORD /*float*/ fTemp[ 256 ];
    DWORD /*float*/ fVID;
    DWORD /*float*/ fCPUSpeed;
    DWORD /*float*/ fFSBSpeed;
    DWORD /*float*/ fMultipier; 
    DWORD /*float*/ sCPUName[ 100 ];
    unsigned char ucFahrenheit;
    unsigned char ucDeltaToTjMax;
    } CORE_TEMP_SHARED_DATA;

static void readCoreTempData( void )
    {
    HANDLE hCoreTempData;
    CORE_TEMP_SHARED_DATA *coreTempDataPtr;

    if( ( hCoreTempData = OpenFileMapping( FILE_MAP_READ, FALSE,
                                           "CoreTempMappingObject" ) ) == NULL )
        return;
    if( ( coreTempDataPtr = ( CORE_TEMP_SHARED_DATA * ) \
            MapViewOfFile( hCoreTempData, FILE_MAP_READ, 0, 0, 0 ) ) != NULL )
        {
        MESSAGE_DATA msgData;
        static const int quality = 15;

        setMessageData( &msgData, coreTempDataPtr,
                        sizeof( CORE_TEMP_SHARED_DATA ) );
        krnlSendMessage( SYSTEM_OBJECT_HANDLE, IMESSAGE_SETATTRIBUTE_S,
                         &msgData, CRYPT_IATTRIBUTE_ENTROPY );
        krnlSendMessage( SYSTEM_OBJECT_HANDLE, IMESSAGE_SETATTRIBUTE,
                         ( MESSAGE_CAST ) &quality,
                         CRYPT_IATTRIBUTE_ENTROPY_QUALITY );
        UnmapViewOfFile( coreTempDataPtr );
        }
    CloseHandle( hCoreTempData );
    }

/* GPU-Z data structures, from forums.techpowerup.com.  This uses an 
   incredibly wasteful memory layout so we end up having to submit around 
   200kB of data, unfortunately we can't easily pick out the few values of 
   interest because the 'data' entries push out the values of interest, the 
   'sensors' entries, to the end of the memory block */

#define MAX_RECORDS     128

#pragma pack(push, 1)

typedef struct {
    WCHAR key[ 256 ];
    WCHAR value[ 256 ];
    } GPUZ_RECORD;

typedef struct {
    WCHAR name[ 256 ];
    WCHAR unit[ 8 ];
    UINT32 digits;
    double value;
    } GPUZ_SENSOR_RECORD;

typedef struct {
    UINT32 version;     /* Version number, should be 1 */
    volatile LONG busy; /* Data-update flag */
    UINT32 lastUpdate;  /* GetTickCount() of last update */
    GPUZ_RECORD data[ MAX_RECORDS ];
    GPUZ_SENSOR_RECORD sensors[ MAX_RECORDS ];
    } GPUZ_SH_MEM;

#pragma pack(pop)

static void readGPUZData( void )
    {
    HANDLE hGPUZData;
    const GPUZ_SH_MEM *gpuzDataPtr;

    if( ( hGPUZData = OpenFileMapping( FILE_MAP_READ, FALSE,
                                       "GPUZShMem" ) ) == NULL )
        return;
    if( ( gpuzDataPtr = ( GPUZ_SH_MEM * ) \
          MapViewOfFile( hGPUZData, FILE_MAP_READ, 0, 0, 0 ) ) != NULL )
        {
        if( gpuzDataPtr->version == 1 )
            {
            MESSAGE_DATA msgData;
            static const int quality = 10;

            setMessageData( &msgData, ( MESSAGE_CAST ) gpuzDataPtr, 
                            sizeof( GPUZ_SH_MEM ) );
            krnlSendMessage( SYSTEM_OBJECT_HANDLE, IMESSAGE_SETATTRIBUTE_S,
                             &msgData, CRYPT_IATTRIBUTE_ENTROPY );
            krnlSendMessage( SYSTEM_OBJECT_HANDLE, IMESSAGE_SETATTRIBUTE,
                             ( MESSAGE_CAST ) &quality,
                             CRYPT_IATTRIBUTE_ENTROPY_QUALITY );
            }
        UnmapViewOfFile( gpuzDataPtr );
        }
    CloseHandle( hGPUZData );
    }

/****************************************************************************
*                                                                           *
*                           Hardware Configuration Data                     *
*                                                                           *
****************************************************************************/

/* Read PnP configuration data.  This is mostly static per machine, but
   differs somewhat across machines.  We have to define the values ourselves
   here due to a combination of some of the values and functions not
   existing at the time VC++ 6.0 was released */

typedef void * HDEVINFO;

#define DIGCF_PRESENT       0x02
#define DIGCF_ALLCLASSES    0x04

#define SPDRP_HARDWAREID    0x01

typedef struct _SP_DEVINFO_DATA {
    DWORD cbSize;
    GUID classGuid;
    DWORD devInst;
    ULONG *reserved;
    } SP_DEVINFO_DATA, *PSP_DEVINFO_DATA;

typedef BOOL ( WINAPI *SETUPDIDESTROYDEVICEINFOLIST )( HDEVINFO DeviceInfoSet );
typedef BOOL ( WINAPI *SETUPDIENUMDEVICEINFO )( HDEVINFO DeviceInfoSet,
                                                DWORD MemberIndex,
                                                PSP_DEVINFO_DATA DeviceInfoData );
typedef HDEVINFO ( WINAPI *SETUPDIGETCLASSDEVS )( /*CONST LPGUID*/ void *ClassGuid,
                                                  /*PCTSTR*/ void *Enumerator,
                                                  HWND hwndParent, DWORD Flags );
typedef BOOL ( WINAPI *SETUPDIGETDEVICEREGISTRYPROPERTY )( HDEVINFO DeviceInfoSet,
                                                PSP_DEVINFO_DATA DeviceInfoData,
                                                DWORD Property, PDWORD PropertyRegDataType,
                                                PBYTE PropertyBuffer,
                                                DWORD PropertyBufferSize, PDWORD RequiredSize );

static void readPnPData( void )
    {
    HANDLE hSetupAPI;
    HDEVINFO hDevInfo;
    SETUPDIDESTROYDEVICEINFOLIST pSetupDiDestroyDeviceInfoList = NULL;
    SETUPDIENUMDEVICEINFO pSetupDiEnumDeviceInfo = NULL;
    SETUPDIGETCLASSDEVS pSetupDiGetClassDevs = NULL;
    SETUPDIGETDEVICEREGISTRYPROPERTY pSetupDiGetDeviceRegistryProperty = NULL;

    if( ( hSetupAPI = DynamicLoad( "SetupAPI.dll" ) ) == NULL )
        return;

    /* Get pointers to the PnP functions.  Although the get class-devs
       and get device registry functions look like standard functions,
       they're actually macros that are mapped to (depending on the build
       type) xxxA or xxxW, so we access them under the straight-ASCII-
       function name */
    pSetupDiDestroyDeviceInfoList = ( SETUPDIDESTROYDEVICEINFOLIST ) \
                GetProcAddress( hSetupAPI, "SetupDiDestroyDeviceInfoList" );
    pSetupDiEnumDeviceInfo = ( SETUPDIENUMDEVICEINFO ) \
                GetProcAddress( hSetupAPI, "SetupDiEnumDeviceInfo" );
    pSetupDiGetClassDevs = ( SETUPDIGETCLASSDEVS ) \
                GetProcAddress( hSetupAPI, "SetupDiGetClassDevsA" );
    pSetupDiGetDeviceRegistryProperty = ( SETUPDIGETDEVICEREGISTRYPROPERTY ) \
                GetProcAddress( hSetupAPI, "SetupDiGetDeviceRegistryPropertyA" );
    if( pSetupDiDestroyDeviceInfoList == NULL || \
        pSetupDiEnumDeviceInfo == NULL || pSetupDiGetClassDevs == NULL || \
        pSetupDiGetDeviceRegistryProperty == NULL )
        {
        DynamicUnload( hSetupAPI );
        return;
        }

