newlib/winsup/cygwin/miscfuncs.cc

/* miscfuncs.cc: misc funcs that don't belong anywhere else

This file is part of Cygwin.

This software is a copyrighted work licensed under the terms of the
Cygwin license.  Please consult the file "CYGWIN_LICENSE" for
details. */

#include "winsup.h"
#include "miscfuncs.h"
#include <ntsecapi.h>
#include <sys/uio.h>
#include <sys/param.h>
#include "ntdll.h"
#include "path.h"
#include "fhandler.h"
#include "exception.h"
#include "tls_pbuf.h"

int __reg2
check_invalid_virtual_addr (const void *s, unsigned sz)
{
  MEMORY_BASIC_INFORMATION mbuf;
  const void *end;

  for (end = (char *) s + sz; s < end;
       s = (char *) mbuf.BaseAddress + mbuf.RegionSize)
    if (!VirtualQuery (s, &mbuf, sizeof mbuf))
      return EINVAL;
  return 0;
}

static char __attribute__ ((noinline))
dummytest (volatile char *p)
{
  return *p;
}

ssize_t
check_iovec (const struct iovec *iov, int iovcnt, bool forwrite)
{
  if (iovcnt < 0 || iovcnt > IOV_MAX)
    {
      set_errno (EINVAL);
      return -1;
    }

  __try
    {

      size_t tot = 0;

      while (iovcnt > 0)
	{
	  if (iov->iov_len > SSIZE_MAX || (tot += iov->iov_len) > SSIZE_MAX)
	    {
	      set_errno (EINVAL);
	      __leave;
	    }

	  volatile char *p = ((char *) iov->iov_base) + iov->iov_len - 1;
	  if (!iov->iov_len)
	    /* nothing to do */;
	  else if (!forwrite)
	    *p  = dummytest (p);
	  else
	    dummytest (p);

	  iov++;
	  iovcnt--;
	}

      if (tot <= SSIZE_MAX)
	return (ssize_t) tot;

      set_errno (EINVAL);
    }
  __except (EFAULT)
  __endtry
  return -1;
}

/* Try hard to schedule another thread.  
   Remember not to call this in a lock condition or you'll potentially
   suffer starvation.  */
void
yield ()
{
  /* MSDN implies that Sleep will force scheduling of other threads.
     Unlike SwitchToThread() the documentation does not mention other
     cpus so, presumably (hah!), this + using a lower priority will
     stall this thread temporarily and cause another to run.
     (stackoverflow and others seem to confirm that setting this thread
     to a lower priority and calling Sleep with a 0 paramenter will
     have this desired effect)

     CV 2017-03-08: Drop lowering the priority.  It leads to potential
		    starvation and it should not be necessary anymore
		    since Server 2003.  See the MSDN Sleep man page. */
  Sleep (0L);
}

/* Get a default value for the nice factor.  When changing these values,
   have a look into the below function nice_to_winprio.  The values must
   match the layout of the static "priority" array. */
int
winprio_to_nice (DWORD prio)
{
  switch (prio)
    {
      case REALTIME_PRIORITY_CLASS:
	return -20;
      case HIGH_PRIORITY_CLASS:
	return -16;
      case ABOVE_NORMAL_PRIORITY_CLASS:
	return -8;
      case NORMAL_PRIORITY_CLASS:
	return 0;
      case BELOW_NORMAL_PRIORITY_CLASS:
	return 8;
      case IDLE_PRIORITY_CLASS:
	return 16;
    }
  return 0;
}

/* Get a Win32 priority matching the incoming nice factor.  The incoming
   nice is limited to the interval [-NZERO,NZERO-1]. */
DWORD
nice_to_winprio (int &nice)
{
  static const DWORD priority[] =
    {
      REALTIME_PRIORITY_CLASS,		/*  0 */
      HIGH_PRIORITY_CLASS,		/*  1 */
      HIGH_PRIORITY_CLASS,
      HIGH_PRIORITY_CLASS,
      HIGH_PRIORITY_CLASS,
      HIGH_PRIORITY_CLASS,
      HIGH_PRIORITY_CLASS,
      HIGH_PRIORITY_CLASS,		/*  7 */
      ABOVE_NORMAL_PRIORITY_CLASS,	/*  8 */
      ABOVE_NORMAL_PRIORITY_CLASS,
      ABOVE_NORMAL_PRIORITY_CLASS,
      ABOVE_NORMAL_PRIORITY_CLASS,
      ABOVE_NORMAL_PRIORITY_CLASS,
      ABOVE_NORMAL_PRIORITY_CLASS,
      ABOVE_NORMAL_PRIORITY_CLASS,
      ABOVE_NORMAL_PRIORITY_CLASS,	/* 15 */
      NORMAL_PRIORITY_CLASS,		/* 16 */
      NORMAL_PRIORITY_CLASS,
      NORMAL_PRIORITY_CLASS,
      NORMAL_PRIORITY_CLASS,
      NORMAL_PRIORITY_CLASS,
      NORMAL_PRIORITY_CLASS,
      NORMAL_PRIORITY_CLASS,
      NORMAL_PRIORITY_CLASS,		/* 23 */
      BELOW_NORMAL_PRIORITY_CLASS,	/* 24 */
      BELOW_NORMAL_PRIORITY_CLASS,
      BELOW_NORMAL_PRIORITY_CLASS,
      BELOW_NORMAL_PRIORITY_CLASS,
      BELOW_NORMAL_PRIORITY_CLASS,
      BELOW_NORMAL_PRIORITY_CLASS,
      BELOW_NORMAL_PRIORITY_CLASS,
      BELOW_NORMAL_PRIORITY_CLASS,	/* 31 */
      IDLE_PRIORITY_CLASS,		/* 32 */
      IDLE_PRIORITY_CLASS,
      IDLE_PRIORITY_CLASS,
      IDLE_PRIORITY_CLASS,
      IDLE_PRIORITY_CLASS,
      IDLE_PRIORITY_CLASS,
      IDLE_PRIORITY_CLASS,
      IDLE_PRIORITY_CLASS		/* 39 */
    };
  if (nice < -NZERO)
    nice = -NZERO;
  else if (nice > NZERO - 1)
    nice = NZERO - 1;
  DWORD prio = priority[nice + NZERO];
  return prio;
}

