Re: [PATCH bpf] bpf: verifier: prevent userspace memory access

[Puranjay Mohan]

Ilya Leoshkevich <iii@xxxxxxxxxxxxx> writes:

> On Wed, Mar 20, 2024 at 11:08:00PM -0700, Alexei Starovoitov wrote:
>> On Wed, Mar 20, 2024 at 3:55 AM Puranjay Mohan <puranjay12@gmailcom> wrote:
>> >
>> > The JITs need to implement bpf_arch_uaddress_limit() to define where
>> > the userspace addresses end for that architecture or TASK_SIZE is taken
>> > as default.
>> >
>> > The implementation is as follows:
>> >
>> > REG_AX =  SRC_REG
>> > if(offset)
>> >         REG_AX += offset;
>> > REG_AX >>= 32;
>> > if (REG_AX <= (uaddress_limit >> 32))
>> >         DST_REG = 0;
>> > else
>> >         DST_REG = *(size *)(SRC_REG + offset);
>> 
>> The patch looks good, but it seems to be causing s390 CI failures.
>> 
>> Ilya,
>> could you help us understand is this check needed on s390
>> and if so, what should be the uaddress_limit ?
>
> s390x does not define ARCH_HAS_NON_OVERLAPPING_ADDRESS_SPACE.
> Userspace and kernel run in completely different and isolated address
> spaces, so it's not possible to determine from a pointer value whether
> it's a user or a kernel pointer.
> But the good news is that whatever you deference without using
> special instruction sequences will refer to the kernel address space.
> So I wonder if we could somehow disable this check on s390x altogether?
> And if we are not sure whether it's a valid pointer, use BPF_PROBE_MEM
> as always.

Thanks for the details. I understand that s390x doesn't need this extra
check because all normal accesses are in the kernel address space and they
will be marked with BPF_PROBE_MEM by the verifier if the pointer is
untrusted.

I have sent v2 of this patch with this check disabled on s390x
https://lore.kernel.org/bpf/20240321101058.68530-1-puranjay12@xxxxxxxxx/

Thanks,
Puranjay