    /* Get info on all PnP devices */
    hDevInfo = pSetupDiGetClassDevs( NULL, NULL, NULL,
                                     DIGCF_PRESENT | DIGCF_ALLCLASSES );
    if( hDevInfo != INVALID_HANDLE_VALUE )
        {
        SP_DEVINFO_DATA devInfoData;
        RANDOM_STATE randomState;
        BYTE buffer[ RANDOM_BUFSIZE + 8 ];
        BYTE pnpBuffer[ 512 + 8 ];
        DWORD cbPnPBuffer;
        int deviceCount, iterationCount, status;

        /* Enumerate all PnP devices */
        status = initRandomData( randomState, buffer, RANDOM_BUFSIZE );
        if( cryptStatusError( status ) )
            retIntError_Void();
        memset( &devInfoData, 0, sizeof( devInfoData ) );
        devInfoData.cbSize = sizeof( SP_DEVINFO_DATA );
        for( deviceCount = 0, iterationCount = 0;
             pSetupDiEnumDeviceInfo( hDevInfo, deviceCount, 
                                     &devInfoData ) && \
                iterationCount < FAILSAFE_ITERATIONS_MAX; 
             deviceCount++, iterationCount++ )
            {
            if( pSetupDiGetDeviceRegistryProperty( hDevInfo, &devInfoData,
                                                   SPDRP_HARDWAREID, NULL,
                                                   pnpBuffer, 512, &cbPnPBuffer ) )
                addRandomData( randomState, pnpBuffer, cbPnPBuffer );
            }
        ENSURES_V( iterationCount < FAILSAFE_ITERATIONS_MAX );
        pSetupDiDestroyDeviceInfoList( hDevInfo );
        endRandomData( randomState, 5 );
        }

    DynamicUnload( hSetupAPI );
    }

/****************************************************************************
*                                                                           *
*                                   Fast Poll                               *
*                                                                           *
****************************************************************************/

/* The shared Win32 fast poll routine */

void fastPoll( void )
    {
    static BOOLEAN addedFixedItems = FALSE;
    FILETIME  creationTime, exitTime, kernelTime, userTime;
    SIZE_T minimumWorkingSetSize, maximumWorkingSetSize;
    MEMORYSTATUS memoryStatus;
    HANDLE handle;
    POINT point;
    RANDOM_STATE randomState;
    BYTE buffer[ RANDOM_BUFSIZE + 8 ];
    int trngValue = 0, status;

    if( krnlIsExiting() )
        return;

    status = initRandomData( randomState, buffer, RANDOM_BUFSIZE );
    if( cryptStatusError( status ) )
        retIntError_Void();

    /* Get various basic pieces of system information: Handle of active
       window, handle of window with mouse capture, handle of clipboard owner
       handle of start of clpboard viewer list, pseudohandle of current
       process, current process ID, pseudohandle of current thread, current
       thread ID, handle of desktop window, handle  of window with keyboard
       focus, whether system queue has any events, cursor position for last
       message, 1 ms time for last message, handle of window with clipboard
       open, handle of process heap, handle of procs window station, types of
       events in input queue, and milliseconds since Windows was started.
       Since a HWND/HANDLE can be a 64-bit value on a 64-bit platform, we 
       have to use a mapping macro that discards the high 32 bits (which
       presumably won't be of much interest anyway).
       
       Note that there's another call that does almost the same thing as 
       GetTickCount(), timeGetTime() from the multimedia API, however this 
       has a fairly constant delta between ticks while GetTickCount() 
       wanders all over the place (while still averaging a fixed value).  So 
       timeGetTime() (which uses KeQueryInterruptTime()) isn't necessarily 
       more accurate than GetTickCount() (which uses KeQueryTickCount()), 
       just more regular.  Since we're interested in unpredictability, we 
       use GetTickCount() and not timeGetTime() */
    addRandomHandle( randomState, GetActiveWindow() );
    addRandomHandle( randomState, GetCapture() );
    addRandomHandle( randomState, GetClipboardOwner() );
    addRandomHandle( randomState, GetClipboardViewer() );
    addRandomHandle( randomState, GetCurrentProcess() );
    addRandomValue( randomState, GetCurrentProcessId() );
    addRandomHandle( randomState, GetCurrentThread() );
    addRandomValue( randomState, GetCurrentThreadId() );
    addRandomHandle( randomState, GetDesktopWindow() );
    addRandomHandle( randomState, GetFocus() );
    addRandomValue( randomState, GetInputState() );
    addRandomValue( randomState, GetMessagePos() );
    addRandomValue( randomState, GetMessageTime() );
    addRandomHandle( randomState, GetOpenClipboardWindow() );
    addRandomHandle( randomState, GetProcessHeap() );
    addRandomHandle( randomState, GetProcessWindowStation() );
    addRandomValue( randomState, GetTickCount() );
    if( krnlIsExiting() )
        return;

    /* Calling the following function can cause problems in some cases in
       that a calling application eventually stops getting events from its
       event loop, so we can't (safely) use it as an entropy source */
/*  addRandomValue( randomState, GetQueueStatus( QS_ALLEVENTS ) ); */

    /* Get multiword system information: Current caret position, current
       mouse cursor position */
    GetCaretPos( &point );
    addRandomData( randomState, &point, sizeof( POINT ) );
    GetCursorPos( &point );
    addRandomData( randomState, &point, sizeof( POINT ) );

    /* Get percent of memory in use, bytes of physical memory, bytes of free
       physical memory, bytes in paging file, free bytes in paging file, user
       bytes of address space, and free user bytes */
    memoryStatus.dwLength = sizeof( MEMORYSTATUS );
    GlobalMemoryStatus( &memoryStatus );
    addRandomData( randomState, &memoryStatus, sizeof( MEMORYSTATUS ) );

    /* Get thread and process creation time, exit time, time in kernel mode,
       and time in user mode in 100ns intervals */
    handle = GetCurrentThread();
    GetThreadTimes( handle, &creationTime, &exitTime, &kernelTime, &userTime );
    addRandomData( randomState, &creationTime, sizeof( FILETIME ) );
    addRandomData( randomState, &exitTime, sizeof( FILETIME ) );
    addRandomData( randomState, &kernelTime, sizeof( FILETIME ) );
    addRandomData( randomState, &userTime, sizeof( FILETIME ) );
    handle = GetCurrentProcess();
    GetProcessTimes( handle, &creationTime, &exitTime, &kernelTime, &userTime );
    addRandomData( randomState, &creationTime, sizeof( FILETIME ) );
    addRandomData( randomState, &exitTime, sizeof( FILETIME ) );
    addRandomData( randomState, &kernelTime, sizeof( FILETIME ) );
    addRandomData( randomState, &userTime, sizeof( FILETIME ) );

    /* Get the minimum and maximum working set size for the current process */
    GetProcessWorkingSetSize( handle, &minimumWorkingSetSize,
                              &maximumWorkingSetSize );
    addRandomValue( randomState, minimumWorkingSetSize );
    addRandomValue( randomState, maximumWorkingSetSize );

    /* The following are fixed for the lifetime of the process so we only
       add them once */
    if( !addedFixedItems )
        {
        STARTUPINFO startupInfo;

        /* Get name of desktop, console window title, new window position and
           size, window flags, and handles for stdin, stdout, and stderr */
        startupInfo.cb = sizeof( STARTUPINFO );
        GetStartupInfo( &startupInfo );
        addRandomData( randomState, &startupInfo, sizeof( STARTUPINFO ) );

        /* Remember that we've got the fixed info */
        addedFixedItems = TRUE;
        }

    /* The performance of QPC varies depending on the architecture it's
       running on and on the OS, the MS documentation is vague about the
       details because it varies so much.  Under Win9x/ME it reads the
       1.193180 MHz PIC timer.  Under NT/Win2K/XP it may or may not read the
       64-bit TSC depending on the HAL and assorted other circumstances,
       generally on machines with a uniprocessor HAL
       KeQueryPerformanceCounter() uses a 3.579545MHz timer and on machines
       with a multiprocessor or APIC HAL it uses the TSC (the exact time
       source is controlled by the HalpUse8254 flag in the kernel).  Since
       Vista-era machines are generally running on multiprocessor HALs we
       usually get the TSC under Vista.

       This choice of time sources is somewhat peculiar because on a
       multiprocessor machine it's theoretically possible to get completely
       different TSC readings depending on which CPU you're currently
       running on, while for uniprocessor machines it's not a problem.
       However the kernel synchronises the TSCs across CPUs at boot time if 
       a multiprocessor HAL is used (it resets the TSC as part of its system 
       init) so this shouldn't really be a problem.  Under WinCE it's 
       completely platform-dependant, if there's no hardware performance 
       counter available it uses the 1ms system timer.