/* Minimal overlapped pipe I/O implementation for signal and commune stuff. */

BOOL WINAPI
CreatePipeOverlapped (PHANDLE hr, PHANDLE hw, LPSECURITY_ATTRIBUTES sa)
{
  int ret = fhandler_pipe::create (sa, hr, hw, 0, NULL,
				   FILE_FLAG_OVERLAPPED);
  if (ret)
    SetLastError (ret);
  return ret == 0;
}

BOOL WINAPI
ReadPipeOverlapped (HANDLE h, PVOID buf, DWORD len, LPDWORD ret_len,
		    DWORD timeout)
{
  OVERLAPPED ov;
  BOOL ret;

  memset (&ov, 0, sizeof ov);
  ov.hEvent = CreateEvent (NULL, TRUE, FALSE, NULL);
  ret = ReadFile (h, buf, len, NULL, &ov);
  if (ret || GetLastError () == ERROR_IO_PENDING)
    {
      if (!ret && WaitForSingleObject (ov.hEvent, timeout) != WAIT_OBJECT_0)
	CancelIo (h);
      ret = GetOverlappedResult (h, &ov, ret_len, FALSE);
    }
  CloseHandle (ov.hEvent);
  return ret;
}

BOOL WINAPI
WritePipeOverlapped (HANDLE h, LPCVOID buf, DWORD len, LPDWORD ret_len,
		     DWORD timeout)
{
  OVERLAPPED ov;
  BOOL ret;

  memset (&ov, 0, sizeof ov);
  ov.hEvent = CreateEvent (NULL, TRUE, FALSE, NULL);
  ret = WriteFile (h, buf, len, NULL, &ov);
  if (ret || GetLastError () == ERROR_IO_PENDING)
    {
      if (!ret && WaitForSingleObject (ov.hEvent, timeout) != WAIT_OBJECT_0)
	CancelIo (h);
      ret = GetOverlappedResult (h, &ov, ret_len, FALSE);
    }
  CloseHandle (ov.hEvent);
  return ret;
}

bool
NT_readline::init (POBJECT_ATTRIBUTES attr, PCHAR in_buf, ULONG in_buflen)
{
  NTSTATUS status;
  IO_STATUS_BLOCK io;

  status = NtOpenFile (&fh, SYNCHRONIZE | FILE_READ_DATA, attr, &io,
                       FILE_SHARE_VALID_FLAGS,
                       FILE_SYNCHRONOUS_IO_NONALERT
                       | FILE_OPEN_FOR_BACKUP_INTENT);
  if (!NT_SUCCESS (status))
    {
      paranoid_printf ("NtOpenFile(%S) failed, status %y",
		       attr->ObjectName, status);
      return false;
    }
  buf = in_buf;
  buflen = in_buflen;
  got = end = buf;
  len = 0;
  line = 1;
  return true;
}

PCHAR 
NT_readline::gets ()
{
  IO_STATUS_BLOCK io;

  while (true)
    {
      /* len == 0 indicates we have to read from the file. */
      if (!len)
	{
	  if (!NT_SUCCESS (NtReadFile (fh, NULL, NULL, NULL, &io, got,
				       (buflen - 2) - (got - buf), NULL, NULL)))
	    return NULL;
	  len = io.Information;
	  /* Set end marker. */
	  got[len] = got[len + 1] = '\0';
	  /* Set len to the absolute len of bytes in buf. */
	  len += got - buf;
	  /* Reset got to start reading at the start of the buffer again. */
	  got = end = buf;
	}
      else
	{
	  got = end + 1;
	  ++line;
	}
      /* Still some valid full line? */
      if (got < buf + len)
	{
	  if ((end = strchr (got, '\n')))
	    {
	      end[end[-1] == '\r' ? -1 : 0] = '\0';
	      return got;
	    }
	  /* Last line missing a \n at EOF? */
	  if (len < buflen - 2)
	    {
	      len = 0;
	      return got;
	    }
	}
      /* We have to read once more.  Move remaining bytes to the start of
         the buffer and reposition got so that it points to the end of
         the remaining bytes. */
      len = buf + len - got;
      memmove (buf, got, len);
      got = buf + len;
      buf[len] = buf[len + 1] = '\0';
      len = 0;
    }
}

