path: root/src/OS/os.c-cygwin

/* $Cambridge: exim/src/OS/os.c-cygwin,v 1.3 2005/08/02 09:09:27 ph10 Exp $

/*************************************************
*     Exim - an Internet mail transport agent    *
*************************************************/

/* Cygwin-specific code. December 2002
   This is concatenated onto the generic src/os.c file.

   This code was supplied by Pierre A. Humblet <Pierre.Humblet@ieee.org>
*/

/* We need a special mkdir that
   allows names starting with // */
#undef mkdir
int cygwin_mkdir( const char *path, mode_t mode )
{
  const char * p = path;
  if (*p == '/') while(*(p+1) == '/') p++;
  return mkdir(p, mode);
}

/* We have strsignal but cannot use #define
   because types don't match */
#define OS_STRSIGNAL /* src/os.c need not provide it */
char * os_strsignal(int sig)
{
  return (char *) strsignal(sig);
}

#ifndef COMPILE_UTILITY /* Utilities don't need special code */
#ifdef INCLUDE_MINIRES
#include "../minires/minires.c"
#include "../minires/os-interface.c"
#endif

#ifdef INCLUDE_PAM
#include "../pam/pam.c"
#endif

unsigned int cygwin_WinVersion;

/* Conflict between Windows definitions and others */
#ifdef NOERROR
#undef NOERROR
#endif
#ifdef DELETE
#undef DELETE
#endif

#include <windows.h>
#include <sys/cygwin.h>

/* Special static variables */
static BOOL cygwin_debug = FALSE;
static int privileged = 1; /* when not privileged, setuid = noop */

#undef setuid
int cygwin_setuid(uid_t uid )
{
  int res;
  if (privileged <= 0) return 0;
  else {
    res = setuid(uid);
    if (cygwin_debug)
      fprintf(stderr, "setuid %lu %lu %d pid: %d\n",
              uid, getuid(),res, getpid());
  }
  return res;
}

#undef setgid
int cygwin_setgid(gid_t gid )
{
  int res;
  if (privileged <= 0) return 0;
  else {
    res = setgid(gid);
    if (cygwin_debug)
      fprintf(stderr, "setgid %lu %lu %d pid: %d\n",
              gid, getgid(), res, getpid());
  }
  return res;
}

/* Background processes run at lower priority */
static void cygwin_setpriority()
{
  if (!SetPriorityClass(GetCurrentProcess(), BELOW_NORMAL_PRIORITY_CLASS))
    SetPriorityClass(GetCurrentProcess(), IDLE_PRIORITY_CLASS);
  return;
}


/* GetVersion()
   MSB: 1 for 95/98/ME; Next 7: build number, except for 95/98/ME
   Next byte: 0
   Next byte: minor version of OS
   Low  byte: major version of OS (3 or 4 for for NT, 5 for 2000 and XP) */
#define VERSION_IS_58M(x) (x & 0x80000000) /* 95, 98, Me   */
#define VERSION_IS_NT(x)  ((x & 0XFF) < 5) /* NT 4 or 3.51 */

/*
  Routine to find if process or thread is privileged
*/

enum {
  CREATE_BIT = 1,
  RESTORE_BIT = 2
};

static DWORD get_privileges ()
{
  char buffer[1024];
  DWORD i, length;
  HANDLE hToken = NULL;
  PTOKEN_PRIVILEGES privs;
  LUID cluid, rluid;
  DWORD ret = 0;

  privs = (PTOKEN_PRIVILEGES) buffer;

  if (OpenProcessToken (GetCurrentProcess(), TOKEN_QUERY, &hToken)
      && LookupPrivilegeValue (NULL, SE_CREATE_TOKEN_NAME, &cluid)
      && LookupPrivilegeValue(NULL, SE_RESTORE_NAME, &rluid)
      && (GetTokenInformation( hToken, TokenPrivileges,
                               privs, sizeof (buffer), &length)
          || (GetLastError () == ERROR_INSUFFICIENT_BUFFER
              && (privs = (PTOKEN_PRIVILEGES) alloca (length))
              && GetTokenInformation(hToken, TokenPrivileges,
                                     privs, length, &length)))) {
    for (i = 0; i < privs->PrivilegeCount; i++) {
      if (privs->Privileges[i].Luid.QuadPart == cluid.QuadPart)
        ret |= CREATE_BIT;
      else if (privs->Privileges[i].Luid.QuadPart == rluid.QuadPart)
        ret |= RESTORE_BIT;
      else continue;
      if (ret == (CREATE_BIT | RESTORE_BIT))
        break;
    }
  }
  else
    fprintf(stderr, "has_create_token_privilege %ld\n", GetLastError());