       Another feature of the TSC (although it doesn't really affect us 
       here) is that mobile CPUs will turn off the TSC when they idle, 
       newer Pentiums and Athlons with advanced power management/clock
       throttling will change the rate of the counter when they clock-
       throttle (to match the current CPU speed), and hyperthreading 
       Pentiums will turn it off when both threads are idle (this more or 
       less makes sense since the CPU will be in the halted state and not 
       executing any instructions to count).  What makes this even more
       exciting is that the resulting handling of the TSC is almost entirely
       nondeterministic, the only thing that's guaranteed is that the
       counter is monotonically incrementing.  For example when a newer
       processor with power-saving features enabled ramps down its core
       clock the TSC rate is lowered to correspond to the clock rate
       (assuming that the BIOS sets this up correctly, which isn't always
       the case).  However various events like cache probes can cause the
       core to temporarily return to the original rate to process the
       event before dropping back to the throttled rate, which can cause
       TSC drift across multiple cores.  Hardware-specific events like
       AMDs STPCLK signalling ("shut 'er down ma, she's glowing red") can
       reach different cores at different times and lead to ramping up
       and down and different times and for different durations, leading
       to further drift.  Newer AMD CPUs fix this by providing drift-
       invariant TSCs if the TscInvariant CPUID feature flag is set.

       As a result, if we're on a system with multiple cores and it doesn't
       have the AMD TscInvariant fix then the TSC skew can also be a source
       of entropy.  To some extent this is just timing the context switch 
       time across CPUs (which in itself is a source of some entropy), but 
       what's left is the TSC skew across cores.  If this is if interest we
       can do it with code something like the following.  Note that this is 
       only sample code, there are various complications such as the fact 
       that another thread may change the process affinity mask between the
       call to the initial GetProcessAffinityMask() and the final
       SetThreadAffinityMask() to restore the original mask, even if we 
       insert another GetProcessAffinityMask() just before the 
       SetThreadAffinityMask() there's still a small race condition present 
       but no real way to avoid it without OS API-level support:

        DWORD processAfinityMask, systemAffinityMask, mask = 0x10000;

        // Get the available processors that the thread can be scheduled on
        if( !GetProcessAffinityMask( GetCurrentProcess(), processAfinityMask, 
                                     systemAffinityMask ) )
            return;

        // Move the thread to each processor and read the TSC
        while( mask != 0 )
            {
            mask >>= 1;
            if( !( mask & processAfinityMask ) )
                continue;

            SetThreadAffinityMask( GetCurrentThread(), mask );
            // RDTSC
            }

        // Restore the original thread affinity
        SetThreadAffinityMask( GetCurrentThread(), processAfinityMask );

       Finally, there's the HPET, but this is only supported in Vista and
       it's not clear how to access it from user-level APIs rather than as
       part of the multimedia subsystem.

       To make things unambiguous, we call RDTSC directly if possible and 
       fall back to QPC in the highly unlikely situation where this isn't 
       present */
#if defined( __WIN64__ )
    {
    unsigned __int64 value;
    
    /* x86-64 always has a TSC, and the read is supported as an intrinsic */
    value = __rdtsc();
    addRandomData( randomState, &value, sizeof( __int64 ) );
    }
#else
  #ifndef NO_ASM
    if( getSysVar( SYSVAR_HWCAP ) & HWCAP_FLAG_RDTSC )
        {
        unsigned long value;

        __asm {
            xor eax, eax
            xor edx, edx        /* Tell VC++ that EDX:EAX will be trashed */
            rdtsc
            mov [value], eax    /* Ignore high 32 bits, which are > 1s res */
            }
        addRandomValue( randomState, value );
        }
  #endif /* NO_ASM */
    else
        {
        LARGE_INTEGER performanceCount;

        if( QueryPerformanceCounter( &performanceCount ) )
            addRandomData( randomState, &performanceCount,
                           sizeof( LARGE_INTEGER ) );
        else
            {
            /* Millisecond accuracy at best... */
            addRandomValue( randomState, GetTickCount() );
            }
        }
#endif /* Win32 vs. Win64 */

    /* If there's a hardware RNG present, read data from it.  We check that
       the RNG is still present/enabled on each fetch since it could (at 
       least in theory) be disabled by the OS between fetches.  We also read 
       the data into an explicitly dword-aligned buffer (which the standard 
       buffer should be anyway, but we make it explicit here just to be 
       safe).  Note that we have to force alignment using a LONGLONG rather 
       than a #pragma pack, since chars don't need alignment it would have 
       no effect on the BYTE [] member */
#ifndef NO_ASM
    if( getSysVar( SYSVAR_HWCAP ) & HWCAP_FLAG_XSTORE )
        {
        struct alignStruct {
            LONGLONG dummy1;        /* Force alignment of following member */
            BYTE buffer[ 64 ];
            };
        struct alignStruct *rngBuffer = ( struct alignStruct * ) buffer;
        void *bufPtr = rngBuffer->buffer;   /* Get it into a form asm can handle */
        int byteCount = 0;

        __asm {
            push es
            xor ecx, ecx        /* Tell VC++ that ECX will be trashed */
            mov eax, 0xC0000001 /* Centaur extended feature flags */
            cpuid
            and edx, 01100b
            cmp edx, 01100b     /* Check for RNG present + enabled flags */
            jne rngDisabled     /* RNG was disabled after our initial check */
            push ds
            pop es
            mov edi, bufPtr     /* ES:EDI = buffer */
            xor edx, edx        /* Fetch 8 bytes */
            xstore_rng
            and eax, 011111b    /* Get count of bytes returned */
            jz rngDisabled      /* Nothing read, exit */
            mov [byteCount], eax
        rngDisabled:
            pop es
            }
        if( byteCount > 0 )
            {
            addRandomData( randomState, bufPtr, byteCount );
            trngValue = 45;
            }
        }
    if( getSysVar( SYSVAR_HWCAP ) & HWCAP_FLAG_RDRAND )
        {
        unsigned long buffer[ 8 + 8 ];
        int byteCount = 0;

        __asm {
            xor eax, eax        /* Tell VC++ that EAX will be trashed */
            xor ecx, ecx
        trngLoop:
            rdrand_eax
            jnc trngExit        /* TRNG result bad, exit with byteCount = 0 */
            mov [buffer+ecx], eax
            add ecx, 4
            cmp ecx, 32         /* Fill 32 bytes worth */
            jl trngLoop
            mov [byteCount], ecx
        trngExit:
            }
        if( byteCount > 0 )
            {
            addRandomData( randomState, buffer, byteCount );
            trngValue = 45;
            }
        }
    if( getSysVar( SYSVAR_HWCAP ) & HWCAP_FLAG_TRNG )
        {
        unsigned long value = 0;

        __asm {
            xor eax, eax
            xor edx, edx        /* Tell VC++ that EDX:EAX will be trashed */
            mov ecx, 0x58002006 /* GLD_MSR_CTRL */
            rdmsr
            and eax, 0110000000000b 
            cmp eax, 0010000000000b /* Check whether TRNG is enabled */
            jne trngDisabled    /* TRNG isn't enabled */
        trngDisabled:
            }
        if( value )
            {
            addRandomValue( randomState, value );
            trngValue = 20;
            }
        }
#endif /* NO_ASM */

    /* Flush any remaining data through.  Quality = int( 33 1/3 % ) + any 
       additional quality from the TRNG if there's one present */
    endRandomData( randomState, 34 + trngValue );
    }

/****************************************************************************
*                                                                           *
*                                   Slow Poll                               *
*                                                                           *
****************************************************************************/

/* If we're running under Win64 there's no need to include Win95/98 
   backwards-compatibility features */

