Re: git: e85eaa930862 - main - Have rtld query the page size from the kernel

On 11/19/24 22:30, Konstantin Belousov wrote:
> On Thu, Nov 14, 2024 at 09:47:30AM -0500, John Baldwin wrote:
>> On 11/13/24 12:10, Jessica Clarke wrote:
>>> On 13 Nov 2024, at 19:44, John Baldwin <jhb@FreeBSD.org> wrote:
>>>>
>>>> On 4/7/22 07:38, Andrew Turner wrote:
>>>>> The branch main has been updated by andrew:
>>>>> URL: https://cgit.FreeBSD.org/src/commit/?id=e85eaa930862d5b4dc917bc31e8d7254a693635d
>>>>> commit e85eaa930862d5b4dc917bc31e8d7254a693635d
>>>>> Author:     Andrew Turner <andrew@FreeBSD.org>
>>>>> AuthorDate: 2022-04-04 15:05:40 +0000
>>>>> Commit:     Andrew Turner <andrew@FreeBSD.org>
>>>>> CommitDate: 2022-04-07 14:37:37 +0000
>>>>>       Have rtld query the page size from the kernel
>>>>>            To allow for a dynamic page size on arm64 have the runtime linker
>>>>>       query the kernel for the currentl page size.
>>>>>            Reviewed by:    kib
>>>>>       Sponsored by:   The FreeBSD Foundation
>>>>>       Differential Revision: https://reviews.freebsd.org/D34765
>>>>
>>>> This broke relro handling for rtld.  The reason is that init_pagesizes() is
>>>> called after parsing the program headers for rltd in init_rtld().  As a result,
>>>> page_size is 0 when rtld_round_page() is called so the relro_size is 0.  The
>>>> RTLD_INIT_EARLY_PAGESIZES case was for ia64, and in the early case it's probably
>>>> not safe to call sysctl?  If it is safe to call sysctl, we could just always
>>>> init pagesizes early?
>>>
>>> It looks like there are a few things going on:
>>>
>>> 1. relocate_object calls obj_enforce_relro if !obj->mainprog, so will
>>> try to enforce RELRO for RTLD itself whilst page_size is 0
>>>
>>> 2. init_rtld later calls obj_enforce_relro for obj_rtld, after
>>> page_size has been initialised
>>>
>>> 3. init_rtld is careful to avoid using global variables until it’s
>>> called relocate_objects for RTLD itself, but by hiding accesses to
>>> page_size away in rtld_*_page that’s no longer true (definitely not
>>> true in the case of text relocations, for example, though whether it
>>> also occurs for other cases we care more about I don’t know)
>>>
>>> So I think there are a couple of things to fix:
>>>
>>> 1. Stop accessing page_size prior to relocate_objects returning for
>>> RTLD itself
>>>
>>> 2. Stop enforcing RELRO twice for RTLD (e.g. add && obj != rtldobj to
>>> relocate_object’s case)
>>>
>>> At least, that’s what I’ve inferred from reading the code.
>>>
>>> Though, to be honest, things might be rather nicer if we just made
>>> .rtld_start responsible for relocating RTLD itself prior to calling
>>> init_rtld, that’s what we have to do for CHERI, as do arm, powerpc and
>>> powerpc64, and it means you can use globals from the start in init_rtld.
>>
>> I've done 2) locally which fixed my immediate issue.
> Can you provide some more info please?

I have a local patch that supports mutiple PT_GNU_RELRO segments, and to
do that it removes relro_* from Obj_Entry and just walks the list of
phdrs in obj_remap_relro().  However, that defeats the order you mention
below of the timing of parse_rtld_phdr().  I hadn't realized until your
paragraph below that that local change was part of why this was exposed.
It would still be more correct, strictly speaking, to skip obj_enforce_relro
for rtldobj, and if we ever supported multiple PT_GNU_RELRO upstream we
would want that change.

I had also missed the later explicit call to obj_enforce_relro() for obj_rtld
when I sent my original e-mail, so nothing is broken in FreeBSD.

> obj_enforce_relro() only acts if obj_relro_size is > 0, which requires
> the call to parse_rtld_phdr() to fill it.  This is indeed called after the
> rtld is relocated by relocate_object(s), but so it was before the Andrew'
> change.
> 
> I can see a problem if plt is mapped ro and requires relro to remap it
> rw before relocation, but I do not believe this is the case for rtld and
> any static linker.
> 
> And what is the arch?

I was testing on a CHERI arch since that is my use case for multiple
PT_GNU_RELRO sections, though the change to obj_remap_relro() is applicable
to all architectures.

> The RTLD_INIT_PAGESIZES_EARLY symbol indeed should be eliminated, I have
> the patch. But let's sort out the page size issue as well.

I do think we can probably just axe that code entirely.

>> I agree though that having all arches implement rtld_relocate_nonplt_self()
>> which is called before rtld() would be nicer.

This might still be a nice thing to do.

-- 
John Baldwin