binutils ld and new PT_GNU_PROPERTY segment

[Fangrui Song]

On 2020-02-19, H.J. Lu wrote:
>On Wed, Feb 19, 2020 at 10:27 AM Fangrui Song <i@maskray.me> wrote:
>>
>> On 2020-02-19, H.J. Lu wrote:
>> >On Wed, Feb 19, 2020 at 5:17 AM Mark Wielaard <mark@klomp.org> wrote:
>> >>
>> >> On Wed, 2020-02-19 at 04:28 -0800, H.J. Lu wrote:
>> >> > On Wed, Feb 19, 2020 at 4:02 AM Mark Wielaard <mark@klomp.org> wrote:
>> >> > > > https://patchwork.kernel.org/patch/11285409/
>> >> > > >
>> >> > > > It is for both x86 and arm64.
>> >> > >
>> >> > > So that is not upstream in the mainline kernel? Why can't that patch
>> >> > > use the existing PT_NOTE segment? That would make it compatible with
>> >> > > existing binaries that don't have this PT_GNU_PROPERTY program header.
>> >> >
>> >> > Kernel loader is one of motivations of PT_GNU_PROPERTY.  Kernel loader
>> >> > only wants to check PT_XXX.
>> >>
>> >> So they can check PT_NOTE because it provides the same information and
>> >> is already available in existing binaries.
>> >>
>> >
>> >Please take a look at glibc note.gnu.property parser.  It is very complicated to
>> >check for invalid .note.gnu.property sections generated by the older
>> >linkers with
>> >the new object.  Kernel loader doesn't want to do it.
>>
>> One way to make things follow the spirit of https://sourceware.org/ml/gnu-gabi/2018-q4/msg00036.html
>>
>> * Define SHT_GNU_PROPERTY
>> * Set sh_type(.note.gnu.property) to SHT_GNU_PROPERTY
>> * Place SHT_GNU_PROPERTY sections in a PT_GNU_PROPERTY segment
>>
>> The generated PT_NOTE will not include .note.gnu.property, so the scheme is compatible with old loaders (ld.so, gdb, Linux, etc).
>> New loaders should interpret PT_GNU_PROPERTY, instead of PT_NOTE.
>>    ( https://patchwork.kernel.org/patch/11285409/ needs no change)
>>
>> This way linkers can keep treating SHT_NOTE sections as opaque and apply "Rules for Linking Unrecognized Sections" (http://www.sco.com/developers/gabi/latest/ch4.sheader.html ) when combining SHT_NOTE sections. At least for lld, there will be no special rules for input SHT_NOTE sections.
>>
>> I will be happy to make changes to lld and LLVM binary utilities if this
>> scheme reaches consensus.
>
>It is kind of too late now.

Better late than never. It is never late to fix the section type if we do intend to fix it.

Loaders don't read sections => the section type change is backward compatible.

On 2020-02-20, H.J. Lu wrote:
>On Thu, Feb 20, 2020 at 1:37 AM Mark Wielaard <mark@klomp.org> wrote:
>>
>> Hi,
>>
>> On Wed, 2020-02-19 at 14:17 -0800, H.J. Lu wrote:
>> > On Wed, Feb 19, 2020 at 1:46 PM Mark Wielaard <mark@klomp.org> wrote:
>> > > This code isn't in the kernel yet. So either it gets changed to use the
>> > > existing scheme with gnu property notes found through PT_NOTE to work
>> > > with existing binaries. Then there is no need for PT_GNU_PROPERTY
>> > > headers.
>> > >
>> > > Or some future kernel will start using PT_GNU_PROPERTY headers to find
>> > > the gnu property notes. But that means it won't work with existing
>> > > binaries that do not have that header. So there is no backwards
>> > > compatibility anyway and we can define SHT_GNU_PROPERTY like above.
>> > >
>> > > So this actually seems the perfect time to make this decision.
>> >
>> > Binaries with .note.gnu.property section have been put into many
>> > OS releases.  We must support them.

We can teach newer assemblers to emit SHT_GNU_PROPERTY.
Newer linkers can support both SHT_GNU_PROPERTY/SHT_NOTE .note.gnu.property

At some point in the future, linkers can drop support for SHT_NOTE .note.gnu.property
Then it will become a graceful degradation: the old SHT_NOTE object files will not be
different from older object files without .note.gnu.property

>> OK. Then it is option 1. The kernel will need to support PT_NOTE for
>> parsing the properties, since such older binaries won't have a
>> PT_GNU_PROPERTY program header. Then we can simply get rid of
>> PT_GNU_PROPERTY since nobody uses it and all information is already
>> available through the PT_NOTE segment.
>>
>
>Kernel loader only checks ld.so and static executable.  Re-link them with
>newer linker will get PT_GNU_PROPERTY.  But ld.so needs to check
>PT_NOTE for older binaries.

The current PT_GNU_PROPERTY usage is all about hints. They are "nice to have" but not
"necessary to have". I don't see any problem teaching newer loaders to forget
PT_NOTE, if we do think PT_GNU_PROPERTY is the way forward.

The currently mixed status is annoying:

glibc: PT_NOTE
Proposed Linux kernel patch: PT_GNU_PROPERTY