[PATCH 2/4] Make the backtrace-data test helper more robust

[PATCH 2/4] Make the backtrace-data test helper more robust

[Ulf Hermann]

When unwinding by frame pointer the unwinder might ask for invalid
addresses. We don't have to fail the test in this case. In fact
any broken dwarf information can lead to requests for invalid
addresses, also without frame pointer unwinding.

Signed-off-by: Ulf Hermann <ulf.hermann@qt.io>
---
  tests/ChangeLog        |  6 ++++++
  tests/backtrace-data.c | 16 +++++++++++++---
  2 files changed, 19 insertions(+), 3 deletions(-)

diff --git a/tests/ChangeLog b/tests/ChangeLog
index 7031cf5..7040ac8 100644
--- a/tests/ChangeLog
+++ b/tests/ChangeLog
@@ -1,3 +1,9 @@
+2017-02-09  Ulf Hermann  <ulf.hermann@qt.io>
+
+	* backtrace-data.c: Don't assert that symbols are found.
+	The unwinder is allowed to ask for invalid addresses. We deny
+	such requests, rather than make the test fail.
+
  2016-11-17  Mark Wielaard  <mjw@redhat.com>
  
  	* run-readelf-s.sh: Add --symbols=.dynsym and --symbols=.symtab tests.
diff --git a/tests/backtrace-data.c b/tests/backtrace-data.c
index b7158da..a387d8f 100644
--- a/tests/backtrace-data.c
+++ b/tests/backtrace-data.c
@@ -76,10 +76,15 @@ memory_read (Dwfl *dwfl, Dwarf_Addr addr, Dwarf_Word *result,
  
    errno = 0;
    long l = ptrace (PTRACE_PEEKDATA, child, (void *) (uintptr_t) addr, NULL);
-  assert (errno == 0);
+
+  // The unwinder can ask for an invalid address.
+  // Don't assert on that but just politely refuse.
+  if (errno != 0) {
+      errno = 0;
+      return false;
+  }
    *result = l;
  
-  /* We could also return false for failed ptrace.  */
    return true;
  }
  
@@ -103,7 +108,8 @@ maps_lookup (pid_t pid, Dwarf_Addr addr, GElf_Addr *basep)
        unsigned long start, end, offset;
        i = fscanf (f, "%lx-%lx %*s %lx %*x:%*x %*x", &start, &end, &offset);
        assert (errno == 0);
-      assert (i == 3);
+      if (i != 3)
+          break;
        char *filename = strdup ("");
        assert (filename);
        size_t filename_len = 0;
@@ -131,6 +137,8 @@ maps_lookup (pid_t pid, Dwarf_Addr addr, GElf_Addr *basep)
  	}
        free (filename);
      }
+  *basep = 0;
+  return NULL;
  }
  
  /* Add module containing ADDR to the DWFL address space.
@@ -145,6 +153,8 @@ report_module (Dwfl *dwfl, pid_t child, Dwarf_Addr addr)
  {
    GElf_Addr base;
    char *long_name = maps_lookup (child, addr, &base);
+  if (!long_name)
+      return NULL; // not found
    Dwfl_Module *mod = dwfl_report_elf (dwfl, long_name, long_name, -1,
  				      base, false /* add_p_vaddr */);
    assert (mod);
-- 
2.1.4

Re: [PATCH 2/4] Make the backtrace-data test helper more robust

[Mark Wielaard]

On Fri, 2017-02-10 at 15:19 +0100, Ulf Hermann wrote:
> When unwinding by frame pointer the unwinder might ask for invalid
> addresses. We don't have to fail the test in this case. In fact
> any broken dwarf information can lead to requests for invalid
> addresses, also without frame pointer unwinding.
> 
> Signed-off-by: Ulf Hermann <ulf.hermann@qt.io>

Thanks. This makes sense to me. Applied.

Note the patch had the same issue as the previous one
(lines starting with a space had the first space doubled).

Cheers,

Mark