  if (hToken)
    CloseHandle(hToken);

  return ret;
}

/* We use a special routine to initialize
    cygwin_init is called from the OS_INIT macro in main(). */

void cygwin_init(int argc, char ** argv, void * rup,
                 void * eup, void * egp, void * cup, void * cgp)
{
  int i;
  uid_t myuid, systemuid;
  gid_t mygid, adminsgid;
  struct passwd * pwp;
  char *cygenv, win32_path[MAX_PATH];
  SID(1, SystemSid, SECURITY_LOCAL_SYSTEM_RID);
  SID(2, AdminsSid, SECURITY_BUILTIN_DOMAIN_RID, DOMAIN_ALIAS_RID_ADMINS);
  DWORD priv_flags;

  myuid = getuid();
  mygid = getgid();
  cygwin_WinVersion = GetVersion();
  if ((cygenv = getenv("CYGWIN")) == NULL) cygenv = "";
  /* Produce some debugging on stderr,
     cannot yet use exim's debug functions.
     Exim does not use -c and ignores -n.
     Set lower priority for daemons */
  for (i = 1; i < argc; i++) {
    if (argv[i][0] == '-') {
      if (argv[i][1] == 'c') {
        argv[i][1] = 'n';  /* Replace -c by -n */
        cygwin_debug = TRUE;
        fprintf(stderr, "CYGWIN = \"%s\".", cygenv);
        cygwin_conv_to_win32_path("/", win32_path);
        fprintf(stderr, " Root / mapped to %s.\n", win32_path);
      }
      else if (argv[i][1] == 'b' && argv[i][2] == 'd')
        cygwin_setpriority();
    }
  }
  if (VERSION_IS_58M(cygwin_WinVersion)) {
    * (uid_t *) rup = myuid;  /* Pretend we are root */
    * (uid_t *) eup = myuid;  /* ... and exim */
    * (gid_t *) egp = mygid;
    return;
  }
  /* Nt/2000/XP
     We initially set the exim uid & gid to those of the "real exim",
       or to the root uid (SYSTEM) and exim gid (ADMINS),
     If privileged, we setuid to those.
     We always set the configure uid to the system uid.
     We always set the root uid to the real uid
       to avoid useless execs following forks.
     If not privileged and unable to chown,
       we set the exim uid to our uid.
     If unprivileged, we fake all subsequent setuid. */

  priv_flags = get_privileges ();
  privileged = !!(priv_flags & CREATE_BIT);

  /* Get the system and admins uid from their sids,
     or use the default values from the Makefile. */
  if ((systemuid = cygwin_internal(CW_GET_UID_FROM_SID, & SystemSid)) == -1)
    systemuid = * (uid_t *) eup;
  if ((adminsgid = cygwin_internal(CW_GET_GID_FROM_SID, & AdminsSid)) == -1)
    adminsgid = * (gid_t *) egp;

  if ((pwp = getpwnam("exim")) != NULL) {
    * (uid_t *) eup = pwp->pw_uid;  /* Set it according to passwd */
    * (gid_t *) egp = pwp->pw_gid;
  }
  else {
    * (uid_t *) eup = systemuid;
    * (gid_t *) egp = adminsgid;
  }

  /* Set the configuration uid and gid to the system uid and admins gid.
     Note that exim uid is also accepted as owner of exim.conf. */
  * (uid_t *) cup = systemuid;
  * (gid_t *) cgp = adminsgid;

  if (privileged) {             /* Can setuid */
    if (cygwin_setgid(* (gid_t *) egp) /* Setuid to exim */
        || cygwin_setuid(* (uid_t *) eup))
      privileged = -1;          /* Problem... Perhaps not in 544 */
  }

  /* Pretend we are root to avoid useless execs.
     We are limited by file access rights */
  * (uid_t *) rup = getuid ();