/* Return an address from the import jmp table of main program.  */
void * __reg1
__import_address (void *imp)
{
  __try
    {
      if (*((uint16_t *) imp) == 0x25ff)
	{
	  const char *ptr = (const char *) imp;
#ifdef __x86_64__
	  const uintptr_t *jmpto = (uintptr_t *)
				   (ptr + 6 + *(int32_t *)(ptr + 2));
#else
	  const uintptr_t *jmpto = (uintptr_t *) *((uintptr_t *) (ptr + 2));
#endif
	  return (void *) *jmpto;
	}
    }
  __except (NO_ERROR) {}
  __endtry
  return NULL;
}

/* Helper function to generate the correct caller address.  For external
   calls, the return address on the stack is _sigbe.  In that case the
   actual caller return address is on the cygtls stack.  Use this function
   via the macro caller_return_address. */
extern "C" void _sigbe ();

void *
__caller_return_address (void *builtin_ret_addr)
{
  return builtin_ret_addr == &_sigbe
	 ? (void *) _my_tls.retaddr () : builtin_ret_addr;
}

/* CygwinCreateThread.

   Replacement function for CreateThread.  What we do here is to remove
   parameters we don't use and instead to add parameters we need to make
   the function pthreads compatible. */

struct pthread_wrapper_arg
{
  LPTHREAD_START_ROUTINE func;
  PVOID arg;
  PBYTE stackaddr;
  PBYTE stackbase;
  PBYTE stacklimit;
  ULONG guardsize;
};

DWORD WINAPI
pthread_wrapper (PVOID arg)
{
  /* Just plain paranoia. */
  if (!arg)
    return ERROR_INVALID_PARAMETER;

  /* The process is now threaded.  Note for later usage by arc4random. */
  __isthreaded = 1;

  /* Fetch thread wrapper info and free from cygheap. */
  pthread_wrapper_arg wrapper_arg = *(pthread_wrapper_arg *) arg;
  cfree (arg);