#ifndef __WIN64__

/* Type definitions for function pointers to call Toolhelp32 functions */

#ifdef _MSC_VER
  #if VC_GE_2005( _MSC_VER )
    #define THREAD_ID   ULONG_PTR
  #else
    #define THREAD_ID   DWORD
  #endif /* VC++ < 2005 */
#endif /* VC++ */

typedef BOOL ( WINAPI *MODULEWALK )( HANDLE hSnapshot, LPMODULEENTRY32 lpme );
typedef BOOL ( WINAPI *THREADWALK )( HANDLE hSnapshot, LPTHREADENTRY32 lpte );
typedef BOOL ( WINAPI *PROCESSWALK )( HANDLE hSnapshot, LPPROCESSENTRY32 lppe );
typedef BOOL ( WINAPI *HEAPLISTWALK )( HANDLE hSnapshot, LPHEAPLIST32 lphl );
typedef BOOL ( WINAPI *HEAPFIRST )( LPHEAPENTRY32 lphe, DWORD th32ProcessID, THREAD_ID th32HeapID );
typedef BOOL ( WINAPI *HEAPNEXT )( LPHEAPENTRY32 lphe );
typedef HANDLE ( WINAPI *CREATESNAPSHOT )( DWORD dwFlags, DWORD th32ProcessID );

/* Global function pointers. These are necessary because the functions need to
   be dynamically linked since only the Win95 kernel currently contains them.
   Explicitly linking to them will make the program unloadable under NT */

static CREATESNAPSHOT pCreateToolhelp32Snapshot = NULL;
static MODULEWALK pModule32First = NULL;
static MODULEWALK pModule32Next = NULL;
static PROCESSWALK pProcess32First = NULL;
static PROCESSWALK pProcess32Next = NULL;
static THREADWALK pThread32First = NULL;
static THREADWALK pThread32Next = NULL;
static HEAPLISTWALK pHeap32ListFirst = NULL;
static HEAPLISTWALK pHeap32ListNext = NULL;
static HEAPFIRST pHeap32First = NULL;
static HEAPNEXT pHeap32Next = NULL;

/* Since there are a significant number of ToolHelp data blocks, we use a
   larger-than-usual intermediate buffer to cut down on kernel traffic */

#define BIG_RANDOM_BUFSIZE  ( RANDOM_BUFSIZE * 4 )
#if BIG_RANDOM_BUFSIZE > MAX_INTLENGTH_SHORT
  #error BIG_RANDOM_BUFSIZE exceeds randomness accumulator size
#endif /* RANDOM_BUFSIZE > MAX_INTLENGTH_SHORT */

static void slowPollWin95( void )
    {
    static BOOLEAN addedFixedItems = FALSE;
    PROCESSENTRY32 pe32;
    THREADENTRY32 te32;
    MODULEENTRY32 me32;
    HEAPLIST32 hl32;
    HANDLE hSnapshot;
    RANDOM_STATE randomState;
    BYTE buffer[ BIG_RANDOM_BUFSIZE + 8 ];
    int iterationCount, status;

    /* The following are fixed for the lifetime of the process so we only
       add them once */
    if( !addedFixedItems )
        {
        readPnPData();
        addedFixedItems = TRUE;
        }

    /* Initialize the Toolhelp32 function pointers if necessary */
    if( pCreateToolhelp32Snapshot == NULL )
        {
        HANDLE hKernel;

        /* Obtain the module handle of the kernel to retrieve the addresses
           of the Toolhelp32 functions */
        if( ( hKernel = GetModuleHandle( "Kernel32.dll" ) ) == NULL )
            return;

        /* Now get pointers to the functions */
        pCreateToolhelp32Snapshot = ( CREATESNAPSHOT ) GetProcAddress( hKernel,
                                                    "CreateToolhelp32Snapshot" );
        pModule32First = ( MODULEWALK ) GetProcAddress( hKernel,
                                                    "Module32First" );
        pModule32Next = ( MODULEWALK ) GetProcAddress( hKernel,
                                                    "Module32Next" );
        pProcess32First = ( PROCESSWALK ) GetProcAddress( hKernel,
                                                    "Process32First" );
        pProcess32Next = ( PROCESSWALK ) GetProcAddress( hKernel,
                                                    "Process32Next" );
        pThread32First = ( THREADWALK ) GetProcAddress( hKernel,
                                                    "Thread32First" );
        pThread32Next = ( THREADWALK ) GetProcAddress( hKernel,
                                                    "Thread32Next" );
        pHeap32ListFirst = ( HEAPLISTWALK ) GetProcAddress( hKernel,
                                                    "Heap32ListFirst" );
        pHeap32ListNext = ( HEAPLISTWALK ) GetProcAddress( hKernel,
                                                    "Heap32ListNext" );
        pHeap32First = ( HEAPFIRST ) GetProcAddress( hKernel,
                                                    "Heap32First" );
        pHeap32Next = ( HEAPNEXT ) GetProcAddress( hKernel,
                                                    "Heap32Next" );

        /* Make sure we got valid pointers for every Toolhelp32 function */
        if( pModule32First == NULL || pModule32Next == NULL || \
            pProcess32First == NULL || pProcess32Next == NULL || \
            pThread32First == NULL || pThread32Next == NULL || \
            pHeap32ListFirst == NULL || pHeap32ListNext == NULL || \
            pHeap32First == NULL || pHeap32Next == NULL || \
            pCreateToolhelp32Snapshot == NULL )
            {
            /* Mark the main function as unavailable in case for future
               reference */
            pCreateToolhelp32Snapshot = NULL;
            return;
            }
        }
    if( krnlIsExiting() )
        return;

    status = initRandomData( randomState, buffer, BIG_RANDOM_BUFSIZE );
    if( cryptStatusError( status ) )
        retIntError_Void();

    /* Take a snapshot of everything we can get to that's currently in the
       system */
    hSnapshot = pCreateToolhelp32Snapshot( TH32CS_SNAPALL, 0 );
    if( !hSnapshot )
        return;

    /* Walk through the local heap.  We have to be careful to not spend
       excessive amounts of time on this if we're linked into a large
       application with a great many heaps and/or heap blocks, since the
       heap-traversal functions are rather slow.  Fortunately this is
       quite rare under Win95/98, since it implies a large/long-running
       server app that would be run under NT/Win2K/et al rather than 
       Win95 (the performance of the mapped ToolHelp32 helper functions 
       under these OSes is even worse than under Win95, fortunately we 
       don't have to use them there).

       Ideally in order to prevent excessive delays we'd count the number
       of heaps and ensure that no_heaps * no_heap_blocks doesn't exceed
       some maximum value, however this requires two passes of (slow) heap
       traversal rather than one, which doesn't help the situation much.
       To provide at least some protection, we limit the total number of
       heaps and heap entries traversed, although this leads to slightly
       suboptimal performance if we have a small number of deep heaps
       rather than the current large number of shallow heaps.

       There is however a second consideration that needs to be taken into
       account when doing this, which is that the heap-management functions
       aren't completely thread-safe, so that under (very rare) conditions
       of heavy allocation/deallocation this can cause problems when calling
       HeapNext().  By limiting the amount of time that we spend in each
       heap, we can reduce our exposure somewhat */
    hl32.dwSize = sizeof( HEAPLIST32 );
    if( pHeap32ListFirst( hSnapshot, &hl32 ) )
        {
        int listCount = 0;

        do
            {
            HEAPENTRY32 he32;

            /* First add the information from the basic Heaplist32
               structure */
            if( krnlIsExiting() )
                {
                CloseHandle( hSnapshot );
                return;
                }
            addRandomData( randomState, &hl32, sizeof( HEAPLIST32 ) );

            /* Now walk through the heap blocks getting information
               on each of them */
            he32.dwSize = sizeof( HEAPENTRY32 );
            if( pHeap32First( &he32, hl32.th32ProcessID, hl32.th32HeapID ) )
                {
                int entryCount = 0;

                do
                    {
                    if( krnlIsExiting() )
                        {
                        CloseHandle( hSnapshot );
                        return;
                        }
                    addRandomData( randomState, &he32,
                                   sizeof( HEAPENTRY32 ) );
                    }
                while( entryCount++ < 20 && pHeap32Next( &he32 ) );
                }
            }
        while( listCount++ < 20 && pHeap32ListNext( hSnapshot, &hl32 ) );
        }

    /* Walk through all processes */
    pe32.dwSize = sizeof( PROCESSENTRY32 );
    iterationCount = 0;
    if( pProcess32First( hSnapshot, &pe32 ) )
        {
        do
            {
            if( krnlIsExiting() )
                {
                CloseHandle( hSnapshot );
                return;
                }
            addRandomData( randomState, &pe32, sizeof( PROCESSENTRY32 ) );
            }
        while( pProcess32Next( hSnapshot, &pe32 ) && \
               iterationCount++ < FAILSAFE_ITERATIONS_LARGE );
        }