  /* If we have not setuid to exim and cannot chown,
     set the exim uid to our uid to avoid chown failures */
  if (privileged <= 0 && !(priv_flags & RESTORE_BIT))
    * (uid_t *) eup = * (uid_t *) rup;

  if (cygwin_debug) {
    fprintf(stderr, "Starting uid %ld, gid %ld, ntsec %lu, privileged %d.\n",
            myuid, mygid, cygwin_internal(CW_CHECK_NTSEC, NULL), privileged);
    fprintf(stderr, "root_uid %ld, exim_uid %ld, exim_gid %ld, config_uid %ld, config_gid %ld.\n",
            * (uid_t *) rup, * (uid_t *) eup, * (gid_t *) egp, * (uid_t *) cup, * (gid_t *) cgp);
  }
  return;
}

#ifndef OS_LOAD_AVERAGE /* Can be set on command line */
#define OS_LOAD_AVERAGE /* src/os.c need not provide it */

/*****************************************************************
 *
 Functions for average load measurements

 There are two methods, which work only on NT.

 The first one uses the HKEY_PERFORMANCE_DATA registry to
 get performance data. It is complex but well documented
 and works on all NT versions.

 The second one uses NtQuerySystemInformation.
 Its use is discouraged starting with WinXP.

 Until 4.43, the Cygwin port of exim was using the first
 method.

*****************************************************************/
#define PERF_METHOD2

/* Structure to compute the load average efficiently */
typedef struct {
  DWORD Lock;
  unsigned long long Time100ns;   /* Last measurement time */
  unsigned long long IdleCount;   /* Latest cumulative idle time */
  unsigned long long LastCounter; /* Last measurement counter */
  unsigned long long PerfFreq;    /* Perf counter frequency */
  int LastLoad;                   /* Last reported load, or -1 */
#ifdef PERF_METHOD1
  PPERF_DATA_BLOCK PerfData;      /* Pointer to a buffer to get the data */
  DWORD BufferSize;               /* Size of PerfData */
  LPSTR * NamesArray;             /* Temporary (malloc) buffer for index */
#endif
} cygwin_perf_t;

static struct {
   HANDLE handle;
   pid_t pid;
   cygwin_perf_t *perf;
} cygwin_load = {NULL, 0, NULL};

#ifdef PERF_METHOD1
/*************************************************************
 METHOD 1

 Obtaining statistics in Windows is done at a low level by
 calling registry functions, in particular the key
 HKEY_PERFORMANCE_DATA on NT and successors.
 Something equivalent exists on Win95, see Microsoft article
 HOWTO: Access the Performance Registry Under Windows 95 (KB 174631)
 but it is not implemented here.

 The list of objects to be polled is specified in the string
 passed to RegQueryValueEx in ReadStat() below.
 On NT, all objects are polled even if info about only one is
 required. This is fixed in Windows 2000. See articles
 INFO: Perflib Calling Close Procedure in Windows 2000 (KB 270127)
 INFO: Performance Data Changes Between Windows NT 4.0 and Windows
 2000 (KB 296523)

 It is unclear to me how the counters are primarily identified.
 Whether it's by name strings or by the offset of their strings
 as mapped in X:\Winnt\system32\perfc009.dat [or equivalently as
 reported by the registry functions in GetNameStrings( ) below].
 Microsoft documentation seems to say that both methods should
 work.

 In the interest of speed and language independence, the main
 code below relies on offsets. However if debug is enabled, the
 code verifies that the names of the corresponding strings are
 as expected.

*****************************************************************/

/* Object and counter indices and names */
#define PROCESSOR_OBJECT_INDEX 238
#define PROCESSOR_OBJECT_STRING "238"
#define PROCESSOR_OBJECT_NAME "Processor"
#define PROCESSOR_TIME_COUNTER 6
#define PROCESSOR_TIME_NAME "% Processor Time"

#define BYTEINCREMENT 800    /* Block to add to PerfData */

/*****************************************************************
 *
 Macros to navigate through the performance data.