  /* Set stack values in TEB */
  PTEB teb = NtCurrentTeb ();
  teb->Tib.StackBase = wrapper_arg.stackbase;
  teb->Tib.StackLimit = wrapper_arg.stacklimit ?: wrapper_arg.stackaddr;
  /* Set DeallocationStack value.  If we have an application-provided stack,
     we set DeallocationStack to NULL, so NtTerminateThread does not deallocate
     any stack.  If we created the stack in CygwinCreateThread, we set
     DeallocationStack to the stackaddr of our own stack, so it's automatically
     deallocated when the thread is terminated. */
  PBYTE dealloc_addr = (PBYTE) teb->DeallocationStack;
  teb->DeallocationStack = wrapper_arg.stacklimit ? wrapper_arg.stackaddr
						  : NULL;
  /* Store the OS-provided DeallocationStack address in wrapper_arg.stackaddr.
     The below assembler code will release the OS stack after switching to our
     new stack. */
  wrapper_arg.stackaddr = dealloc_addr;
  /* Set thread stack guarantee matching the guardsize.
     Note that the guardsize is one page bigger than the guarantee. */
  if (wrapper_arg.guardsize > wincap.def_guard_page_size ())
    {
      wrapper_arg.guardsize -= wincap.page_size ();
      SetThreadStackGuarantee (&wrapper_arg.guardsize);
    }
  /* Initialize new _cygtls. */
  _my_tls.init_thread (wrapper_arg.stackbase - CYGTLS_PADSIZE,
		       (DWORD (*)(void*, void*)) wrapper_arg.func);
#ifdef __i386__
  /* Copy exception list over to new stack.  I'm not quite sure how the
     exception list is extended by Windows itself.  What's clear is that it
     always grows downwards and that it starts right at the stackbase.
     Therefore we first count the number of exception records and place
     the copy at the stackbase, too, so there's still a lot of room to
     extend the list up to where our _cygtls region starts. */
  _exception_list *old_start = (_exception_list *) teb->Tib.ExceptionList;
  unsigned count = 0;
  teb->Tib.ExceptionList = NULL;
  for (_exception_list *e_ptr = old_start;
       e_ptr && e_ptr != (_exception_list *) -1;
       e_ptr = e_ptr->prev)
    ++count;
  if (count)
    {
      _exception_list *new_start = (_exception_list *) wrapper_arg.stackbase
						       - count;
      teb->Tib.ExceptionList = (struct _EXCEPTION_REGISTRATION_RECORD *)
			       new_start;
      while (true)
	{
	  new_start->handler = old_start->handler;
	  if (old_start->prev == (_exception_list *) -1)
	    {
	      new_start->prev = (_exception_list *) -1;
	      break;
	    }
	  new_start->prev = new_start + 1;
	  new_start = new_start->prev;
	  old_start = old_start->prev;
	}
    }
#endif
#ifdef __x86_64__
  __asm__ ("\n\
	   leaq  %[WRAPPER_ARG], %%rbx	# Load &wrapper_arg into rbx	\n\
	   movq  (%%rbx), %%r12		# Load thread func into r12	\n\
	   movq  8(%%rbx), %%r13	# Load thread arg into r13	\n\
	   movq  16(%%rbx), %%rcx	# Load stackaddr into rcx	\n\
	   movq  24(%%rbx), %%rsp	# Load stackbase into rsp	\n\
	   subq  %[CYGTLS], %%rsp	# Subtract CYGTLS_PADSIZE	\n\
	   				# (here we are 16 bytes aligned)\n\
	   subq  $32, %%rsp		# Subtract another 32 bytes	\n\
	   				# (shadow space for arg regs)	\n\
	   xorq  %%rbp, %%rbp		# Set rbp to 0			\n\
	   # We moved to the new stack.					\n\
	   # Now it's safe to release the OS stack.			\n\
	   movl  $0x8000, %%r8d		# dwFreeType: MEM_RELEASE	\n\
	   xorl  %%edx, %%edx		# dwSize:     0			\n\
	   # dwAddress is already in the correct arg register rcx	\n\
	   call  VirtualFree						\n\
	   # All set.  We can copy the thread arg from the safe		\n\
	   # register r13 and then just call the function.		\n\
	   movq  %%r13, %%rcx		# Move thread arg to 1st arg reg\n\
	   call  *%%r12			# Call thread func		\n"
	   : : [WRAPPER_ARG] "o" (wrapper_arg),
	       [CYGTLS] "i" (CYGTLS_PADSIZE));
#else
  __asm__ ("\n\
	   leal  %[WRAPPER_ARG], %%ebx	# Load &wrapper_arg into ebx	\n\
	   movl  (%%ebx), %%eax		# Load thread func into eax	\n\
	   movl  4(%%ebx), %%ecx	# Load thread arg into ecx	\n\
	   movl  8(%%ebx), %%edx	# Load stackaddr into edx	\n\
	   movl  12(%%ebx), %%ebx	# Load stackbase into ebx	\n\
	   subl  %[CYGTLS], %%ebx	# Subtract CYGTLS_PADSIZE	\n\
	   subl  $4, %%ebx		# Subtract another 4 bytes	\n\
	   movl  %%ebx, %%esp		# Set esp			\n\
	   xorl  %%ebp, %%ebp		# Set ebp to 0			\n\
	   # Make gcc 3.x happy and align the stack so that it is	\n\
	   # 16 byte aligned right before the final call opcode.	\n\
	   andl  $-16, %%esp		# 16 byte align			\n\
	   addl  $-12, %%esp		# 12 bytes + 4 byte arg = 16	\n\
	   # Now we moved to the new stack.  Save thread func address	\n\
	   # and thread arg on new stack				\n\
	   pushl %%ecx			# Push thread arg onto stack	\n\
	   pushl %%eax			# Push thread func onto stack	\n\
	   # Now it's safe to release the OS stack.			\n\
	   pushl $0x8000		# dwFreeType: MEM_RELEASE	\n\
	   pushl $0x0			# dwSize:     0			\n\
	   pushl %%edx			# lpAddress:  stackaddr		\n\
	   call _VirtualFree@12		# Shoot				\n\
	   # All set.  We can pop the thread function address from	\n\
	   # the stack and call it.  The thread arg is still on the	\n\
	   # stack in the expected spot.				\n\
	   popl  %%eax			# Pop thread_func address	\n\
	   call  *%%eax			# Call thread func		\n"
	   : : [WRAPPER_ARG] "o" (wrapper_arg),
	       [CYGTLS] "i" (CYGTLS_PADSIZE));
#endif
  /* pthread::thread_init_wrapper calls pthread::exit, which
     in turn calls ExitThread, so we should never arrive here. */
  api_fatal ("Dumb thinko in pthread handling.  Whip the developer.");
}

#ifdef __x86_64__
/* The memory region used for thread stacks */
#define THREAD_STORAGE_LOW	0x080000000L
#define THREAD_STORAGE_HIGH	0x100000000L
/* We provide the stacks always in 1 Megabyte slots */
#define THREAD_STACK_SLOT	0x100000L	/* 1 Meg */
/* Maximum stack size returned from the pool. */
#define THREAD_STACK_MAX	0x20000000L	/* 512 Megs */

class thread_allocator
{
  UINT_PTR current;
public:
  thread_allocator () : current (THREAD_STORAGE_HIGH) {}
  PVOID alloc (SIZE_T size)
  {
    SIZE_T real_size = roundup2 (size, THREAD_STACK_SLOT);
    BOOL overflow = FALSE;
    PVOID real_stackaddr = NULL;

