Re: [PATCH] x86: Add a GNU_PROPERTY_X86_ISA_1_USED note if needed

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

On Fri, Jul 20, 2018 at 6:31 AM, Rich Felker <dalias@libc.org> wrote:
> On Fri, Jul 20, 2018 at 05:24:08AM -0700, H.J. Lu wrote:
>> On Fri, Jul 20, 2018 at 2:30 AM, Szabolcs Nagy <szabolcs.nagy@arm.com> wrote:
>> > On 19/07/18 22:33, H.J. Lu wrote:
>> >>
>> >> Here is the updated patch.  Any comments?
>> >>
>> >>
>> >> H.J.
>> >> --
>> >> When -z separate-code, which is enabled by default for Linux/x86, is
>> >> used to create executable, ld won't place any data in the code-only
>> >> PT_LOAD segment.  If there are no data sections placed before the
>> >> code-only PT_LOAD segment, the program headers won't be mapped into
>> >> any PT_LOAD segment.  When the executable tries to access it (based
>> >> on the program header address passed in AT_PHDR), it will lead to
>> >> segfault.  This patch inserts a GNU_PROPERTY_X86_ISA_1_USED note if
>> >> there may be no data sections before the text section so that the
>> >> first PT_LOAD segment won't be code-only and will contain the program
>> >> header.
>> >
>> >
>> > i agree with Rich's comment on the bugzilla ticket:
>> > simply always adding a .phdrs section that covers
>> > the program headers would ensure that this kind of
>> > bug is not introduced on other targets and seems
>> > less hackish than adding a note.
>>
>> We can explore it.
>>
>> > i don't know why there is no such section already,
>> > or if there is any drawback adding such a section,
>> > is the requirement to access program headers in a
>> > static linked executable a linux only thing?
>>
>> The program headers aren't required to be mapped into
>> process image.  It may not be correct for Linux kermel
>> to set AT_PHDR to an unmapped address.  But I'd like
>> to support the existing kernel.  I am going to check in
>> my change ASIS and backport it to 2.31 branch.
>
> They are required to be mapped, for at least these reasons:
>
> 1. PT_TLS is the only way the runtime can find the program's TLS
>    image.
>
> 2. They're exposed as a public interface via dl_iterate_phdr, and the
>    AT_PHDR auxv entry.

Only this one is relevant here.   The others don't need program headers
mapped into process image.

> 3. PT_GNU_EH_FRAME is the way (or at least one of the ways) to find
>    unwind information.
>
> 4. PT_LOAD segments are needed for dladdr to determine if an address
>    is within the bounds of an object's memory.
>
> 5. PT_GNU_RELRO is necessary to mprotect the relro zone read-only
>    after processing relocations.
>
> 6. PT_DYNAMIC is the way to find _DYNAMIC in a dynamic-linked program
>    and the cleanest way to find the load address for static PIE
>    programs.
>
> It's true that *some* of the things in the program headers are mainly
> needed prior to passing control from the kernel into the mapped
> program, but *a lot* are needed at runtime to implement important
> features.
>
> Some of them could be replaced for static linking with auto-generated
> symbols, but that's not existing practice, and it's probably good that
> it's not since compilers will tend to make transformations that
> invalidate start/end symbol hacks.
>
> Rich



-- 
H.J.