    /* Walk through all threads */
    te32.dwSize = sizeof( THREADENTRY32 );
    iterationCount = 0;
    if( pThread32First( hSnapshot, &te32 ) )
        {
        do
            {
            if( krnlIsExiting() )
                {
                CloseHandle( hSnapshot );
                return;
                }
            addRandomData( randomState, &te32, sizeof( THREADENTRY32 ) );
            }
        while( pThread32Next( hSnapshot, &te32 ) && \
               iterationCount++ < FAILSAFE_ITERATIONS_LARGE );
        }

    /* Walk through all modules associated with the process */
    me32.dwSize = sizeof( MODULEENTRY32 );
    iterationCount = 0;
    if( pModule32First( hSnapshot, &me32 ) )
        {
        do
            {
            if( krnlIsExiting() )
                {
                CloseHandle( hSnapshot );
                return;
                }
            addRandomData( randomState, &me32, sizeof( MODULEENTRY32 ) );
            }
        while( pModule32Next( hSnapshot, &me32 ) && \
               iterationCount++ < FAILSAFE_ITERATIONS_LARGE );
        }

    /* Clean up the snapshot */
    CloseHandle( hSnapshot );
    if( krnlIsExiting() )
        return;

    /* Flush any remaining data through */
    endRandomData( randomState, 100 );
    }

/* Perform a thread-safe slow poll for Windows 95 */

unsigned __stdcall threadSafeSlowPollWin95( void *dummy )
    {
    UNUSED_ARG( dummy );

    slowPollWin95();
    _endthreadex( 0 );
    return( 0 );
    }
#endif /* __WIN64__ */

/* Type definitions for function pointers to call NetAPI32 functions */

typedef DWORD ( WINAPI *NETSTATISTICSGET )( LPWSTR szServer, LPWSTR szService,
                                            DWORD dwLevel, DWORD dwOptions,
                                            LPBYTE *lpBuffer );
typedef DWORD ( WINAPI *NETAPIBUFFERSIZE )( LPVOID lpBuffer, LPDWORD cbBuffer );
typedef DWORD ( WINAPI *NETAPIBUFFERFREE )( LPVOID lpBuffer );

/* Type definitions for functions to call native NT functions */

typedef DWORD ( WINAPI *NTQUERYSYSTEMINFORMATION )( DWORD systemInformationClass,
                                PVOID systemInformation,
                                ULONG systemInformationLength,
                                PULONG returnLength );
typedef DWORD ( WINAPI *NTQUERYINFORMATIONPROCESS )( HANDLE processHandle,
                                DWORD processInformationClass,
                                PVOID processInformation,
                                ULONG processInformationLength,
                                PULONG returnLength );
typedef DWORD ( WINAPI *NTPOWERINFORMATION )( DWORD powerInformationClass,
                                PVOID inputBuffer, ULONG inputBufferLength,
                                PVOID outputBuffer, ULONG outputBufferLength );

/* Global function pointers. These are necessary because the functions need to
   be dynamically linked since only the WinNT kernel currently contains them.
   Explicitly linking to them will make the program unloadable under Win95 */

static NETSTATISTICSGET pNetStatisticsGet = NULL;
static NETAPIBUFFERSIZE pNetApiBufferSize = NULL;
static NETAPIBUFFERFREE pNetApiBufferFree = NULL;
static NTQUERYSYSTEMINFORMATION pNtQuerySystemInformation = NULL;
static NTQUERYINFORMATIONPROCESS pNtQueryInformationProcess = NULL;
static NTPOWERINFORMATION pNtPowerInformation = NULL;

/* When we query the performance counters, we allocate an initial buffer and
   then reallocate it as required until RegQueryValueEx() stops returning
   ERROR_MORE_DATA.  The following values define the initial buffer size and
   step size by which the buffer is increased */

#define PERFORMANCE_BUFFER_SIZE     65536   /* Start at 64K */
#define PERFORMANCE_BUFFER_STEP     16384   /* Step by 16K */

static void registryPoll( void )
    {
    static int cbPerfData = PERFORMANCE_BUFFER_SIZE;
    PPERF_DATA_BLOCK pPerfData;
    MESSAGE_DATA msgData;
    DWORD dwSize, dwStatus;
    int iterations = 0, status;

    /* Wait for any async keyset driver binding to complete.  You may be
       wondering what this call is doing here... the reason why it's 
       necessary is because RegQueryValueEx() will hang indefinitely if the 
       async driver bind is in progress.  The problem occurs in the dynamic 
       loading and linking of driver DLL's, which work as follows:

        hDriver = DynamicLoad( DRIVERNAME );
        pFunction1 = ( TYPE_FUNC1 ) GetProcAddress( hDriver, NAME_FUNC1 );
        pFunction2 = ( TYPE_FUNC1 ) GetProcAddress( hDriver, NAME_FUNC2 );

       If RegQueryValueEx() is called while the GetProcAddress()'s are in
       progress, it will hang indefinitely.  This is probably due to some
       synchronisation problem in the NT kernel where the GetProcAddress()
       calls affect something like a module reference count or function
       reference count while RegQueryValueEx() is trying to take a snapshot of
       the statistics, which include the reference counts.  Because of this,
       we have to wait until any async driver bind has completed before we can
       call RegQueryValueEx() */
    if( !krnlWaitSemaphore( SEMAPHORE_DRIVERBIND ) )
        {
        /* The kernel is shutting down, bail out */
        return;
        }

    /* Get information from the system performance counters.  This can take a
       few seconds to do.  In some environments the call to RegQueryValueEx()
       can produce an access violation at some random time in the future, in
       some cases adding a short delay after the following code block makes
       the problem go away.  This problem is extremely difficult to
       reproduce, I haven't been able to get it to occur despite running it
       on a number of machines.  MS knowledge base article Q178887 covers
       this type of problem, it's typically caused by an external driver or
       other program that adds its own values under the
       HKEY_PERFORMANCE_DATA key.  The NT kernel, via Advapi32.dll, calls the
       required external module to map in the data inside an SEH try/except
       block, so problems in the module's collect function don't pop up until
       after it has finished, so the fault appears to occur in Advapi32.dll.
       There may be problems in the NT kernel as well though, a low-level
       memory checker indicated that ExpandEnvironmentStrings() in
       Kernel32.dll, called an interminable number of calls down inside
       RegQueryValueEx(), was overwriting memory (it wrote twice the
       allocated size of a buffer to a buffer allocated by the NT kernel).
       OTOH this could be coming from the external module calling back into
       the kernel, which eventually causes the problem described above.

       Possibly as an extension of the problem that the krnlWaitSemaphore()
       call above works around, running two instances of cryptlib (e.g. two
       applications that use it) under NT4 can result in one of them hanging
       in the RegQueryValueEx() call.  This happens only under NT4 and is
       hard to reproduce in any consistent manner.

       One workaround that helps a bit is to read the registry as a remote
       (rather than local) registry, it's possible that the use of a network
       RPC call isolates the calling app from the problem in that whatever
       service handles the RPC is taking the hit and not affecting the
       calling app.  Since this would require another round of extensive
       testing to verify and the NT native API call is working fine, we'll
       stick with the native API call for now.

       Some versions of NT4 had a problem where the amount of data returned
       was mis-reported and would never settle down, because of this the code
       below includes a safety-catch that bails out after 10 attempts have
       been made, this results in no data being returned but at does ensure
       that the thread will terminate.

       In addition to these problems the code in RegQueryValueEx() that
       estimates the amount of memory required to return the performance
       counter information isn't very accurate (it's much worse than the
       "slightly-inaccurate" level that the MS docs warn about, it's usually
       wildly off) since it always returns a worst-case estimate which is
       usually nowhere near the actual amount required.  For example it may
       report that 128K of memory is required, but only return 64K of data.

       Even worse than the registry-based performance counters is the
       performance data helper (PDH) shim that tries to make the counters
       look like the old Win16 API (which is also used by Win95).  Under NT
       this can consume tens of MB of memory and huge amounts of CPU time
       while it gathers its data, and even running once can still consume
       about 1/2MB of memory.