 *****************************************************************/
#define FirstObject(PerfData)\
  ((PPERF_OBJECT_TYPE)((PBYTE)PerfData + PerfData->HeaderLength))
#define NextObject(PerfObj)\
  ((PPERF_OBJECT_TYPE)((PBYTE)PerfObj + PerfObj->TotalByteLength))
#define ObjectCounterBlock(PerfObj)\
  ((PPERF_COUNTER_BLOCK)(PBYTE)PerfObj + PerfObj->DefinitionLength )
#define FirstInstance(PerfObj )\
  ((PPERF_INSTANCE_DEFINITION)((PBYTE)PerfObj + PerfObj->DefinitionLength))
#define InstanceCounterBlock(PerfInst)\
  ((PPERF_COUNTER_BLOCK) ((PBYTE)PerfInst + PerfInst->ByteLength ))
#define NextInstance(PerfInst )\
  ((PPERF_INSTANCE_DEFINITION)((PBYTE)InstanceCounterBlock(PerfInst) + \
        InstanceCounterBlock(PerfInst)->ByteLength) )
#define FirstCounter(PerfObj)\
  ((PPERF_COUNTER_DEFINITION) ((PBYTE)PerfObj + PerfObj->HeaderLength))
#define NextCounter(PerfCntr)\
  ((PPERF_COUNTER_DEFINITION)((PBYTE)PerfCntr + PerfCntr->ByteLength))

/*****************************************************************
 *
 Load the counter and object names from the registry
 to cygwin_load.perf->NameStrings
 and index them in cygwin_load.perf->NamesArray

 NameStrings seems to be taken from the file
 X:\Winnt\system32\perfc009.dat

 This is used only for name verification during initialization,
 if DEBUG(D_load) is TRUE.

*****************************************************************/
static BOOL GetNameStrings( )
{
  HKEY hKeyPerflib;      // handle to registry key
  DWORD dwArraySize;     // size for array
  DWORD dwNamesSize;     // size for strings
  LPSTR lpCurrentString; // pointer for enumerating data strings
  DWORD dwCounter;       // current counter index
  LONG  res;

  /* Get the number of Counter items into dwArraySize. */
  if ((res = RegOpenKeyEx( HKEY_LOCAL_MACHINE,
                           "SOFTWARE\\Microsoft\\Windows NT\\CurrentVersion\\Perflib",
                           0,
                           KEY_QUERY_VALUE, /* KEY_READ, */
                           &hKeyPerflib))
      != ERROR_SUCCESS) {
    DEBUG(D_load) debug_printf("RegOpenKeyEx (1): error %ld (Windows)\n", res);
    return FALSE;
  }
  dwNamesSize = sizeof(dwArraySize); /* Temporary reuse */
  if ((res = RegQueryValueEx( hKeyPerflib,
                              "Last Counter",
                              NULL,
                              NULL,
                              (LPBYTE) &dwArraySize,
                              &dwNamesSize ))
      != ERROR_SUCCESS) {
    DEBUG(D_load) debug_printf("RegQueryValueEx (1): error %ld (Windows)\n", res);
    return FALSE;
  }
  RegCloseKey( hKeyPerflib );
  /* Open the key containing the counter and object names. */
  if ((res = RegOpenKeyEx( HKEY_LOCAL_MACHINE,
                           "SOFTWARE\\Microsoft\\Windows NT\\CurrentVersion\\Perflib\\009",
                           0,
                           KEY_READ,
                           &hKeyPerflib))
      != ERROR_SUCCESS) {
    DEBUG(D_load) debug_printf("RegOpenKeyEx (2): error %ld (Windows)\n", res);
    return FALSE;
  }
  /* Get the size of the Counter value in the key
     and then read the value in the tail of NamesArray */
  dwNamesSize = 0;
  lpCurrentString = NULL;
  while (1) {
    res = RegQueryValueEx( hKeyPerflib,
                           "Counter",
                           NULL,
                           NULL,
                           (unsigned char *) lpCurrentString,
                           &dwNamesSize);
    if ((res == ERROR_SUCCESS) && /* Bug (NT 4.0): SUCCESS was returned on first call */
        (cygwin_load.perf->NamesArray != NULL)) break;
    if ((res == ERROR_SUCCESS) || /* but cygwin_load.perf->NamesArrays == NULL */
        (res == ERROR_MORE_DATA)) {
      /* Allocate memory BOTH for the names array and for the counter and object names */
      if ((cygwin_load.perf->NamesArray =
           (LPSTR *) malloc( (dwArraySize + 1) * sizeof(LPSTR) + dwNamesSize * sizeof(CHAR)))
          != NULL) {
        /* Point to area for the counter and object names */
        lpCurrentString = (LPSTR) & cygwin_load.perf->NamesArray[dwArraySize + 1];
        continue;
      }
      DEBUG(D_load) debug_printf("Malloc: errno %d (%s)\n", errno, strerror(errno));
    }
    else { /* Serious error */
      DEBUG(D_load) debug_printf("RegQueryValueEx (2): error %ld (Windows)\n", res);
    }
    return FALSE;
  }
  RegCloseKey( hKeyPerflib );
  /* Index the names into an array. */
  while (*lpCurrentString) {
    dwCounter = atol( lpCurrentString );
    lpCurrentString += (lstrlen(lpCurrentString)+1);
    cygwin_load.perf->NamesArray[dwCounter] = lpCurrentString;
    lpCurrentString += (strlen(lpCurrentString)+1);
  }
  return TRUE;
}