    /* If an application requests a monster stack, we fulfill this request
       from outside of our pool, top down. */
    if (real_size > THREAD_STACK_MAX)
      return VirtualAlloc (NULL, real_size, MEM_RESERVE | MEM_TOP_DOWN,
			   PAGE_READWRITE);
    /* Simple round-robin.  Keep looping until VirtualAlloc succeeded, or
       until we overflowed and hit the current address. */
    for (UINT_PTR addr = current - real_size;
	 !real_stackaddr && (!overflow || addr >= current);
	 addr -= THREAD_STACK_SLOT)
      {
	if (addr < THREAD_STORAGE_LOW)
	  {
	    addr = THREAD_STORAGE_HIGH - real_size;
	    overflow = TRUE;
	  }
	real_stackaddr = VirtualAlloc ((PVOID) addr, real_size,
				       MEM_RESERVE, PAGE_READWRITE);
	if (!real_stackaddr)
	  {
	    /* So we couldn't grab this space.  Let's check the state.
	       If this area is free, simply try the next lower 1 Meg slot.
	       Otherwise, shift the next try down to the AllocationBase
	       of the current address, minus the requested slot size.
	       Add THREAD_STACK_SLOT since that's subtracted in the next
	       run of the loop anyway. */
	    MEMORY_BASIC_INFORMATION mbi;
	    VirtualQuery ((PVOID) addr, &mbi, sizeof mbi);
	    if (mbi.State != MEM_FREE)
	      addr = (UINT_PTR) mbi.AllocationBase - real_size
						    + THREAD_STACK_SLOT;
	  }
      }
    /* If we got an address, remember it for the next allocation attempt. */
    if (real_stackaddr)
      current = (UINT_PTR) real_stackaddr;
    else
      set_errno (EAGAIN);
    return real_stackaddr;
  }
};

thread_allocator thr_alloc NO_COPY;

/* Just set up a system-like main thread stack from the pthread stack area
   maintained by the thr_alloc class.  See the description in the x86_64-only
   code in _dll_crt0 to understand why we have to do this. */
PVOID
create_new_main_thread_stack (PVOID &allocationbase, SIZE_T parent_commitsize)
{
  PIMAGE_DOS_HEADER dosheader;
  PIMAGE_NT_HEADERS ntheader;
  SIZE_T stacksize;
  ULONG guardsize;
  SIZE_T commitsize;
  PBYTE stacklimit;

  dosheader = (PIMAGE_DOS_HEADER) GetModuleHandle (NULL);
  ntheader = (PIMAGE_NT_HEADERS)
	     ((PBYTE) dosheader + dosheader->e_lfanew);
  stacksize = ntheader->OptionalHeader.SizeOfStackReserve;
  stacksize = roundup2 (stacksize, wincap.allocation_granularity ());

  allocationbase
	= thr_alloc.alloc (ntheader->OptionalHeader.SizeOfStackReserve);
  guardsize = wincap.def_guard_page_size ();
  if (parent_commitsize)
    commitsize = (SIZE_T) parent_commitsize;
  else
    commitsize = ntheader->OptionalHeader.SizeOfStackCommit;
  commitsize = roundup2 (commitsize, wincap.page_size ());
  if (commitsize > stacksize - guardsize - wincap.page_size ())
    commitsize = stacksize - guardsize - wincap.page_size ();
  stacklimit = (PBYTE) allocationbase + stacksize - commitsize - guardsize;
  /* Setup guardpage. */
  if (!VirtualAlloc (stacklimit, guardsize,
		     MEM_COMMIT, PAGE_READWRITE | PAGE_GUARD))
    return NULL;
  /* Setup committed region. */
  stacklimit += guardsize;
  if (!VirtualAlloc (stacklimit, commitsize, MEM_COMMIT, PAGE_READWRITE))
    return NULL;
  NtCurrentTeb()->Tib.StackBase = ((PBYTE) allocationbase + stacksize);
  NtCurrentTeb()->Tib.StackLimit = stacklimit;
  return ((PBYTE) allocationbase + stacksize - 16);
}
#endif

HANDLE WINAPI
CygwinCreateThread (LPTHREAD_START_ROUTINE thread_func, PVOID thread_arg,
		    PVOID stackaddr, ULONG stacksize, ULONG guardsize,
		    DWORD creation_flags, LPDWORD thread_id)
{
  PVOID real_stackaddr = NULL;
  ULONG real_stacksize = 0;
  ULONG real_guardsize = 0;
  pthread_wrapper_arg *wrapper_arg;
  HANDLE thread = NULL;

  wrapper_arg = (pthread_wrapper_arg *) ccalloc (HEAP_STR, 1,
						sizeof *wrapper_arg);
  if (!wrapper_arg)
    {
      SetLastError (ERROR_OUTOFMEMORY);
      return NULL;
    }
  wrapper_arg->func = thread_func;
  wrapper_arg->arg = thread_arg;

  if (stackaddr)
    {
      /* If the application provided the stack, just use it.  There won't
	 be any stack overflow handling! */
      wrapper_arg->stackaddr = (PBYTE) stackaddr;
      wrapper_arg->stackbase = (PBYTE) stackaddr + stacksize;
    }
  else
    {
      PBYTE real_stacklimit;