       "Windows NT is a thing of genuine beauty, if you're seriously into 
        genuine ugliness.  It's like a woman with a history of insanity in 
        the family, only worse" -- Hans Chloride, "Why I Love Windows NT" */
    pPerfData = ( PPERF_DATA_BLOCK ) clAlloc( "slowPollNT", cbPerfData );
    while( pPerfData != NULL && iterations++ < 10 )
        {
        dwSize = cbPerfData;
        dwStatus = RegQueryValueEx( HKEY_PERFORMANCE_DATA, "Global", NULL,
                                    NULL, ( LPBYTE ) pPerfData, &dwSize );
        if( dwStatus == ERROR_SUCCESS )
            {
            if( !memcmp( pPerfData->Signature, L"PERF", 8 ) )
                {
                static const int quality = 100;

                setMessageData( &msgData, pPerfData, dwSize );
                status = krnlSendMessage( SYSTEM_OBJECT_HANDLE,
                                          IMESSAGE_SETATTRIBUTE_S, &msgData,
                                          CRYPT_IATTRIBUTE_ENTROPY );
                if( cryptStatusOK( status ) )
                    krnlSendMessage( SYSTEM_OBJECT_HANDLE,
                                     IMESSAGE_SETATTRIBUTE,
                                     ( MESSAGE_CAST ) &quality,
                                     CRYPT_IATTRIBUTE_ENTROPY_QUALITY );
                }
            clFree( "slowPollWinNT", pPerfData );
            pPerfData = NULL;
            }
        else
            {
            if( dwStatus == ERROR_MORE_DATA )
                {
                PPERF_DATA_BLOCK pTempPerfData;
                
                cbPerfData += PERFORMANCE_BUFFER_STEP;
                pTempPerfData = ( PPERF_DATA_BLOCK ) realloc( pPerfData, cbPerfData );
                if( pTempPerfData == NULL )
                    {
                    /* The realloc failed, free the original block and force 
                       a loop exit */
                    clFree( "slowPollWinNT", pPerfData );
                    pPerfData = NULL;
                    }
                else
                    pPerfData = pTempPerfData;
                }
            }
        }

    /* Although this isn't documented in the Win32 API docs, it's necessary to
       explicitly close the HKEY_PERFORMANCE_DATA key after use (it's
       implicitly opened on the first call to RegQueryValueEx()).  If this
       isn't done then any system components that provide performance data
       can't be removed or changed while the handle remains active */
    RegCloseKey( HKEY_PERFORMANCE_DATA );
    }

static void slowPollWinNT( void )
    {
    static BOOLEAN addedFixedItems = FALSE;
    static int isWorkstation = CRYPT_ERROR;
    MESSAGE_DATA msgData;
    HANDLE hDevice;
    LPBYTE lpBuffer;
    ULONG ulSize;
    DWORD dwType, dwSize, dwResult;
    void *buffer;
    int nDrive, noResults = 0, status;

    /* Find out whether this is an NT server or workstation if necessary */
    if( isWorkstation == CRYPT_ERROR )
        {
        HKEY hKey;

        if( RegOpenKeyEx( HKEY_LOCAL_MACHINE,
                          "SYSTEM\\CurrentControlSet\\Control\\ProductOptions",
                          0, KEY_READ, &hKey ) == ERROR_SUCCESS )
            {
            BYTE szValue[ 32 + 8 ];
            dwSize = 32;

            isWorkstation = TRUE;
            if( RegQueryValueEx( hKey, "ProductType", 0, NULL, szValue,
                                 &dwSize ) == ERROR_SUCCESS && \
                dwSize >= 5 && strCompare( szValue, "WinNT", 5 ) )
                {
                /* Note: There are (at least) three cases for ProductType:
                   WinNT = NT Workstation, ServerNT = NT Server, LanmanNT =
                   NT Server acting as a Domain Controller */
                isWorkstation = FALSE;
                }

            RegCloseKey( hKey );
            }
        }

    /* The following are fixed for the lifetime of the process so we only
       add them once */
    if( !addedFixedItems )
        {
        readPnPData();
        addedFixedItems = TRUE;
        }

    /* Initialize the NetAPI32 function pointers if necessary */
    if( hNetAPI32 == NULL )
        {
        /* Obtain a handle to the module containing the Lan Manager functions */
        if( ( hNetAPI32 = DynamicLoad( "NetAPI32.dll" ) ) != NULL )
            {
            /* Now get pointers to the functions */
            pNetStatisticsGet = ( NETSTATISTICSGET ) GetProcAddress( hNetAPI32,
                                                        "NetStatisticsGet" );
            pNetApiBufferSize = ( NETAPIBUFFERSIZE ) GetProcAddress( hNetAPI32,
                                                        "NetApiBufferSize" );
            pNetApiBufferFree = ( NETAPIBUFFERFREE ) GetProcAddress( hNetAPI32,
                                                        "NetApiBufferFree" );

            /* Make sure we got valid pointers for every NetAPI32 function */
            if( pNetStatisticsGet == NULL ||
                pNetApiBufferSize == NULL ||
                pNetApiBufferFree == NULL )
                {
                /* Free the library reference and reset the static handle */
                DynamicUnload( hNetAPI32 );
                hNetAPI32 = NULL;
                }
            }
        }

    /* Initialize the NT kernel native API function pointers if necessary */
    if( hNTAPI == NULL && \
        ( hNTAPI = GetModuleHandle( "NTDll.dll" ) ) != NULL )
        {
        /* Get a pointer to the NT native information query functions */
        pNtQuerySystemInformation = ( NTQUERYSYSTEMINFORMATION ) \
                        GetProcAddress( hNTAPI, "NtQuerySystemInformation" );
        pNtQueryInformationProcess = ( NTQUERYINFORMATIONPROCESS ) \
                        GetProcAddress( hNTAPI, "NtQueryInformationProcess" );
        if( pNtQuerySystemInformation == NULL || \
            pNtQueryInformationProcess == NULL )
            hNTAPI = NULL;
        pNtPowerInformation = ( NTPOWERINFORMATION ) \
                        GetProcAddress( hNTAPI, "NtPowerInformation" );
        }
    if( krnlIsExiting() )
        return;

    /* Get network statistics.  Note: Both NT Workstation and NT Server by
       default will be running both the workstation and server services.  The
       heuristic below is probably useful though on the assumption that the
       majority of the network traffic will be via the appropriate service. In
       any case the network statistics return almost no randomness */
    if( hNetAPI32 != NULL &&
        pNetStatisticsGet( NULL,
                           isWorkstation ? L"LanmanWorkstation" : L"LanmanServer",
                           0, 0, &lpBuffer ) == 0 )
        {
        pNetApiBufferSize( lpBuffer, &dwSize );
        setMessageData( &msgData, lpBuffer, dwSize );
        krnlSendMessage( SYSTEM_OBJECT_HANDLE, IMESSAGE_SETATTRIBUTE_S,
                         &msgData, CRYPT_IATTRIBUTE_ENTROPY );
        pNetApiBufferFree( lpBuffer );
        }

    /* Get disk I/O statistics for all the hard drives */
    for( nDrive = 0; nDrive < FAILSAFE_ITERATIONS_MED; nDrive++ )
        {
        BYTE diskPerformance[ 256 + 8 ];
        char szDevice[ 32 + 8 ];

        /* Check whether we can access this device */
        sprintf_s( szDevice, 32, "\\\\.\\PhysicalDrive%d", nDrive );
        hDevice = CreateFile( szDevice, 0, FILE_SHARE_READ | FILE_SHARE_WRITE,
                              NULL, OPEN_EXISTING, 0, NULL );
        if( hDevice == INVALID_HANDLE_VALUE )
            break;

        /* Note: This only works if the user has turned on the disk
           performance counters with 'diskperf -y'.  These counters were
           traditionally disabled under NT, although in newer installs of 
           Win2K and newer the physical disk object is enabled by default 
           while the logical disk object is disabled by default 
           ('diskperf -yv' turns on the counters for logical drives in this 
           case, since they're already on for physical drives).
           