/*****************************************************************
 *
 Find the value of the Processor Time counter

*****************************************************************/
static BOOL ReadTimeCtr(PPERF_OBJECT_TYPE PerfObj,
                        PPERF_COUNTER_DEFINITION CurCntr,
                        PPERF_COUNTER_BLOCK PtrToCntr,
                        unsigned long long * TimePtr){
  int j;
  /* Scan all counters. */
  for( j = 0; j < PerfObj->NumCounters; j++ ) {
    if (CurCntr->CounterNameTitleIndex == PROCESSOR_TIME_COUNTER) {
      /* Verify it is really the proc time counter */
      if ((CurCntr->CounterType != PERF_100NSEC_TIMER_INV) || /* Wrong type */
          ((cygwin_load.perf->NamesArray != NULL) &&                  /* Verify name */
           (strcmp(cygwin_load.perf->NamesArray[CurCntr->CounterNameTitleIndex],
                   PROCESSOR_TIME_NAME)))) {
        log_write(0, LOG_MAIN|LOG_PANIC,
                  "Incorrect Perf counter type or name %x %s",
                  (unsigned) CurCntr->CounterType,
                  cygwin_load.perf->NamesArray[CurCntr->CounterNameTitleIndex]);
        return FALSE;
      }
      *TimePtr += *(unsigned long long int *) ((PBYTE) PtrToCntr + CurCntr->CounterOffset);
      return TRUE; /* return TRUE as soon as we found the counter */
    }
    /* Get the next counter. */
    CurCntr = NextCounter( CurCntr );
  }
  return FALSE;
}

/*****************************************************************
 *
 ReadStat()
 Measures current Time100ns and IdleCount
 Return TRUE if success.

 *****************************************************************/
static BOOL ReadStat(unsigned long long int *Time100nsPtr,
                     unsigned long long int * IdleCountPtr)
{
  PPERF_OBJECT_TYPE PerfObj;
  PPERF_INSTANCE_DEFINITION PerfInst;
  PPERF_COUNTER_DEFINITION PerfCntr;
  PPERF_COUNTER_BLOCK PtrToCntr;
  DWORD i, k, res;

