guile-dbi cores in mark_db_handle

[Andrew Gaylard]

Hi guile people,

I'm experiencing cores in guile-dbi.

This is on Solaris-10u10 on SPARC, with guile and guile-dbi compiled 
with -O2.
When compiled with -O0, the problem disappears.  I've also been unable to
replicate the problem on Linux x86_64. This suggests to me that the
underlying cause could be timing-related (my guile app does use threads).

This is what gdb shows:

Program terminated with signal 10, Bus error.
#0  0xffffffff75601c8c in mark_db_handle (g_db_handle_smob=0x10115efc0) 
at guile-dbi.c:114
114       scm_gc_mark(g_db_handle->bcknd);
(gdb) p g_db_handle
$1 = (gdbi_db_handle_t *) 0x304

That 0x304 is SCM_EOL. The only part of guile-dbi.c which looks like it 
could be
causing this is free_db_handle, which does this:

SCM_SETCDR (g_db_handle_smob, SCM_EOL);

If this is the cause, then why is mark_db_handle being called after the 
object
has already been free'd?

I can prevent the cores by adding this to the start of mark_db_handle:

if (g_db_handle == (gdbi_db_handle_t*) SCM_EOL) return SCM_UNSPECIFIED;

However, I'm concerned that this isn't the right fix; what if something 
really does need
to be marked, and isn't?  -- it'll be GC'd when it shouldn't, right?

--
Andrew