Bug 1351737

link to build in ppc koji - http://ppc.koji.fedoraproject.org/koji/taskinfo?taskID=3501464

From what I can see it used exactly the same buildroot (same NVRs for packages in the buildroot) as the build in primary koji.

And because krb5-1.14.1-7.fc25 now also fails in the current rawhide buildroot, we should be able to find the differences in the buildroot.

(rawhide, krb5-1.14.1-7.fc25.src.rpm) http://ppc.koji.fedoraproject.org/koji/taskinfo?taskID=3501356
(f24, krb5-1.14.1-8.fc25.src.rpm) http://ppc.koji.fedoraproject.org/koji/taskinfo?taskID=3501244
(rawhide, krb5-1.14.1-8.fc25.src.rpm) http://ppc.koji.fedoraproject.org/koji/taskinfo?taskID=3501128

Only difference between f24 and f25/rawhide is the package split, but that came in with -7 in f25/rawhide and built fine at the time.  So something's changed under me, I think...

Package split also built fine on rhel-7's ppc, ppc64, and ppc64le builders.

The problem seems to happen when converting a string to encryption type. For some reason, strcasecmp is not returning the right result.

Check alias 'des3-cbc-sha1-kd' against 'des3-cbc-sha1-kd': 0(cmp) 49(case_cmp)

This is the printf I added in enctype_util.c, just before the checking for the alias. The (cmp) result is returned by strcmp while (case_cmp) comes from strcasecmp.

I confirm that this build issue is related to changes added in glibc-2.23.90-24.fc25 package update.

Have done local rebuild of krb5-1.14.1-8.fc25 with glibc versions
glibc-2.23.90-22.fc25, glibc-2.23.90-23.fc25 and glibc-2.23.90-24.fc25 locally with no any additional changes made. krb5 builds successfully with glibc-2.23.90-22.fc25 and glibc-2.23.90-23.fc25 but fails with glibc-2.23.90-24.fc25 .

We suspect that it might be related to ppc64(le) specific changes made with commit https://sourceware.org/git/?p=glibc.git;a=commit;h=c8376f3e07602aaef9cb843bb73cb5f2b860634a in glibc.

Switching component to glibc.

(In reply to Rafael Fonseca from comment #4)
> The problem seems to happen when converting a string to encryption type. For
> some reason, strcasecmp is not returning the right result.
> 
> Check alias 'des3-cbc-sha1-kd' against 'des3-cbc-sha1-kd': 0(cmp)
> 49(case_cmp)

Is this under the C locale, or some other locale?

Could you please specify the architecture where you saw this?  Fedora no longer supports powerpc (32-bit).

Okay, based comment 1, it happens at least on ppc64le.

(In reply to Florian Weimer from comment #7)
> (In reply to Rafael Fonseca from comment #4)
> > The problem seems to happen when converting a string to encryption type. For
> > some reason, strcasecmp is not returning the right result.
> > 
> > Check alias 'des3-cbc-sha1-kd' against 'des3-cbc-sha1-kd': 0(cmp)
> > 49(case_cmp)
> 
> Is this under the C locale, or some other locale?
> 
> Could you please specify the architecture where you saw this?  Fedora no

It occurs on both pcc64 and ppc64le.
Scratch build for latest krb5 is at http://ppc.koji.fedoraproject.org/koji/taskinfo?taskID=3501128

Did ppc64(le) scratch build of glibc glibc-2.23.90-25.fc25 with reverted commit https://sourceware.org/git/?p=glibc.git;a=commit;h=c8376f3e07602aaef9cb843bb73cb5f2b860634a. Scratch build can be found at http://ppc.koji.fedoraproject.org/koji/taskinfo?taskID=3509465 . 

Updating local test machine (ppc64 guest) with new glibc scratch build glibc-2.23.90-26 leads to krb5 builds successfully.

I'm trying to use this test program to reproduce the issue without krb5.  So far, no luck.

#include <string.h>
#include <stdlib.h>
#include <stdbool.h>
#include <malloc.h>
#include <locale.h>

int
main (void)
{
  if (setlocale (LC_ALL, "") == NULL)
    {
      printf ("error: setlocale: %m\n");
      return 1;
    }

  enum { max_align = 128 };
  const char *str = "des3-cbc-sha1-kd";
  size_t blob_size = max_align + strlen (str) + 1;
  char *left = memalign (max_align, blob_size);
  char *right = memalign (max_align, blob_size);
  if (left == NULL || right == NULL)
    {
      printf ("error: out of memory\n");
      return 1;
    }

  bool errors = false;
  for (int left_align = 0; left_align < max_align; ++left_align)
    {
      memset (left, 'L', blob_size);
      strcpy (left + left_align, str);
      for (int right_align = 0; right_align < max_align; ++right_align)
        {
          memset (right, 'R', blob_size);
          strcpy (right + right_align, str);
          if (strcasecmp (left + left_align,
                          right + right_align) != 0)
            {
              printf ("error: mismatch; left_align = %d, right_align = %d\n",
                      left_align, right_align);
              errors = true;
            }
        }
    }

  free (right);
  free (left);
  return errors;
}

(In reply to Rafael Fonseca from comment #4)
> The problem seems to happen when converting a string to encryption type. For
> some reason, strcasecmp is not returning the right result.
> 
> Check alias 'des3-cbc-sha1-kd' against 'des3-cbc-sha1-kd': 0(cmp)
> 49(case_cmp)
> 
> This is the printf I added in enctype_util.c, just before the checking for
> the alias. The (cmp) result is returned by strcmp while (case_cmp) comes
> from strcasecmp.

Could you please post the patch that you used for debugging?  Thanks.

Created attachment 1176050 [details]
krb5 debugging printfs

These are the modifications I made to krb5.

Created attachment 1176051 [details]
krb5 debugging printfs

This is the right version.

Created attachment 1176053 [details]
Output from debugging

This is the output when using the modifications above.

I reported this upstream.  It does not appear to be an alignment issue.

This bug only ever affected rawhide, where it has been fixed.