Re: Alignment and sizes of note sections in 64-bit ELF objects

["H.J. Lu" <hjl.tools@gmail.com>]

On Mon, Nov 20, 2017 at 7:27 AM, Suprateeka R Hegde
<hegdesmailbox@gmail.com> wrote:
> (I took a while to recollect the discussion)
>
> On 16-Nov-2017 10:11 PM, H.J. Lu wrote:
>> On Mon, Apr 3, 2017 at 11:48 AM, Suprateeka R Hegde
>> <hegdesmailbox@gmail.com> wrote:
>>> On 03-Apr-2017 08:49 PM, Ian Lance Taylor wrote:
>>>> On Mon, Apr 3, 2017 at 8:16 AM, H.J. Lu <hjl.tools@gmail.com> wrote:
>>>>> According to gABI, in 64-bit objects, each note entry is an array of 8-bye
>>>>> words in the format of the target processor.  But I got
>>>>>
>>>>>   [ 2] .note.ABI-tag     NOTE            0000000000000254 000254
>>>>> 000020 00   A  0   0  4
>>>>>   [ 3] .note.gnu.build-id NOTE            0000000000000274 000274
>>>>> 000024 00   A  0   0  4
>>>>>
>>>>> on Linux/x86-64.  .note.ABI-tag size is 32, but it isn't aligned at 8 bytes.
>>>>> .note.gnu.build-id size is 36, which isn't multiple of 8 bytes.  Should
>>>>> note sections in 64-bit ELF objects be multiple of 8 bytes as well as
>>>>> aligned to 8 bytes?
>>>
>>> On HP-UX, it is 8 byte alignment. And the size becomes multiple of 8
>>> bytes automatically, as we ensure 8 byte alignment for the next note
>>> entry. This is as per the gABI.
>>>
>>>> This is the comment I wrote in gold when I looked into this:
>>>>
>>>>   // Authorities all agree that the values in a .note field should
>>>>   // be aligned on 4-byte boundaries for 32-bit binaries.  However,
>>>>   // they differ on what the alignment is for 64-bit binaries.
>>>>   // The GABI says unambiguously they take 8-byte alignment:
>>>>   //    http://sco.com/developers/gabi/latest/ch5.pheader.html#note_section
>>>>   // Other documentation says alignment should always be 4 bytes:
>>>>   //    http://www.netbsd.org/docs/kernel/elf-notes.html#note-format
>>>
>>> Thats interesting. But why should it differ from GABI?
>>>
>>> May be because: In both ILP32 and LP64 model, integer is 4 bytes. And
>>> the documentation says "integer" in parenthesis for the fields.
>>>
>>> For example:
>>> Name Size
>>>         4 bytes (integer)
>>> Desc Size
>>>         4 bytes (integer)
>>>
>>> In contrast to GABI, NetBSD seems to have attributed these fields with
>>> "integer".
>>>
>>
>> Linux has been using 4 byte alignment for .note.ABI-tag note and
>> .note.gnu.build-id note in 64-bit objects.  We can't change their
>> alignment to 8 bytes.   Can we add a footnote in gABI to address it?
>
> It should be more clear than that. See below.
>
>> In reality, this isn't a real issue since all notes in one PT_LOAD segment
>> must have the same alignment which equals to p_align.
>
> p_align of PT_LOAD or PT_NOTE?
>
>>  Note parser
>> can use p_align of PT_LOAD segment for note alignment,
>
> Why not p_align of PT_NOTE? p_align of PT_LOAD seems to be (on my
> Ubuntu) set to the 2MiB pagesize value.
>
>> instead of
>> assuming alignment based on ELF file class.
>
> The gABI description of ELF class based alignment may be because gABI
> does not talk anything about PT_NOTE actually being part of PT_LOAD.
> PT_NOTE could be a separate segment on its own outside PT_LOAD, though
> almost all implementations make it part of PT_LOAD.
>
>> BTW, should we document that all notes in one PT_LOAD segment
>> must have the same alignment which equals to p_align?
>
> Why not p_align of PT_NOTE? And p_align of PT_NOTE may be set to the
> maximum of all sh_addralign of all SHT_NOTE that make up the PT_NOTE.

Typo.  It should be PT_NOTE, instead of PT_LOAD.


-- 
H.J.