Re: [PATCH 2/3] btf: fix 'extern const void' variables [PR106773]

[David Faust <david.faust@oracle.com>]

On 12/8/22 23:34, Indu Bhagat wrote:
> Looks OK to me overall. Minor comments below.
> 
> Thanks
> 
> On 12/7/22 12:57, David Faust wrote:
>> The eBPF loader expects to find BTF_KIND_VAR records for references to
>> extern const void symbols. We were mistakenly identifing these as
>> unsupported types, and as a result skipping emitting VAR records for
>> them.
>>
>> In addition, the internal DWARF representation from which BTF is
>> produced does not generate 'const' modifier DIEs for the void type,
>> which meant in BTF the 'const' qualifier was dropped for 'extern const
>> void' variables. This patch also adds support for generating a const
>> void type in BTF to correct emission for these variables.
>>
>> 	PR target/106773
>>
>> gcc/
>>
>> 	* btfout.cc (btf_collect_datasec): Correct size of void entries.
>> 	(btf_dvd_emit_preprocess_cb): Do not skip emitting variables which
>> 	refer to void types.
>> 	(btf_init_postprocess): Create 'const void' type record if needed and
>> 	adjust variables to refer to it as appropriate.
>>
>> gcc/testsuite/
>>
>> 	* gcc.dg/debug/btf/btf-pr106773.c: New test.
>> ---
>>   gcc/btfout.cc                                 | 44 +++++++++++++++++--
>>   gcc/testsuite/gcc.dg/debug/btf/btf-pr106773.c | 25 +++++++++++
>>   2 files changed, 65 insertions(+), 4 deletions(-)
>>   create mode 100644 gcc/testsuite/gcc.dg/debug/btf/btf-pr106773.c
>>
>> diff --git a/gcc/btfout.cc b/gcc/btfout.cc
>> index a1c6266a7db..05f3a3f9b6e 100644
>> --- a/gcc/btfout.cc
>> +++ b/gcc/btfout.cc
>> @@ -354,6 +354,8 @@ btf_collect_datasec (ctf_container_ref ctfc)
>>         tree size = DECL_SIZE_UNIT (node->decl);
>>         if (tree_fits_uhwi_p (size))
>>   	info.size = tree_to_uhwi (size);
>> +      else if (VOID_TYPE_P (TREE_TYPE (node->decl)))
>> +	info.size = 1;
>>   
>>         /* Offset is left as 0 at compile time, to be filled in by loaders such
>>   	 as libbpf.  */
>> @@ -439,7 +441,7 @@ btf_dvd_emit_preprocess_cb (ctf_dvdef_ref *slot, ctf_container_ref arg_ctfc)
>>     ctf_dvdef_ref var = (ctf_dvdef_ref) * slot;
>>   
>>     /* Do not add variables which refer to unsupported types.  */
>> -  if (btf_removed_type_p (var->dvd_type))
>> +  if (!voids.contains (var->dvd_type) && btf_removed_type_p (var->dvd_type))
>>       return 1;
>>   
>>     arg_ctfc->ctfc_vars_list[num_vars_added] = var;
>> @@ -1073,15 +1075,49 @@ btf_init_postprocess (void)
>>   {
>>     ctf_container_ref tu_ctfc = ctf_get_tu_ctfc ();
>>   
>> -  size_t i;
>> -  size_t num_ctf_types = tu_ctfc->ctfc_types->elements ();
>> -
>>     holes.create (0);
>>     voids.create (0);
>>   
>>     num_types_added = 0;
>>     num_types_created = 0;
>>   
>> +  /* Workaround for 'const void' variables. These variables are sometimes used
>> +     in eBPF programs to address kernel symbols. DWARF does not generate const
>> +     qualifier on void type, so we would incorrectly emit these variables
>> +     without the const qualifier.
>> +     Unfortunately we need the TREE node to know it was const, and we need
>> +     to create the const modifier type (if needed) now, before making the types
>> +     list. So we can't avoid iterating with FOR_EACH_VARIABLE here, and then
>> +     again when creating the DATASEC entries.  */
> 
> "Dot, space, space, new sentence." in 3 places.
> 
> 
>> +  ctf_id_t constvoid_id = CTF_NULL_TYPEID;
>> +  varpool_node *var;
>> +  FOR_EACH_VARIABLE (var)
>> +    {
>> +      if (!var->decl)
>> +	continue;
>> +
>> +      tree type = TREE_TYPE (var->decl);
>> +      if (type && VOID_TYPE_P (type) && TYPE_READONLY (type))
>> +	{
>> +	  dw_die_ref die = lookup_decl_die (var->decl);
>> +	  if (die == NULL)
>> +	    continue;
>> +
>> +	  ctf_dvdef_ref dvd = ctf_dvd_lookup (tu_ctfc, die);
>> +	  if (dvd == NULL)
>> +	    continue;
>> +
>> +	  /* Create the 'const' modifier type for void.  */
>> +	  if (constvoid_id == CTF_NULL_TYPEID)
>> +	    constvoid_id = ctf_add_reftype (tu_ctfc, CTF_ADD_ROOT,
>> +					    dvd->dvd_type, CTF_K_CONST, NULL);
> 
> No de-duplication of the const void type.  I assume libbpf will take 
> care of this eventually.

