Re: [PATCH] dlopen: Fix issues related to NODELETE handling and relocations

[Florian Weimer <fweimer at redhat dot com>]

* Carlos O'Donell:

>> +/* Globally exported.  Set by the main program to true before
>> +   termination, and used by tst-dlopen-nodelete-reloc-mod2.so to
>> +   trigger marking his module as NODELETE (and also for its destructor
>
> s/his/this/g

Fixed.

>> +static void __attribute__ ((destructor))
>> +fini (void)
>> +{
>> +  /* This object is never loaded completely.  */
>
> OK, because we never finish up doing the work because mod16 fails.
>
>> +  puts ("error: tst-dlopen-nodelete-reloc-mod15.so destructor invoked");
>
> Why doesn't this fail the test?

Thinko on my part.  Fixed.

>> diff --git a/elf/tst-dlopen-nodelete-reloc-mod16.c b/elf/tst-dlopen-nodelete-reloc-mod16.c
>> new file mode 100644
>> index 0000000000..f836f04fb5
>> --- /dev/null
>> +++ b/elf/tst-dlopen-nodelete-reloc-mod16.c

>> +int
>> +global_function_mod15 (void)
>
> OK. This interposes global_function_mod15 in mod15.

Ugh.  Fixed as well.


>> diff --git a/elf/tst-dlopen-nodelete-reloc-mod5.c b/elf/tst-dlopen-nodelete-reloc-mod5.c
>> new file mode 100644
>> index 0000000000..f876fa0f97

>> +/* Defined in tst-dlopen-nodelete-reloc-mod3.so.  The dependency is
>> +   expressed via DT_NEEDED on the intermedia DSO
>
> s/intermedia/intermediate/g

Fixed.
>> +   Sixth test: NODELETE status is retained after relocation failure
>> +   with unique symbol dependency.  The object graph ensures that the
>> +   unique symbol binding is processed before the dlopen failure.
>> +
>> +                                        DT_NEEDED
>> +     mod17  --(DT_NEEDED)--> mod15 --(unique symbol)--> mod14
>> +       \                       ^                  (RTLD_NODELETE)
>> +        \                 (DT_NEEDED)
>> +         \                     |
>> +          `---(DT_NEEDED)--> mod16
>> +                       (fails to relocate)
>> +
>> +   mod14 must remain NODELETE after opening mod17 failed.  */
>
> OK. Correct.

I added:

   mod14 is loaded first, and the loading mod17 is attempted.

You indicated off-list that you wouldn't object pushing this now, so
I've done that after retesting.

Thanks,
Florian