  /* Get the performance data for the Processor object
     There is no need to open a key.
     We may need to blindly increase the buffer size.
     BufferSize does not return info but may be changed */
  while (1) {
    DWORD BufferSize = cygwin_load.perf->BufferSize;
    res = RegQueryValueEx( HKEY_PERFORMANCE_DATA,
                           PROCESSOR_OBJECT_STRING,
                           NULL,
                           NULL,
                           (LPBYTE) cygwin_load.perf->PerfData,
                           &BufferSize );
    if (res == ERROR_SUCCESS) break;
    if (res == ERROR_MORE_DATA ) {
      /* Increment if necessary to get a buffer that is big enough. */
      cygwin_load.perf->BufferSize += BYTEINCREMENT;
      if ((cygwin_load.perf->PerfData =
           (PPERF_DATA_BLOCK) realloc( cygwin_load.perf->PerfData, cygwin_load.perf->BufferSize ))
          != NULL) continue;
      DEBUG(D_load) debug_printf("Malloc: errno %d (%s)\n", errno, strerror(errno));
    }
    else { /* Serious error */
      DEBUG(D_load) debug_printf("RegQueryValueEx (3): error %ld (Windows)\n", res);
    }
    return FALSE;
  }
  /* Initialize the counters */
  *Time100nsPtr = 0;
  *IdleCountPtr = 0;
  /* We should only have one object, but write general code just in case. */
  PerfObj = FirstObject( cygwin_load.perf->PerfData );
  for( i = 0; i < cygwin_load.perf->PerfData->NumObjectTypes; i++ ) {
    /* We are only interested in the processor object */
    if ( PerfObj->ObjectNameTitleIndex == PROCESSOR_OBJECT_INDEX) {
      /* Possibly verify it is really the Processor object. */
      if ((cygwin_load.perf->NamesArray != NULL) &&
          (strcmp(cygwin_load.perf->NamesArray[PerfObj->ObjectNameTitleIndex],
                  PROCESSOR_OBJECT_NAME))) {
        log_write(0, LOG_MAIN|LOG_PANIC,
                  "Incorrect Perf object name %s",
                  cygwin_load.perf->NamesArray[PerfObj->ObjectNameTitleIndex]);
        return FALSE;
      }
      /* Get the first counter */
      PerfCntr = FirstCounter( PerfObj );
      /* See if the object has instances.
         It should, but write general code. */
      if( PerfObj->NumInstances != PERF_NO_INSTANCES ) {
        PerfInst = FirstInstance( PerfObj );
        for( k = 0; k < PerfObj->NumInstances; k++ ) {
          /* There can be several processors.
             Accumulate both the Time100ns and the idle counter.
             Starting with Win2000 there is an instance named "_Total".
             Do not use it.     We only use instances with a single
             character in the name.
             If we examine the object names, we also look at the instance
             names and their lengths and issue reports */
          if ( cygwin_load.perf->NamesArray != NULL) {
            CHAR ascii[30]; /* The name is in unicode */
            wsprintf(ascii,"%.29lS",
                     (char *)((PBYTE)PerfInst + PerfInst->NameOffset));
            log_write(0, LOG_MAIN,
                      "Perf: Found processor instance \"%s\", length %d",
                      ascii, PerfInst->NameLength);
            if ((PerfInst->NameLength != 4) &&
                (strcmp(ascii, "_Total") != 0)) {
              log_write(0, LOG_MAIN|LOG_PANIC,
                        "Perf: WARNING: Unexpected processor instance name");
              return FALSE;
            }
          }
          if (PerfInst->NameLength == 4) {
            *Time100nsPtr += cygwin_load.perf->PerfData->PerfTime100nSec.QuadPart;
            PtrToCntr = InstanceCounterBlock(PerfInst);
            if (! ReadTimeCtr(PerfObj, PerfCntr, PtrToCntr, IdleCountPtr)) {
              return FALSE;
            }
          }
          PerfInst = NextInstance( PerfInst );
        }
        return (*Time100nsPtr != 0); /* Something was read */
      }
      else { /* No instance, just the counter data */
        *Time100nsPtr = cygwin_load.perf->PerfData->PerfTime100nSec.QuadPart;
        PtrToCntr = ObjectCounterBlock(PerfObj);
        return ReadTimeCtr(PerfObj, PerfCntr, PtrToCntr, IdleCountPtr);
      }
    }
    PerfObj = NextObject( PerfObj );
  }
  return FALSE; /* Did not find the Processor object */
}

#elif defined(PERF_METHOD2)

/*************************************************************
  METHOD 2

  Uses NtQuerySystemInformation.
  This requires definitions that are not part of
  standard include files.
*************************************************************/
#include <ntdef.h>

typedef enum _SYSTEM_INFORMATION_CLASS
{
  SystemBasicInformation = 0,
  SystemPerformanceInformation = 2,
  SystemTimeOfDayInformation = 3,
  SystemProcessesAndThreadsInformation = 5,
  SystemProcessorTimes = 8,
  SystemPagefileInformation = 18,
  /* There are a lot more of these... */
} SYSTEM_INFORMATION_CLASS;

typedef struct _SYSTEM_BASIC_INFORMATION
{
  ULONG Unknown;
  ULONG MaximumIncrement;
  ULONG PhysicalPageSize;
  ULONG NumberOfPhysicalPages;
  ULONG LowestPhysicalPage;
  ULONG HighestPhysicalPage;
  ULONG AllocationGranularity;
  ULONG LowestUserAddress;
  ULONG HighestUserAddress;
  ULONG ActiveProcessors;
  UCHAR NumberProcessors;
} SYSTEM_BASIC_INFORMATION, *PSYSTEM_BASIC_INFORMATION;