Hm, not sure I follow. Where is the duplication? The const void type is
only created once here, for the first such variable which needs it, and
reused for subsequent variables. And it does not already exist in the
CTF which we are translating into BTF.

In any case, yes libbpf can handle duplicated types. Though it would
still be good to minimize that where we can to not bloat the BTF info.

> 
>> +	  dvd->dvd_type = constvoid_id;
>> +	}
>> +    }
>> +
>> +  size_t i;
>> +  size_t num_ctf_types = tu_ctfc->ctfc_types->elements ();
>> +
>>     if (num_ctf_types)
>>       {
>>         init_btf_id_map (num_ctf_types + 1);
>> diff --git a/gcc/testsuite/gcc.dg/debug/btf/btf-pr106773.c b/gcc/testsuite/gcc.dg/debug/btf/btf-pr106773.c
>> new file mode 100644
>> index 00000000000..f90fa773a4b
>> --- /dev/null
>> +++ b/gcc/testsuite/gcc.dg/debug/btf/btf-pr106773.c
>> @@ -0,0 +1,25 @@
>> +/* Test BTF generation for extern const void symbols.
>> +   BTF_KIND_VAR records should be emitted for such symbols if they are used,
>> +   as well as a corresponding entry in the appropriate DATASEC record.  */
>> +
>> +/* { dg-do compile } */
>> +/* { dg-options "-O0 -gbtf -dA" } */
>> +
>> +/* Expect 1 variable record only for foo, with 'extern' (2) linkage.  */
>> +/* { dg-final { scan-assembler-times "\[\t \]0xe000000\[\t \]+\[^\n\]*btv_info" 1 } } */
>> +/* { dg-final { scan-assembler-times "\[\t \]0x2\[\t \]+\[^\n\]*btv_linkage" 1 } } */
>> +
>> +/* { dg-final { scan-assembler-times "ascii \"foo.0\"\[\t \]+\[^\n\]*btf_string" 1 } } */
>> +
>> +/* { dg-final { scan-assembler-times "0\[\t \]+\[^\n\]*bts_offset" 1 } } */
>> +/* { dg-final { scan-assembler-times "1\[\t \]+\[^\n\]*bts_size" 1 } } */
>> +
>> +extern const void foo __attribute__((weak)) __attribute__((section (".ksyms")));
>> +extern const void bar __attribute__((weak)) __attribute__((section (".ksyms")));
>> +
>> +unsigned long func () {
>> +  unsigned long x = (unsigned long) &foo;
>> +
>> +  return x;
>> +}
>> +
>