      /* If not, we have to create the stack here. */
      real_stacksize = roundup2 (stacksize, wincap.page_size ());
      real_guardsize = roundup2 (guardsize, wincap.page_size ());
      /* Add the guardsize to the stacksize */
      real_stacksize += real_guardsize;
      /* Take dead zone page into account, which always stays uncommited. */
      real_stacksize += wincap.page_size ();
      /* Now roundup the result to the next allocation boundary. */
      real_stacksize = roundup2 (real_stacksize,
				 wincap.allocation_granularity ());
      /* Reserve stack. */
#ifdef __x86_64__
      real_stackaddr = thr_alloc.alloc (real_stacksize);
#else
      /* FIXME? If the TOP_DOWN method tends to collide too much with
	 other stuff, we should provide our own mechanism to find a
	 suitable place for the stack.  Top down from the start of
	 the Cygwin DLL comes to mind. */
      real_stackaddr = VirtualAlloc (NULL, real_stacksize,
				     MEM_RESERVE | MEM_TOP_DOWN,
				     PAGE_READWRITE);
#endif
      if (!real_stackaddr)
	return NULL;
      /* Set up committed region.  We set up the stack like the OS does,
	 with a reserved region, the guard pages, and a commited region.
	 We commit the stack commit size from the executable header, but
	 at least PTHREAD_STACK_MIN (64K). */
      static ULONG exe_commitsize;

      if (!exe_commitsize)
	{
	  PIMAGE_DOS_HEADER dosheader;
	  PIMAGE_NT_HEADERS ntheader;

	  dosheader = (PIMAGE_DOS_HEADER) GetModuleHandle (NULL);
	  ntheader = (PIMAGE_NT_HEADERS)
		     ((PBYTE) dosheader + dosheader->e_lfanew);
	  exe_commitsize = ntheader->OptionalHeader.SizeOfStackCommit;
	  exe_commitsize = roundup2 (exe_commitsize, wincap.page_size ());
	}
      ULONG commitsize = exe_commitsize;
      if (commitsize > real_stacksize - real_guardsize - wincap.page_size ())
	commitsize = real_stacksize - real_guardsize - wincap.page_size ();
      else if (commitsize < PTHREAD_STACK_MIN)
	commitsize = PTHREAD_STACK_MIN;
      real_stacklimit = (PBYTE) real_stackaddr + real_stacksize
			- commitsize - real_guardsize;
      if (!VirtualAlloc (real_stacklimit, real_guardsize, MEM_COMMIT,
			 PAGE_READWRITE | PAGE_GUARD))
	goto err;
      real_stacklimit += real_guardsize;
      if (!VirtualAlloc (real_stacklimit, commitsize, MEM_COMMIT,
			 PAGE_READWRITE))
	goto err;

      wrapper_arg->stackaddr = (PBYTE) real_stackaddr;
      wrapper_arg->stackbase = (PBYTE) real_stackaddr + real_stacksize;
      wrapper_arg->stacklimit = real_stacklimit;
      wrapper_arg->guardsize = real_guardsize;
    }
  /* Use the STACK_SIZE_PARAM_IS_A_RESERVATION parameter so only the
     minimum size for a thread stack is reserved by the OS.  Note that we
     reserve a 256K stack, not 64K, otherwise the thread creation might
     crash the process due to a stack overflow. */
  thread = CreateThread (&sec_none_nih, 4 * PTHREAD_STACK_MIN,
			 pthread_wrapper, wrapper_arg,
			 creation_flags | STACK_SIZE_PARAM_IS_A_RESERVATION,
			 thread_id);

err:
  if (!thread && real_stackaddr)
    {
      /* Don't report the wrong error even though VirtualFree is very unlikely
	 to fail. */
      DWORD err = GetLastError ();
      VirtualFree (real_stackaddr, 0, MEM_RELEASE);
      SetLastError (err);
    }
  return thread;
}

#ifdef __x86_64__
/* These functions are almost verbatim FreeBSD code (even if the header of
   one file mentiones NetBSD), just wrapped in the minimum required code to
   make them work with the MS AMD64 ABI.
   See FreeBSD src/lib/libc/amd64/string/memset.S
   and FreeBSD src/lib/libc/amd64/string/bcopy.S */