typedef struct __attribute__ ((aligned (8))) _SYSTEM_PROCESSOR_TIMES
{
  LARGE_INTEGER IdleTime;
  LARGE_INTEGER KernelTime;
  LARGE_INTEGER UserTime;
  LARGE_INTEGER DpcTime;
  LARGE_INTEGER InterruptTime;
  ULONG InterruptCount;
} SYSTEM_PROCESSOR_TIMES, *PSYSTEM_PROCESSOR_TIMES;

typedef NTSTATUS NTAPI (*NtQuerySystemInformation_t) (SYSTEM_INFORMATION_CLASS, PVOID, ULONG, PULONG);
typedef ULONG NTAPI (*RtlNtStatusToDosError_t) (NTSTATUS);

static NtQuerySystemInformation_t NtQuerySystemInformation;
static RtlNtStatusToDosError_t RtlNtStatusToDosError;

/*****************************************************************
 *
 LoadNtdll()
 Load special functions from the NTDLL
 Return TRUE if success.

 *****************************************************************/

static BOOL LoadNtdll()
{
  HINSTANCE hinstLib;

  if ((hinstLib = LoadLibrary("NTDLL.DLL"))
      && (NtQuerySystemInformation =
          (NtQuerySystemInformation_t) GetProcAddress(hinstLib,
                                                        "NtQuerySystemInformation"))
      && (RtlNtStatusToDosError =
          (RtlNtStatusToDosError_t) GetProcAddress(hinstLib,
                                                     "RtlNtStatusToDosError")))
    return TRUE;

  DEBUG(D_load)
    debug_printf("perf: load: %ld (Windows)\n", GetLastError());
  return FALSE;
}

/*****************************************************************
 *
 ReadStat()
 Measures current Time100ns and IdleCount
 Return TRUE if success.

 *****************************************************************/

static BOOL ReadStat(unsigned long long int *Time100nsPtr,
                     unsigned long long int *IdleCountPtr)
{
  NTSTATUS ret;
  SYSTEM_BASIC_INFORMATION sbi;
  PSYSTEM_PROCESSOR_TIMES spt;

  *Time100nsPtr = *IdleCountPtr = 0;

  if ((ret = NtQuerySystemInformation(SystemBasicInformation,
                                      (PVOID) &sbi, sizeof sbi, NULL))
      != STATUS_SUCCESS) {
    DEBUG(D_load)
      debug_printf("Perf: NtQuerySystemInformation: %lu (Windows)\n",
                   RtlNtStatusToDosError(ret));
  }
  else if (!(spt = (PSYSTEM_PROCESSOR_TIMES) alloca(sizeof(spt[0]) * sbi.NumberProcessors))) {
    DEBUG(D_load)
      debug_printf("Perf: alloca: errno %d (%s)\n", errno, strerror(errno));
  }
  else if ((ret = NtQuerySystemInformation(SystemProcessorTimes, (PVOID) spt,
                                           sizeof spt[0] * sbi.NumberProcessors, NULL))
           != STATUS_SUCCESS) {
    DEBUG(D_load)
      debug_printf("Perf: NtQuerySystemInformation: %lu (Windows)\n",
                   RtlNtStatusToDosError(ret));
  }
  else {
    int i;
    for (i = 0; i < sbi.NumberProcessors; i++) {
      *Time100nsPtr += spt[i].KernelTime.QuadPart;;
      *Time100nsPtr += spt[i].UserTime.QuadPart;
      *IdleCountPtr += spt[i].IdleTime.QuadPart;
    }
    return TRUE;
  }
  return FALSE;
}
#endif /* PERF_METHODX */

/*****************************************************************
 *
 InitLoadAvg()
 Initialize the cygwin_load.perf structure.
 and set cygwin_load.perf->Flag to TRUE if successful.
 This is called the first time os_getloadavg is called
 *****************************************************************/
static void InitLoadAvg(cygwin_perf_t *this)
{
  BOOL success = TRUE;

  /* Get perf frequency and counter */
  QueryPerformanceFrequency((LARGE_INTEGER *)& this->PerfFreq);
  QueryPerformanceCounter((LARGE_INTEGER *)& this->LastCounter);

#ifdef PERF_METHOD1
  DEBUG(D_load) {
    /* Get the name strings through the registry
       to verify that the object and counter numbers
       have the names we expect */
    success = GetNameStrings();
  }
#endif
  /* Get initial values for Time100ns and IdleCount */
  success = success
            && ReadStat( & this->Time100ns,
                         & this->IdleCount);
  /* If success, set the Load to 0, else to -1 */
  if (success) this->LastLoad = 0;
  else {
    log_write(0, LOG_MAIN, "Cannot obtain Load Average");
    this->LastLoad = -1;
  }
#ifdef PERF_METHOD1
  /* Free the buffer created for debug name verification */
  if (this->NamesArray != NULL) {
    free(this->NamesArray);
    this->NamesArray = NULL;
  }
#endif
}


/*****************************************************************
 *
 os_getloadavg()

 Return -1 if not available;
 Return the previous value if less than AVERAGING sec old.
 else return the processor load on a [0 - 1000] scale.