           In addition using the documented DISK_PERFORMANCE data structure 
           to contain the returned data returns ERROR_INSUFFICIENT_BUFFER 
           (which is wrong) and doesn't change dwSize (which is also wrong) 
           so we pass in a larger buffer and pre-set dwSize to a safe 
           value.  Finally, there's a bug in pre-SP4 Win2K in which enabling 
           diskperf, installing a file system filter driver, and then 
           disabling diskperf, causes diskperf to corrupt the registry key 
           HKEY_LOCAL_MACHINE\SYSTEM\CurrentControlSet\Control\Class\
           {71A27CDD-812A-11D0-BEC7-08002BE2092F}\Upper Filters, resulting 
           in a Stop 0x7B bugcheck */
        dwSize = sizeof( diskPerformance );
        if( DeviceIoControl( hDevice, IOCTL_DISK_PERFORMANCE, NULL, 0,
                             &diskPerformance, 256, &dwSize, NULL ) )
            {
            if( krnlIsExiting() )
                {
                CloseHandle( hDevice );
                return;
                }
            setMessageData( &msgData, &diskPerformance, dwSize );
            krnlSendMessage( SYSTEM_OBJECT_HANDLE, IMESSAGE_SETATTRIBUTE_S,
                             &msgData, CRYPT_IATTRIBUTE_ENTROPY );
            }
        CloseHandle( hDevice );
        }
    if( krnlIsExiting() )
        return;

    /* In theory we should be using the Win32 performance query API to obtain
       unpredictable data from the system, however this is so unreliable (see
       the multiple sets of comments in registryPoll()) that it's too risky
       to rely on it except as a fallback in emergencies.  Instead, we rely
       mostly on the NT native API function NtQuerySystemInformation(), which
       has the dual advantages that it doesn't have as many (known) problems
       as the Win32 equivalent and that it doesn't access the data indirectly
       via pseudo-registry keys, which means that it's much faster.  Note
       that the Win32 equivalent actually works almost all of the time, the
       problem is that on one or two systems it can fail in strange ways that
       are never the same and can't be reproduced on any other system, which
       is why we use the native API here.  Microsoft officially documented
       this function in early 2003, so it'll be fairly safe to use */
    if( ( hNTAPI == NULL ) || \
        ( buffer = clAlloc( "slowPollNT", PERFORMANCE_BUFFER_SIZE ) ) == NULL )
        {
        registryPoll();
        return;
        }

    /* Clear the buffer before use.  We have to do this because even though 
       NtQuerySystemInformation() tells us that it's filled in ulSize bytes, 
       it doesn't necessarily mean that it actually has provided that much 
       data */
    memset( buffer, 0, PERFORMANCE_BUFFER_SIZE );

    /* Scan the first 64 possible information types (we don't bother with
       increasing the buffer size as we do with the Win32 version of the
       performance data read, we may miss a few classes but it's no big deal).
       This scan typically yields around 20 pieces of data, there's nothing
       in the range 65...128 so chances are there won't be anything above
       there either */
    for( dwType = 0; dwType < 64; dwType++ )
        {
        /* Some information types are write-only (the IDs are shared with
           a set-information call), we skip these */
        if( dwType == 26 || dwType == 27 || dwType == 38 || \
            dwType == 38 || dwType == 46 || dwType == 47 || \
            dwType == 48 || dwType == 52 )
            continue;

        /* ID 53 = SystemSessionProcessInformation reads input from the
           output buffer, which has to contain a session ID and pointer
           to the actual buffer in which to store the session information.
           Because this isn't a standard query, we skip this */
        if( dwType == 53 )
            continue;

        /* Query the info for this ID.  Some results (for example for
           ID = 6, SystemCallCounts) are only available in checked builds
           of the kernel.  A smaller subcless of results require that
           certain system config flags be set, for example
           SystemObjectInformation requires that the
           FLG_MAINTAIN_OBJECT_TYPELIST be set in NtGlobalFlags.  To avoid
           having to special-case all of these, we try reading each one and
           only use those for which we get a success status */
        dwResult = pNtQuerySystemInformation( dwType, buffer,
                                              PERFORMANCE_BUFFER_SIZE - 2048,
                                              &ulSize );
        if( dwResult != ERROR_SUCCESS )
            continue;

        /* Some calls (e.g. ID = 23, SystemProcessorStatistics, and ID = 24,
           SystemDpcInformation) incorrectly return a length of zero, so we
           manually adjust the length to the correct value */
        if( ulSize == 0 )
            {
            if( dwType == 23 )
                ulSize = 6 * sizeof( ULONG );
            if( dwType == 24 )
                ulSize = 5 * sizeof( ULONG );
            }

        /* If we got some data back, add it to the entropy pool.  Note that 
           just because NtQuerySystemInformation() tells us that it's 
           provided ulSize bytes doesn't necessarily mean that it has, see 
           the comment after the malloc() for details */
        if( ulSize > 0 && ulSize <= PERFORMANCE_BUFFER_SIZE - 2048 )
            {
            if( krnlIsExiting() )
                {
                clFree( "slowPollWinNT", buffer );
                return;
                }
            setMessageData( &msgData, buffer, ulSize );
            status = krnlSendMessage( SYSTEM_OBJECT_HANDLE,
                                      IMESSAGE_SETATTRIBUTE_S, &msgData,
                                      CRYPT_IATTRIBUTE_ENTROPY );
            if( cryptStatusOK( status ) )
                noResults++;
            }
        }

    /* Now do the same for the process information.  This call is rather ugly
       in that it requires an exact length match for the data returned,
       failing with a STATUS_INFO_LENGTH_MISMATCH error code (0xC0000004) if
       the length isn't an exact match */
#if 0   /* This requires compiler-level assistance to handle complex nested 
           structs, alignment issues, and so on.  Without the headers in 
           which the entries are declared it's almost impossible to do */
    for( dwType = 0; dwType < 32; dwType++ )
        {
        static const struct { int type; int size; } processInfo[] = {
            { CRYPT_ERROR, CRYPT_ERROR }
            };
        int i;

        for( i = 0; processInfo[ i ].type != CRYPT_ERROR && \
                    i < FAILSAFE_ITERATIONS_SMALL; i++ )
            {
            /* Query the info for this ID */
            dwResult = pNtQueryInformationProcess( GetCurrentProcess(),
                                            processInfo[ i ].type, buffer,
                                            processInfo[ i ].size, &ulSize );
            if( dwResult != ERROR_SUCCESS )
                continue;

            /* If we got some data back, add it to the entropy pool */
            if( ulSize > 0 && ulSize <= PERFORMANCE_BUFFER_SIZE - 2048 )
                {
                if( krnlIsExiting() )
                    {
                    clFree( "slowPollWinNT", buffer );
                    return;
                    }
                setMessageData( &msgData, buffer, ulSize );
                status = krnlSendMessage( SYSTEM_OBJECT_HANDLE,
                                          IMESSAGE_SETATTRIBUTE_S, &msgData,
                                          CRYPT_IATTRIBUTE_ENTROPY );
                if( cryptStatusOK( status ) )
                    noResults++;
                }
            }
        ENSURES( i < FAILSAFE_ITERATIONS_SMALL );
        }
#endif /* 0 */

    /* Finally do the same for the system power status information.  There 
       are only a limited number of useful information types available so we
       restrict ourselves to the useful types.  In addition since this
       function doesn't return length information we have to hardcode in
       length data */
    if( pNtPowerInformation != NULL )
        {
        static const struct { int type; int size; } powerInfo[] = {
            { 0, 128 }, /* SystemPowerPolicyAc */
            { 1, 128 }, /* SystemPowerPolicyDc */
            { 4, 64 },  /* SystemPowerCapabilities */
            { 5, 48 },  /* SystemBatteryState */
            { 11, 48 }, /* ProcessorInformation */
            { 12, 24 }, /* SystemPowerInformation */
            { CRYPT_ERROR, CRYPT_ERROR }
            };
        int i;

        for( i = 0; powerInfo[ i ].type != CRYPT_ERROR && \
                    i < FAILSAFE_ITERATIONS_MED; i++ )
            {
            /* Query the info for this ID */
            dwResult = pNtPowerInformation( powerInfo[ i ].type, NULL, 0, buffer,
                                            PERFORMANCE_BUFFER_SIZE - 2048 );
            if( dwResult != ERROR_SUCCESS )
                continue;
            if( krnlIsExiting() )
                {
                clFree( "slowPollWinNT", buffer );
                return;
                }
            setMessageData( &msgData, buffer, powerInfo[ i ].size );
            status = krnlSendMessage( SYSTEM_OBJECT_HANDLE,
                                      IMESSAGE_SETATTRIBUTE_S, &msgData,
                                      CRYPT_IATTRIBUTE_ENTROPY );
            if( cryptStatusOK( status ) )
                noResults++;
            }
        ENSURES_V( i < FAILSAFE_ITERATIONS_MED );
        }
    clFree( "slowPollWinNT", buffer );

