[PATCH v4 2/2] x86: Update _dl_tlsdesc_dynamic to preserve caller-saved registers

[Adhemerval Zanella Netto]

On 16/02/24 09:58, H.J. Lu wrote:
> On Fri, Feb 16, 2024 at 4:47 AM Adhemerval Zanella Netto
> <adhemerval.zanella@linaro.org> wrote:
>>
>>
>>
>> On 16/02/24 09:37, H.J. Lu wrote:
>>> On Fri, Feb 16, 2024 at 4:20 AM H.J. Lu <hjl.tools@gmail.com> wrote:
>>>>
>>>> On Fri, Feb 16, 2024 at 4:18 AM Florian Weimer <fweimer@redhat.com> wrote:
>>>>>
>>>>> * H. J. Lu:
>>>>>
>>>>>> On Thu, Feb 15, 2024 at 10:23 PM Florian Weimer <fweimer@redhat.com> wrote:
>>>>>>>
>>>>>>> * Adhemerval Zanella Netto:
>>>>>>>
>>>>>>>>> diff --git a/elf/malloc-for-test.map b/elf/malloc-for-test.map
>>>>>>>>> new file mode 100644
>>>>>>>>> index 0000000000..8437cf4346
>>>>>>>>> --- /dev/null
>>>>>>>>> +++ b/elf/malloc-for-test.map
>>>>>>>>> @@ -0,0 +1,6 @@
>>>>>>>>> +GLIBC_2.0 {
>>>>>>>>
>>>>>>>> You need to use the correct version to override the malloc:
>>>>>>>>
>>>>>>>> $ gdb --args tst-gnu2-tls2 --direct
>>>>>>>> [...]
>>>>>>>> (gdb) b apply_tls
>>>>>>>> (gdb) r
>>>>>>>> Thread 2 "tst-gnu2-tls2" hit Breakpoint 1, apply_tls (p=0x7ffff7bfee80) at tst-gnu2-tls2mod1.c:25
>>>>>>>> 25 {
>>>>>>>> (gdb) b malloc
>>>>>>>> Breakpoint 2 at 0x7ffff7ca8ad0: malloc. (3 locations)
>>>>>>>> (gdb) c
>>>>>>>> Continuing.
>>>>>>>
>>>>>>> Why do we need to set a symbol version here?  I think this can be
>>>>>>> removed.
>>>>>>
>>>>>> malloc-for-test.so provides a fake malloc to clobber caller-save
>>>>>> registers.  Since malloc in ld.so has a symbol version,
>>>>>>
>>>>>>   struct r_found_version version;
>>>>>>   version.name = symbol_version_string (libc, GLIBC_2_0);
>>>>>>   version.hidden = 0;
>>>>>>   version.hash = _dl_elf_hash (version.name);
>>>>>>   version.filename = NULL;
>>>>>>
>>>>>>   void *new_calloc = lookup_malloc_symbol (main_map, "calloc", &version);
>>>>>>   void *new_free = lookup_malloc_symbol (main_map, "free", &version);
>>>>>>   void *new_malloc = lookup_malloc_symbol (main_map, "malloc", &version);
>>>>>>   void *new_realloc = lookup_malloc_symbol (main_map, "realloc", &version);
>>>>>>
>>>>>> malloc in malloc-for-test.so must have the same version.  Otherwise,
>>>>>> it won't be used in the test.
>>>>>
>>>>> I thought that unversioned symbols interpose all versioned symbols.  Has
>>>>> this changed in the dynamic linker?
>>>>
>>>> Only for this case.
>>>>
>>>
>>> I was wrong.  Version isn't needed.  Will fix it.
>>>
>>
>> Indeed, but I think we also improve the register check testing since just clobbering
>> on malloc does not really provide much coverage.
> 
> Why?  The slow path will always call malloc in malloc-for-test.so.
> We can clobber any caller-save registers we need.  Without
> malloc-for-test.so, the test doesn't fail for i386 since malloc in libc.so
> doesn't use any vector registers.
> 

The idea is to check whether _dl_tlsdesc_dynamic call does preserve any
possible caller-saved register and since this test is generic and afaik
all affected ABIs follow the same idea (compiler won't save/restore such
register), clobbering the register will only trigger a possible issue
iff the thread that actually issue the TLS usage does actually use any
possible register.

That's why I think a better coverage would to also clobber the register
before the TLS access, and check if their values does not change over
the TLS access.