 The first time we are called we initialize the counts
 and return 0 or -1.
 The initial load cannot be measured as we use the processor 100%
*****************************************************************/
static SECURITY_ATTRIBUTES sa = {sizeof (SECURITY_ATTRIBUTES), NULL, TRUE};
#define AVERAGING 10

int os_getloadavg()
{
  unsigned long long Time100ns, IdleCount, CurrCounter;
  int value;
  pid_t newpid;

  /* New process.
     Reload the dlls and the file mapping */
  if ((newpid = getpid()) != cygwin_load.pid) {
    BOOL new;
    cygwin_load.pid = newpid;

#ifdef PERF_METHOD2
    if (!LoadNtdll()) {
      log_write(0, LOG_MAIN, "Cannot obtain Load Average");
      cygwin_load.perf = NULL;
      return -1;
    }
#endif

    if ((new = !cygwin_load.handle)) {
      cygwin_load.handle = CreateFileMapping (INVALID_HANDLE_VALUE, &sa, PAGE_READWRITE,
                                              0, sizeof(cygwin_perf_t), NULL);
      DEBUG(D_load)
        debug_printf("Perf: CreateFileMapping: handle %x\n", (unsigned) cygwin_load.handle);
    }
    cygwin_load.perf = (cygwin_perf_t *) MapViewOfFile (cygwin_load.handle,
                                                        FILE_MAP_READ | FILE_MAP_WRITE, 0, 0, 0);
    DEBUG(D_load)
      debug_printf("Perf: MapViewOfFile: addr %x\n", (unsigned) cygwin_load.perf);
    if (new && cygwin_load.perf)
      InitLoadAvg(cygwin_load.perf);
  }

  /* Check if initialized OK */
  if (!cygwin_load.perf || cygwin_load.perf->LastLoad < 0)
    return -1;

  /* If we cannot get the lock, we return 0.
     This is to prevent any lock-up possibility.
     Finding a lock busy is unlikely, and giving up only
     results in an immediate delivery .*/

  if (InterlockedCompareExchange(&cygwin_load.perf->Lock, 1, 0)) {
    DEBUG(D_load)
      debug_printf("Perf: Lock busy\n");
    return 0;
  }

    /* Get the current time (PerfCounter) */
    QueryPerformanceCounter((LARGE_INTEGER *)& CurrCounter);
    /* Calls closer than AVERAGING sec apart use the previous value */
  if (CurrCounter - cygwin_load.perf->LastCounter >
      AVERAGING * cygwin_load.perf->PerfFreq) {
      /* Get Time100ns and IdleCount */
      if (ReadStat( & Time100ns, & IdleCount)) { /* Success */
        /* Return processor load on 1000 scale */
      value = 1000 - ((1000 * (IdleCount - cygwin_load.perf->IdleCount)) /
                      (Time100ns - cygwin_load.perf->Time100ns));
      cygwin_load.perf->Time100ns = Time100ns;
      cygwin_load.perf->IdleCount = IdleCount;
      cygwin_load.perf->LastCounter = CurrCounter;
      cygwin_load.perf->LastLoad = value;
      DEBUG(D_load)
        debug_printf("Perf: New load average %d\n", value);
      }
      else { /* Something bad happened.
                Refuse to measure the load anymore
                but don't bother releasing the buffer */
        log_write(0, LOG_MAIN, "Cannot obtain Load Average");
      cygwin_load.perf->LastLoad = -1;
    }
  }
  else
  DEBUG(D_load)
      debug_printf("Perf: Old load average %d\n", cygwin_load.perf->LastLoad);
  cygwin_load.perf->Lock = 0;
  return cygwin_load.perf->LastLoad;
}
#endif /* OS_LOAD_AVERAGE */
#endif /* COMPILE_UTILITY */