    /* If we got enough data, we can leave now without having to try for a
       Win32-level performance information query */
    if( noResults > 15 )
        {
        static const int quality = 100;

        if( krnlIsExiting() )
            return;
        krnlSendMessage( SYSTEM_OBJECT_HANDLE, IMESSAGE_SETATTRIBUTE,
                         ( MESSAGE_CAST ) &quality,
                         CRYPT_IATTRIBUTE_ENTROPY_QUALITY );
        return;
        }
    if( krnlIsExiting() )
        return;

    /* We couldn't get enough results from the kernel, fall back to the
       somewhat troublesome registry poll */
    registryPoll();
    }

/* Perform a thread-safe slow poll for Windows NT */

unsigned __stdcall threadSafeSlowPollWinNT( void *dummy )
    {
#if 0
    typedef BOOL ( WINAPI *CREATERESTRICTEDTOKEN )( HANDLE ExistingTokenHandle,
                                DWORD Flags, DWORD DisableSidCount,
                                PSID_AND_ATTRIBUTES SidsToDisable,
                                DWORD DeletePrivilegeCount,
                                PLUID_AND_ATTRIBUTES PrivilegesToDelete,
                                DWORD RestrictedSidCount,
                                PSID_AND_ATTRIBUTES SidsToRestrict,
                                PHANDLE NewTokenHandle );
    static CREATERESTRICTEDTOKEN pCreateRestrictedToken = NULL;
    static BOOLEAN isInited = FALSE;
#endif /* 0 */

    UNUSED_ARG( dummy );

    /* If the poll performed any kind of active operation like the Unix one
       rather than just basic data reads it'd be a good idea to drop 
       privileges before we begin, something that can be performed by the
       following code.  Note though that this creates all sorts of 
       complications when cryptlib is running as a service and/or performing
       impersonation, which is why it's disabled by default */
#if 0
    if( !isInited )
        {
        /* Since CreateRestrictedToken() is a Win2K function we can only use
           it on a post-NT4 system, and have to bind it at runtime */
        if( getSysVar( SYSVAR_OSVERSION ) > 4 )
            {
            const HINSTANCE hAdvAPI32 = GetModuleHandle( "AdvAPI32.dll" );

            pCreateRestrictedToken = ( CREATERESTRICTEDTOKEN ) \
                        GetProcAddress( hAdvAPI32, "CreateRestrictedToken" );
            }
        isInited = TRUE;
        }
    if( pCreateRestrictedToken != NULL )
        {
        HANDLE hToken, hNewToken;

        ImpersonateSelf( SecurityImpersonation );
        OpenThreadToken( GetCurrentThread(),
                         TOKEN_ASSIGN_PRIMARY | TOKEN_DUPLICATE | \
                         TOKEN_QUERY | TOKEN_ADJUST_DEFAULT | \
                         TOKEN_IMPERSONATE, TRUE, &hToken );
        CreateRestrictedToken( hToken, DISABLE_MAX_PRIVILEGE, 0, NULL, 0, NULL,
                               0, NULL, &hNewToken );
        SetThreadToken( &hThread, hNewToken );
        }
#endif /* 0 */

    slowPollWinNT();
#if 0
    if( pCreateRestrictedToken != NULL )
        RevertToSelf();
#endif /* 0 */
    _endthreadex( 0 );
    return( 0 );
    }

/* Perform a generic slow poll.  This starts the OS-specific poll in a
   separate thread */

void slowPoll( void )
    {
    if( krnlIsExiting() )
        return;

    /* Read data from various hardware sources */
    readSystemRNG();
    readMBMData();
    readEverestData();
    readSysToolData();
    readRivaTunerData();
    readHMonitorData();
    readATITrayToolsData();
    readCoreTempData();
    readGPUZData();

    /* Start a threaded slow poll.  If a slow poll is already running, we
       just return since there isn't much point in running two of them at the
       same time */
    if( cryptStatusError( krnlEnterMutex( MUTEX_RANDOM ) ) )
        return;
    if( hThread != NULL )
        {
        krnlExitMutex( MUTEX_RANDOM );
        return;
        }
#ifndef __WIN64__
    if( getSysVar( SYSVAR_ISWIN95 ) == TRUE )
        {
        hThread = ( HANDLE ) _beginthreadex( NULL, 0, threadSafeSlowPollWin95,
                                             NULL, 0, &threadID );
        }
    else
#endif /* __WIN64__ */
        {
        /* In theory since the thread is gathering info used (eventually)
           for crypto keys we could set an ACL on the thread to make it
           explicit that no-one else can mess with it:

            void *aclInfo = initACLInfo( THREAD_ALL_ACCESS );

            hThread = ( HANDLE ) _beginthreadex( getACLInfo( aclInfo ),
                                                 0, threadSafeSlowPollWinNT,
                                                 NULL, 0, &threadID );
            freeACLInfo( aclInfo );

          However, although this is supposed to be the default access for
          threads anyway, when used from a service (running under the
          LocalSystem account) under Win2K SP4 and up, the thread creation
          fails with error = 22 (invalid parameter).  Presumably MS patched
          some security hole or other in SP4, which causes the thread
          creation to fail.  Because of this problem, we don't set an ACL for
          the thread */
        hThread = ( HANDLE ) _beginthreadex( NULL, 0,
                                             threadSafeSlowPollWinNT,
                                             NULL, 0, &threadID );
        }
    krnlExitMutex( MUTEX_RANDOM );
    assert( hThread );
    }

/* Wait for the randomness gathering to finish.  Anything that requires the
   gatherer process to have completed gathering entropy should call
   waitforRandomCompletion(), which will block until the background process
   completes */

CHECK_RETVAL \
int waitforRandomCompletion( const BOOLEAN force )
    {
    DWORD dwResult;
    const DWORD timeout = force ? 2000 : 300000L;
    int status;

    /* If there's no polling thread running, there's nothing to do.  Note
       that this isn't entirely thread-safe because someone may start
       another poll after we perform this check, but there's no way to
       handle this without some form of interlock mechanism with the
       randomness mutex and the WaitForSingleObject().  In any case all
       that'll happen is that the caller won't get all of the currently-
       polling entropy */
    if( hThread == NULL )
        return( CRYPT_OK );

    /* Wait for the polling thread to terminate.  If it's a forced shutdown
       we only wait a short amount of time (2s) before we bail out, 
       otherwise we hang around for as long as it takes (with a sanity-check
       upper limit of 5 minutes) */
    dwResult = WaitForSingleObject( hThread, timeout );
    if( dwResult != WAIT_OBJECT_0 )
        {
        /* Since this is a cleanup function there's not much that we can do 
           at this point, although we warn in debug mode */
        DEBUG_DIAG(( "Error %lX waiting for object", dwResult ));
        assert( DEBUG_WARN );
        return( CRYPT_OK );
        }
    assert( dwResult != WAIT_FAILED );  /* Warn in debug mode */

    /* Clean up */
    status = krnlEnterMutex( MUTEX_RANDOM );
    if( cryptStatusError( status ) )
        return( status );
    if( hThread != NULL )
        {
        CloseHandle( hThread );
        hThread = NULL;
        }
    krnlExitMutex( MUTEX_RANDOM );

    return( CRYPT_OK );
    }

/* Initialise and clean up any auxiliary randomness-related objects */

void initRandomPolling( void )
    {
    /* Reset the various module and object handles and status info and
       initialise the system RNG interface if it's present */
    hAdvAPI32 = hNetAPI32 = hThread = NULL;
    initSystemRNG();
    }

void endRandomPolling( void )
    {
    assert( hThread == NULL );
    if( hNetAPI32 )
        {
        DynamicUnload( hNetAPI32 );
        hNetAPI32 = NULL;
        }
    endSystemRNG();
    }