[Libunwind-devel] mincore(2) problems on Linux

[Jason Evans]

I have been testing recent versions of libunwind on Centos 5.2 (x86_64), 
and it appears that mincore(2) does not do what we want.  That is, 
msync(2) and mincore(2) do not give the same results.

While tracking this down, I noticed a longer-standing problem in 
validate_mem() having to do with unaligned access that straddles pages. 
 In practice I wouldn't be surprised if this bug is never triggered, 
but fixing it seems worthwhile nonetheless.  Following is a partial 
patch, but it doesn't actually fix the problem.  Can anyone shed some 
light on whether mincore(2) is viable on Linux?  If not I'll try to come 
up with a more sophisticated autoconf test for mincore(2).

Thanks,
Jason

diff --git a/src/x86_64/Ginit.c b/src/x86_64/Ginit.c
index 14614f1..7942c39 100644
--- a/src/x86_64/Ginit.c
+++ b/src/x86_64/Ginit.c
@@ -33,6 +33,9 @@ WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE 
SOFTWARE.  */
 #include <stdlib.h>
 #include <string.h>
 #include <sys/mman.h>
+#ifdef HAVE_MINCORE
+#include <errno.h>
+#endif

 #include "unwind_i.h"

@@ -93,6 +96,12 @@ validate_mem (unw_word_t addr)
 #ifdef HAVE_MINCORE
   char mvec[2]; /* Unaligned access may cross page boundary */
 #endif
+  size_t len;
+
+  if (PAGE_START(addr + sizeof (unw_word_t) - 1) == PAGE_START(addr))
+    len = PAGE_SIZE;
+  else
+    len = PAGE_SIZE * 2;

   addr = PAGE_START(addr);

@@ -106,11 +115,21 @@ validate_mem (unw_word_t addr)
     }

 #ifdef HAVE_MINCORE
-  if (mincore ((void *) addr, sizeof (unw_word_t), mvec) == -1)
+  while (1)
+    {
+      if (mincore ((void *) addr, len, mvec) == 0)
+        break;
+      switch (errno)
+       {
+       case EAGAIN: break;
+       case ENOMEM: return -1;
+       case EFAULT: case EINVAL: default: abort ();
+       }
+    }
 #else
-  if (msync ((void *) addr, sizeof (unw_word_t), MS_ASYNC) == -1)
-#endif
+  if (msync ((void *) addr, len, MS_ASYNC) == -1)
     return -1;
+#endif

   victim = lga_victim;
   for (i = 0; i < NLGA; i++) {