asm ("								\n\
/*									\n\
 * Written by J.T. Conklin <jtc@NetBSD.org>.				\n\
 * Public domain.							\n\
 * Adapted for NetBSD/x86_64 by						\n\
 * Frank van der Linden <fvdl@wasabisystems.com>			\n\
 */									\n\
									\n\
	.globl	memset							\n\
	.seh_proc memset						\n\
memset:									\n\
	movq	%rsi,8(%rsp)						\n\
	movq	%rdi,16(%rsp)						\n\
	.seh_endprologue						\n\
	movq	%rcx,%rdi						\n\
	movq	%rdx,%rsi						\n\
	movq	%r8,%rdx						\n\
									\n\
	movq    %rsi,%rax						\n\
	andq    $0xff,%rax						\n\
	movq    %rdx,%rcx						\n\
	movq    %rdi,%r11						\n\
									\n\
	cld			/* set fill direction forward */	\n\
									\n\
	/* if the string is too short, it's really not worth the	\n\
	 * overhead of aligning to word boundries, etc.  So we jump to	\n\
	 * a plain unaligned set. */					\n\
	cmpq    $0x0f,%rcx						\n\
	jle     L1							\n\
									\n\
	movb    %al,%ah		/* copy char to all bytes in word */\n\
	movl    %eax,%edx						\n\
	sall    $16,%eax						\n\
	orl     %edx,%eax						\n\
									\n\
	movl    %eax,%edx						\n\
	salq    $32,%rax						\n\
	orq     %rdx,%rax						\n\
									\n\
	movq    %rdi,%rdx	/* compute misalignment */		\n\
	negq    %rdx							\n\
	andq    $7,%rdx							\n\
	movq    %rcx,%r8						\n\
	subq    %rdx,%r8						\n\
									\n\
	movq    %rdx,%rcx	/* set until word aligned */		\n\
	rep								\n\
	stosb								\n\
									\n\
	movq    %r8,%rcx						\n\
	shrq    $3,%rcx		/* set by words */			\n\
	rep								\n\
	stosq								\n\
									\n\
	movq    %r8,%rcx	/* set remainder by bytes */		\n\
	andq    $7,%rcx							\n\
L1:     rep								\n\
	stosb								\n\
	movq    %r11,%rax						\n\
									\n\
	movq	8(%rsp),%rsi						\n\
	movq	16(%rsp),%rdi						\n\
	ret								\n\
	.seh_endproc							\n\
");

asm ("								\n\
/*-									\n\
 * Copyright (c) 1990 The Regents of the University of California.	\n\
 * All rights reserved.							\n\
 *									\n\
 * This code is derived from locore.s.					\n\
 *									\n\
 * Redistribution and use in source and binary forms, with or without	\n\
 * modification, are permitted provided that the following conditions	\n\
 * are met:								\n\
 * 1. Redistributions of source code must retain the above copyright	\n\
 *    notice, this list of conditions and the following disclaimer.	\n\
 * 2. Redistributions in binary form must reproduce the above copyright	\n\
 *    notice, this list of conditions and the following disclaimer in	\n\
 *    the documentation and/or other materials provided with the	\n\
 *    distribution.							\n\
 * 3. Neither the name of the University nor the names of its		\n\
 *    contributors may be used to endorse or promote products derived	\n\
 *    from this software without specific prior written permission.	\n\
 *									\n\
 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS''	\n\
 * AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO,\n\
 * THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A		\n\
 * PARTICULAR PURPOSE ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR	\n\
 * CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL,\n\
 * EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,	\n\
 * PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR	\n\
 * PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY	\n\
 * OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT		\n\
 * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE	\n\
 * USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH	\n\
 * DAMAGE.								\n\
 */									\n\
									\n\
	.seh_proc _memcpy						\n\
_memcpy:								\n\
	movq	%rsi,8(%rsp)						\n\
	movq	%rdi,16(%rsp)						\n\
	.seh_endprologue						\n\
	movq	%rcx,%rdi						\n\
	movq	%rdx,%rsi						\n\
	movq	%r8,%rdx						\n\
									\n\
	movq    %rdx,%rcx						\n\
	movq    %rdi,%r8						\n\
	subq    %rsi,%r8						\n\
	cmpq    %rcx,%r8	/* overlapping? */			\n\
	jb      1f							\n\
	cld                     /* nope, copy forwards. */		\n\
	shrq    $3,%rcx		/* copy by words */			\n\
	rep movsq							\n\
	movq    %rdx,%rcx						\n\
	andq    $7,%rcx		/* any bytes left? */			\n\
	rep movsb							\n\
	jmp	2f							\n\
1:									\n\
	addq    %rcx,%rdi	/* copy backwards. */			\n\
	addq    %rcx,%rsi						\n\
	std								\n\
	andq    $7,%rcx		/* any fractional bytes? */		\n\
	decq    %rdi							\n\
	decq    %rsi							\n\
	rep movsb							\n\
	movq    %rdx,%rcx	/* copy remainder by words */		\n\
	shrq    $3,%rcx							\n\
	subq    $7,%rsi							\n\
	subq    $7,%rdi							\n\
	rep movsq							\n\
	cld								\n\
2:									\n\
	movq	8(%rsp),%rsi						\n\
	movq	16(%rsp),%rdi						\n\
	ret								\n\
	.seh_endproc							\n\
									\n\
	.globl  memmove							\n\
	.seh_proc memmove						\n\
memmove:								\n\
	.seh_endprologue						\n\
	movq	%rcx,%rax	/* return dst */			\n\
	jmp	_memcpy							\n\
	.seh_endproc							\n\
									\n\
	.globl  memcpy							\n\
	.seh_proc memcpy						\n\
memcpy:									\n\
	.seh_endprologue						\n\
	movq	%rcx,%rax	/* return dst */			\n\
	jmp	_memcpy							\n\
	.seh_endproc							\n\
									\n\
	.globl  mempcpy							\n\
	.seh_proc mempcpy						\n\
mempcpy:								\n\
	.seh_endprologue						\n\
	movq	%rcx,%rax	/* return dst  */			\n\
	addq	%r8,%rax	/*         + n */			\n\
	jmp	_memcpy							\n\
	.seh_endproc							\n\
									\n\
	.globl  wmemmove						\n\
	.seh_proc wmemmove						\n\
wmemmove:								\n\
	.seh_endprologue						\n\
	shlq	$1,%r8		/* cnt * sizeof (wchar_t) */		\n\
	movq	%rcx,%rax	/* return dst */			\n\
	jmp	_memcpy							\n\
	.seh_endproc							\n\
									\n\
	.globl  wmemcpy							\n\
	.seh_proc wmemcpy						\n\
wmemcpy:								\n\
	.seh_endprologue						\n\
	shlq	$1,%r8		/* cnt * sizeof (wchar_t) */		\n\
	movq	%rcx,%rax	/* return dst */			\n\
	jmp	_memcpy							\n\
	.seh_endproc							\n\
									\n\
	.globl  wmempcpy						\n\
	.seh_proc wmempcpy						\n\
wmempcpy:								\n\
	.seh_endprologue						\n\
	shlq	$1,%r8		/* cnt * sizeof (wchar_t) */		\n\
	movq	%rcx,%rax	/* return dst */			\n\
	addq	%r8,%rax	/*         + n */			\n\
	jmp	_memcpy							\n\
	.seh_endproc							\n\
");

