diff --git a/winsup/cygwin/common.din b/winsup/cygwin/common.din
index 9584d09..243fd01 100644
--- a/winsup/cygwin/common.din
+++ b/winsup/cygwin/common.din
@@ -269,6 +269,7 @@ ctime SIGFE
 ctime_r SIGFE
 cuserid NOSIGFE
 cwait SIGFE
+cygheap_profthr_all NOSIGFE
 cygwin_attach_handle_to_fd SIGFE
 cygwin_conv_path SIGFE
 cygwin_conv_path_list SIGFE
diff --git a/winsup/cygwin/cygheap.cc b/winsup/cygwin/cygheap.cc
index 6493485..5159d30 100644
--- a/winsup/cygwin/cygheap.cc
+++ b/winsup/cygwin/cygheap.cc
@@ -744,3 +744,16 @@ init_cygheap::find_tls (int sig, bool& issig_wait)
     WaitForSingleObject (t->mutex, INFINITE);
   return t;
 }
+
+/* Called from profil.c to sample all non-main thread PC values for profiling */
+extern "C" void
+cygheap_profthr_all (void (*profthr_byhandle) (HANDLE))
+{
+  int ix = -1;
+  while (++ix < (int) nthreads)
+    {
+      _cygtls *tls = cygheap->threadlist[ix].thread;
+      if (tls->tid)
+	profthr_byhandle (tls->tid->win32_obj_id);
+    }
+}
diff --git a/winsup/cygwin/gmon.c b/winsup/cygwin/gmon.c
index 96b1189..0b7ecc0 100644
--- a/winsup/cygwin/gmon.c
+++ b/winsup/cygwin/gmon.c
@@ -36,6 +36,7 @@ static char rcsid[] = "$OpenBSD: gmon.c,v 1.8 1997/07/23 21:11:27 kstailey Exp $
  * The differences should be within __MINGW32__ guard.
  */
 
+#include <errno.h>
 #include <fcntl.h>
 #include <stdlib.h>
 #include <stdio.h>
@@ -49,6 +50,7 @@ static char rcsid[] = "$OpenBSD: gmon.c,v 1.8 1997/07/23 21:11:27 kstailey Exp $
 
 /* XXX needed? */
 //extern char *minbrk __asm ("minbrk");
+extern int _setmode(int, int);
 
 #ifdef _WIN64
 #define MINUS_ONE_P (-1LL)
@@ -152,6 +154,7 @@ void
 _mcleanup(void)
 {
 	static char gmon_out[] = "gmon.out";
+	static char gmon_template[] = "gmon.outXXXXXX";
 	int fd;
 	int hz;
 	int fromindex;
@@ -222,7 +225,14 @@ _mcleanup(void)
 	proffile = gmon_out;
 #endif
 
-	fd = open(proffile , O_CREAT|O_TRUNC|O_WRONLY|O_BINARY, 0666);
+	fd = open(proffile, O_CREAT|O_EXCL|O_TRUNC|O_WRONLY|O_BINARY, 0666);
+	if (fd < 0 && errno == EEXIST) {
+		fd = mkstemp(gmon_template);
+		if (fd >= 0) {
+			_setmode(fd, O_BINARY);
+			fchmod(fd, 0644);
+		}
+	}
 	if (fd < 0) {
 		perror( proffile );
 		return;
diff --git a/winsup/cygwin/gmon.h b/winsup/cygwin/gmon.h
index 0932ed9..b0fb479 100644
--- a/winsup/cygwin/gmon.h
+++ b/winsup/cygwin/gmon.h
@@ -153,7 +153,7 @@ struct rawarc {
  * The profiling data structures are housed in this structure.
  */
 struct gmonparam {
-	int		state;
+	volatile int	state;
 	u_short		*kcount;
 	size_t		kcountsize;
 	u_short		*froms;
diff --git a/winsup/cygwin/mcount.c b/winsup/cygwin/mcount.c
index fad6728..6111b35 100644
--- a/winsup/cygwin/mcount.c
+++ b/winsup/cygwin/mcount.c
@@ -41,6 +41,7 @@ static char rcsid[] = "$OpenBSD: mcount.c,v 1.6 1997/07/23 21:11:27 kstailey Exp
 #endif
 #include <sys/types.h>
 #include "gmon.h"
+#include "winsup.h"
 
 /*
  * mcount is called on entry to each function compiled with the profiling
@@ -70,11 +71,12 @@ _MCOUNT_DECL (size_t frompc, size_t selfpc)
 	p = &_gmonparam;
 	/*
 	 * check that we are profiling
-	 * and that we aren't recursively invoked.
+	 * and that we aren't recursively invoked by this thread
+	 * or entered anew by any other thread.
 	 */
-	if (p->state != GMON_PROF_ON)
+	if (InterlockedCompareExchange (
+		    &p->state, GMON_PROF_BUSY, GMON_PROF_ON) != GMON_PROF_ON)
 		return;
-	p->state = GMON_PROF_BUSY;
 	/*
 	 * check that frompcindex is a reasonable pc value.
 	 * for example:	signal catchers get called from the stack,
@@ -162,10 +164,10 @@ _MCOUNT_DECL (size_t frompc, size_t selfpc)
 		}
 	}
 done:
-	p->state = GMON_PROF_ON;
+	InterlockedExchange (&p->state, GMON_PROF_ON);
 	return;
 overflow:
-	p->state = GMON_PROF_ERROR;
+	InterlockedExchange (&p->state, GMON_PROF_ERROR);
 	return;
 }
 
diff --git a/winsup/cygwin/profil.c b/winsup/cygwin/profil.c
index eb41c08..f76fda5 100644
--- a/winsup/cygwin/profil.c
+++ b/winsup/cygwin/profil.c
@@ -28,6 +28,8 @@
 /* global profinfo for profil() call */
 static struct profinfo prof;
 
+extern void cygheap_profthr_all (void (*) (HANDLE));
+
 /* Get the pc for thread THR */
 
 static size_t
@@ -65,25 +67,36 @@ print_prof (struct profinfo *p)
 }
 #endif
 
-/* Everytime we wake up use the main thread pc to hash into the cell in the
-   profile buffer ARG. */
+/* Every time we wake up sample the main thread's pc to hash into the cell
+   in the profile buffer ARG.  Then all other pthreads' pc's are sampled.  */
 
-static void CALLBACK profthr_func (LPVOID);
+static void
+profthr_byhandle (HANDLE thr)
+{
+  size_t idx;
+  size_t pc = (size_t) get_thrpc (thr);
+
+  // code assumes there is only one profinfo in play: the static prof up top
+  if (pc >= prof.lowpc && pc < prof.highpc)
+    {
+      idx = PROFIDX (pc, prof.lowpc, prof.scale);
+      prof.counter[idx]++;
+    }
+}
 
 static void CALLBACK
 profthr_func (LPVOID arg)
 {
   struct profinfo *p = (struct profinfo *) arg;
-  size_t pc, idx;
 
   for (;;)
     {
-      pc = (size_t) get_thrpc (p->targthr);
-      if (pc >= p->lowpc && pc < p->highpc)
-	{
-	  idx = PROFIDX (pc, p->lowpc, p->scale);
-	  p->counter[idx]++;
-	}
+      // record profiling sample for main thread
+      profthr_byhandle (p->targthr);
+
+      // record profiling samples for other pthreads, if any
+      cygheap_profthr_all (profthr_byhandle);
+
 #if 0
       print_prof (p);
 #endif