On Mon, Jan 22, 2024 at 03:01:06PM -0600, Eric W. Biederman wrote:
> Kees Cook <keescook@xxxxxxxxxxxx> writes:
>
> > On Mon, Jan 22, 2024 at 10:43:59AM -0600, Eric W. Biederman wrote:
> >> Jan Bujak <j@xxxxxxx> writes:
> >>
> >> > Hi.
> >> >
> >> > I recently updated my kernel and one of my programs started segfaulting.
> >> >
> >> > The issue seems to be related to how the kernel interprets PT_LOAD headers;
> >> > consider the following program headers (from 'readelf' of my reproduction):
> >> >
> >> > Program Headers:
> >> > Type Offset VirtAddr PhysAddr FileSiz MemSiz Flg Align
> >> > LOAD 0x001000 0x10000 0x10000 0x000010 0x000010 R 0x1000
> >> > LOAD 0x002000 0x11000 0x11000 0x000010 0x000010 RW 0x1000
> >> > LOAD 0x002010 0x11010 0x11010 0x000000 0x000004 RW 0x1000
> >> > LOAD 0x003000 0x12000 0x12000 0x0000d2 0x0000d2 R E 0x1000
> >> > LOAD 0x004000 0x20000 0x20000 0x000004 0x000004 RW 0x1000
> >> >
> >> > Old kernels load this ELF file in the following way ('/proc/self/maps'):
> >> >
> >> > 00010000-00011000 r--p 00001000 00:02 131 ./bug-reproduction
> >> > 00011000-00012000 rw-p 00002000 00:02 131 ./bug-reproduction
> >> > 00012000-00013000 r-xp 00003000 00:02 131 ./bug-reproduction
> >> > 00020000-00021000 rw-p 00004000 00:02 131 ./bug-reproduction
> >> >
> >> > And new kernels do it like this:
> >> >
> >> > 00010000-00011000 r--p 00001000 00:02 131 ./bug-reproduction
> >> > 00011000-00012000 rw-p 00000000 00:00 0
> >> > 00012000-00013000 r-xp 00003000 00:02 131 ./bug-reproduction
> >> > 00020000-00021000 rw-p 00004000 00:02 131 ./bug-reproduction
> >> >
> >> > That map between 0x11000 and 0x12000 is the program's '.data' and '.bss'
> >> > sections to which it tries to write to, and since the kernel doesn't map
> >> > them anymore it crashes.
> >> >
> >> > I bisected the issue to the following commit:
> >> >
> >> > commit 585a018627b4d7ed37387211f667916840b5c5ea
> >> > Author: Eric W. Biederman <ebiederm@xxxxxxxxxxxx>
> >> > Date: Thu Sep 28 20:24:29 2023 -0700
> >> >
> >> > binfmt_elf: Support segments with 0 filesz and misaligned starts
> >> >
> >> > I can confirm that with this commit the issue reproduces, and with it
> >> > reverted it doesn't.
> >> >
> >> > I have prepared a minimal reproduction of the problem available here,
> >> > along with all of the scripts I used for bisecting:
> >> >
> >> > https://github.com/koute/linux-elf-loading-bug
> >> >
> >> > You can either compile it from source (requires Rust and LLD), or there's
> >> > a prebuilt binary in 'bin/bug-reproduction` which you can run. (It's tiny,
> >> > so you can easily check with 'objdump -d' that it isn't malicious).
> >> >
> >> > On old kernels this will run fine, and on new kernels it will
> >> > segfault.
> >>
> >> Frankly your ELF binary is buggy, and probably the best fix would be to
> >> fix the linker script that is used to generate your binary.
> >>
> >> The problem is the SYSV ABI defines everything in terms of pages and so
> >> placing two ELF segments on the same page results in undefined behavior.
> >>
> >> The code was fixed to honor your .bss segment and now your .data segment
> >> is being stomped, because you defined them to overlap.
> >>
> >> Ideally your linker script would place both your .data and .bss in
> >> the same segment. That would both fix the issue and give you a more
> >> compact elf binary, while not changing the generated code at all.
> >>
> >>
> >> That said regressions suck and it would be good if we could update the
> >> code to do something reasonable in this case.
> >>
> >> We can perhaps we can update the .bss segment to just memset an existing
> >> page if one has already been mapped. Which would cleanly handle a case
> >> like yours. I need to think about that for a moment to see what the
> >> code would look like to do that.
> >
> > It's the "if one has already been mapped" part which might
> > become expensive...
>
> I am wondering if perhaps we can add MAP_FIXED_NOREPLACE and take
> some appropriate action if there is already a mapping there.
Yeah, in the general case we had to back out MAP_FIXED_NOREPLACE usage
for individual LOADs because there were so many cases of overlapping
LOADs. :( Currently it's only used during the initial mapping (when
"total_size" is set), to avoid colliding with the stack.
But, as you suggest, if we only use it for filesz==0, it could work.
> Such as printing a warning and skipping the action entirely for
> a pure bss segment. That would essentially replicate the previous
> behavior.
Instead of failing, perhaps we just fallback to not using
MAP_FIXED_NOREPLACE and do the memset? (And maybe pr_warn_once?)
> At a minimum adding MAP_FIXED_NOREPLACE should allow us to
> deterministically detect and warn about problems, making it easier
> for people to understand why their binary won't run.
Yeah, it seems like it's the vm_brk_flags() that is clobber the mapping,
so we have to skip that for the MAP_FIXED_NOREPLACE fails on a filesz==0
case?
--
Kees Cook