#endif

/* Signal the thread name to any attached debugger

   (See "How to: Set a Thread Name in Native Code"
   https://msdn.microsoft.com/en-us/library/xcb2z8hs.aspx) */

#define MS_VC_EXCEPTION 0x406D1388

void
SetThreadName(DWORD dwThreadID, const char* threadName)
{
  if (!IsDebuggerPresent ())
    return;

  ULONG_PTR info[] =
    {
      0x1000,                 /* type, must be 0x1000 */
      (ULONG_PTR) threadName, /* pointer to threadname */
      dwThreadID,             /* thread ID (+ flags on x86_64) */
#ifdef _X86_
      0,                      /* flags, must be zero */
#endif
    };

#ifdef _X86_
  /* On x86, for __try/__except to work, we must ensure our exception handler is
     installed, which may not be the case if this is being called during early
     initialization. */
  exception protect;
#endif

  __try
    {
      RaiseException (MS_VC_EXCEPTION, 0, sizeof (info) / sizeof (ULONG_PTR),
		      info);
    }
  __except (NO_ERROR)
  __endtry
}

#define add_size(p,s) ((p) = ((__typeof__(p))((PBYTE)(p)+(s))))

static WORD num_cpu_per_group = 0;
static WORD group_count = 0;

WORD
__get_cpus_per_group (void)
{
  tmp_pathbuf tp;

  if (num_cpu_per_group)
    return num_cpu_per_group;

  num_cpu_per_group = 64;
  group_count = 1;

  PSYSTEM_LOGICAL_PROCESSOR_INFORMATION_EX lpi =
            (PSYSTEM_LOGICAL_PROCESSOR_INFORMATION_EX) tp.c_get ();
  DWORD lpi_size = NT_MAX_PATH;

  /* Fake a SYSTEM_LOGICAL_PROCESSOR_INFORMATION_EX group info block on Vista
     systems.  This may be over the top but if the below code just using
     ActiveProcessorCount turns out to be insufficient, we can build on that. */
  if (!wincap.has_processor_groups ()
      || !GetLogicalProcessorInformationEx (RelationGroup, lpi, &lpi_size))
    {
      lpi_size = sizeof *lpi;
      lpi->Relationship = RelationGroup;
      lpi->Size = lpi_size;
      lpi->Group.MaximumGroupCount = 1;
      lpi->Group.ActiveGroupCount = 1;
      lpi->Group.GroupInfo[0].MaximumProcessorCount = wincap.cpu_count ();
      lpi->Group.GroupInfo[0].ActiveProcessorCount
        = __builtin_popcountl (wincap.cpu_mask ());
      lpi->Group.GroupInfo[0].ActiveProcessorMask = wincap.cpu_mask ();
    }

  PSYSTEM_LOGICAL_PROCESSOR_INFORMATION_EX plpi = lpi;
  for (DWORD size = lpi_size; size > 0;
       size -= plpi->Size, add_size (plpi, plpi->Size))
    if (plpi->Relationship == RelationGroup)
      {
        /* There are systems with a MaximumProcessorCount not reflecting the
	   actually available CPUs.  The ActiveProcessorCount is correct
	   though.  So we just use ActiveProcessorCount for now, hoping for
	   the best. */
        num_cpu_per_group = plpi->Group.GroupInfo[0].ActiveProcessorCount;

	/* Follow that lead to get the group count. */
	group_count = plpi->Group.ActiveGroupCount;
        break;
      }

  return num_cpu_per_group;
}

WORD
__get_group_count (void)
{
  if (group_count == 0)
    (void) __get_cpus_per_group (); // caller should have called this first
  return group_count;
}