Follow @Openwall on Twitter for new release announcements and other news
[<prev] [next>] [<thread-prev] [thread-next>] [day] [month] [year] [list]
Date: Fri, 07 Jan 2022 14:58:54 +0100
From: Florian Weimer <>
To: Rich Felker <>
Cc: Nihal Jere <>,
Subject: Re: Dynamic linker segfault

* Rich Felker:

>> With a p_align < PAGESIZE check in place, portable binaries need to use
>> the value 65536.  When running with page size 4096, the loader cannot
>> know whether p_align was set to this value merely to satisfy the p_align
>> < PAGESIZE check, or because there is actually some section alignment
>> that requires 65536 byte alignment.  There is no kernel interface to
>> request 65536 byte alignment, so the loader has to do extra work to
>> satisfy this request.  And in the first case (no actual 65536 byte
>> alignment requirement), that work is unnecessary.
> Unfortunately it's impossible to distinguish between such an alignment
> requirement and __attribute__((__aligned__(65536))) appearing in
> section contents that went into the segment, so disregarding it is a
> bug (one musl also has) I think.


>> > In any case, do you know if this test file is somehow related to that
>> > work, or is it just a guess? It doesn't seem to be related to me since
>> > it's essentially a "pageless" mapping setup.
>> The glibc test seems to be just buggy: First we verify p_align against
>> the page size, then we use that p_align value to check the alignment of
>> the PT_LOAD segment (mainly file offset congruency).  The p_align check
>> against the page size looks completely optional to me if we check file
>> offset congruency directly against the run-time page size.
>> The ELF specification explicitly describes the p_align values 0 and 1 as
>> valid, indicating no alignment constraint.  So a p_align < PAGESIZE
>> check is buggy in that regard as well.  This also conflicts with your
>> interpretation as p_align as the maximum supported page size.
> The ELF specification describes syntax not semantic requirements on
> the platform to support anything the syntax can represent.

That's not entirely accurate, there are exceptions.  p_align seems to be
one such exception:

    As ``Program Loading'' describes in this chapter of the processor
    supplement, loadable process segments must have congruent values for
    p_vaddr and p_offset, modulo the page size. This member gives the
    value to which the segments are aligned in memory and in the
    file. Values 0 and 1 mean no alignment is required. Otherwise,
    p_align should be a positive, integral power of 2, and p_vaddr
    should equal p_offset, modulo p_align.

Most processor supplements do not discuss p_align unfortunately?

> The semantics of the above can't be honored (without memcpy instead of
> mmap, which is really something you don't want to support) if they
> produce congruences incompatible with the layout, and they can't be
> honored at all if they're incompatible with the permissions
> requirements. Even when they don't conflict with the permissions
> requirements, they may expose unintended mapped data (which for users
> wanting separate-code is problematic, since it introduces gadgets). So
> I'm not convinced it should be supported even when it "can".

The congruency compatibility is a property of the file layout, not the
p_align value.  If the file layout is incompatible, changing p_align
doesn't make the object loadable.

As far as I can tell, p_align is not at all useful, except for (e.g.)
__attribute__((__aligned__(65536))) for global.  But even that wasn't
implemented widely.


Powered by blists - more mailing lists

Confused about mailing lists and their use? Read about mailing lists on Wikipedia and check out these guidelines on proper formatting of your messages.