Re: [PATCH v2 10/14] objtool: Add support for relative switch tables

2023-06-23 Thread Christophe Leroy


Le 22/06/2023 à 13:48, Peter Zijlstra a écrit :
> On Thu, Jun 22, 2023 at 12:54:32PM +0200, Christophe Leroy wrote:
>> On powerpc, switch tables are relative, than means the address of the
>> table is added to the value of the entry in order to get the pointed
>> address: (r10 is the table address, r4 the index in the table)
>>
>>lis r10,0 <== Load r10 with upper part of .rodata address
>>R_PPC_ADDR16_HA .rodata
>>addir10,r10,0 <== Add lower part of .rodata address
>>R_PPC_ADDR16_LO .rodata
>>lwzxr8,r10,r4 <== Read table entry at r10 + r4 into r8
>>add r10,r8,r10<== Add table address to read value
>>mtctr   r10   <== Save calculated address in CTR
>>bctr  <== Branch to address in CTR
>>
>> But for c_jump_tables it is not the case, they contain the
>> pointed address directly:
>>
>>lis r28,0 <== Load r28 with upper .rodata..c_jump_table
>>R_PPC_ADDR16_HA   .rodata..c_jump_table
>>addir28,r28,0 <== Add lower part of 
>> .rodata..c_jump_table
>>R_PPC_ADDR16_LO   .rodata..c_jump_table
>>lwzxr10,r28,r10   <== Read table entry at r10 + r28 into r10
>>mtctr   r10   <== Save read value in CTR
>>bctr  <== Branch to address in CTR
>>
>> Add support to objtool for relative tables, with a flag in order to
>> tell table by table if that table uses relative or absolute addressing.
>>
>> And use correct size for 'long' instead of hard coding a size of '8'.
> 
> Again, see tip/objtool/core...
> 
> This one is going to be a little hard. It would be very good if you can
> avoid growing struct reloc; Josh went through a ton of effort to shrink
> it.
> 
> Would it work to use reloc->sym->is_rel_jumptable ? That still has a few
> spare bits in it's bitfield.
> 

I noticed that the relocation types are different, so the flag is indeed 
not needed. I have now done the following, will it work for x86 ?

diff --git a/tools/objtool/check.c b/tools/objtool/check.c
index ae4a1608d97e..05d789e6d3b6 100644
--- a/tools/objtool/check.c
+++ b/tools/objtool/check.c
@@ -1988,7 +1988,7 @@ static int add_jump_table(struct objtool_file 
*file, struct instruction *insn,
struct symbol *pfunc = insn_func(insn)->pfunc;
struct reloc *table = insn_jump_table(insn);
struct instruction *dest_insn;
-   unsigned int prev_offset = 0;
+   unsigned int offset, prev_offset = 0;
struct reloc *reloc = table;
struct alternative *alt;

@@ -2003,7 +2003,7 @@ static int add_jump_table(struct objtool_file 
*file, struct instruction *insn,
break;

/* Make sure the table entries are consecutive: */
-   if (prev_offset && reloc_offset(reloc) != prev_offset + 8)
+   if (prev_offset && reloc_offset(reloc) != prev_offset + 
elf_addr_size(file->elf))
break;

/* Detect function pointers from contiguous objects: */
@@ -2011,7 +2011,12 @@ static int add_jump_table(struct objtool_file 
*file, struct instruction *insn,
reloc_addend(reloc) == pfunc->offset)
break;

-   dest_insn = find_insn(file, reloc->sym->sec, 
reloc_addend(reloc));
+   if (reloc_type(reloc) == R_ABS32 || reloc_type(reloc) == 
R_ABS64)
+   offset = reloc_addend(reloc);
+   else
+   offset = reloc_addend(reloc) + reloc_offset(table) - 
reloc_offset(reloc);
+
+   dest_insn = find_insn(file, reloc->sym->sec, offset);
if (!dest_insn)
break;


Christophe


Re: [PATCH v2 10/14] objtool: Add support for relative switch tables

2023-06-22 Thread Peter Zijlstra
On Thu, Jun 22, 2023 at 12:54:32PM +0200, Christophe Leroy wrote:
> On powerpc, switch tables are relative, than means the address of the
> table is added to the value of the entry in order to get the pointed
> address: (r10 is the table address, r4 the index in the table)
> 
>   lis r10,0   <== Load r10 with upper part of .rodata address
>   R_PPC_ADDR16_HA .rodata
>   addir10,r10,0   <== Add lower part of .rodata address
>   R_PPC_ADDR16_LO .rodata
>   lwzxr8,r10,r4   <== Read table entry at r10 + r4 into r8
>   add r10,r8,r10  <== Add table address to read value
>   mtctr   r10 <== Save calculated address in CTR
>   bctr<== Branch to address in CTR
> 
> But for c_jump_tables it is not the case, they contain the
> pointed address directly:
> 
>   lis r28,0   <== Load r28 with upper .rodata..c_jump_table
>   R_PPC_ADDR16_HA   .rodata..c_jump_table
>   addir28,r28,0   <== Add lower part of 
> .rodata..c_jump_table
>   R_PPC_ADDR16_LO   .rodata..c_jump_table
>   lwzxr10,r28,r10 <== Read table entry at r10 + r28 into r10
>   mtctr   r10 <== Save read value in CTR
>   bctr<== Branch to address in CTR
> 
> Add support to objtool for relative tables, with a flag in order to
> tell table by table if that table uses relative or absolute addressing.
> 
> And use correct size for 'long' instead of hard coding a size of '8'.

Again, see tip/objtool/core...

This one is going to be a little hard. It would be very good if you can
avoid growing struct reloc; Josh went through a ton of effort to shrink
it.

Would it work to use reloc->sym->is_rel_jumptable ? That still has a few
spare bits in it's bitfield.

> diff --git a/tools/objtool/include/objtool/elf.h 
> b/tools/objtool/include/objtool/elf.h
> index e1ca588eb69d..64aac87a4825 100644
> --- a/tools/objtool/include/objtool/elf.h
> +++ b/tools/objtool/include/objtool/elf.h
> @@ -80,6 +80,7 @@ struct reloc {
>   s64 addend;
>   int idx;
>   bool jump_table_start;
> + bool jump_table_is_rel;
>  };