Re: [PATCH] arm64: Add config to limit user space to 47bits

[Ard Biesheuvel]

On 13 July 2016 at 18:14, Alexander Graf <agraf@xxxxxxx> wrote:
> On 07/13/2016 05:59 PM, Ard Biesheuvel wrote:
>>
>> On 13 July 2016 at 17:42, Alexander Graf <agraf@xxxxxxx> wrote:
>>>
>>> Some user space applications are known to break with 48 bits virtual
>>
>> known by whom? At least I wasn't aware of it, so could you please
>> share some examples?
>
>
> Sure! Known to me so far are:
>
>   * mozjs17
>   * mozjs24
>   * mozjs38
>   * js-1.8.5
>   * java-1.7 (older JITs, fixed in newer ones)
>

Ah, right. The JIT stuff ... I did read about that before, but thanks
for refreshing my memory

> I'm not sure if there are more, but the fact that I've run into this problem
> more than once doesn't make me incredibly happy :).
>

There is little point in reiterating over all the opinions that have
been shared, but for the record, I agree that we should *not* cater
for horrible hacks in JIT code. The only thing we should attempt to
deal with on the kernel side is VA space size related differences
between architectures that have some base in language or ABI specs,
and this is clearly not the case here.

It does bring up another interesting point, though. The size of the
userland VA space simply follows the size of the kernel VA space, and
the latter is chosen to be 48 bits on recent defconfig kernels to
accommodate platforms that have a 512 GB hole in the middle of their
physical address space. If the userland address space is populated
very sparsely, this may have a measurable performance impact
(especially under virt, due to the way each level of translation at
stage 1 triggers a full page table walk at stage 2), and so there may
be a case to be made for lowering TASK_SIZE for other reasons than the
one given in this patch. Or at least be able to interrogate the OS for
the size of the VA space ...

>>
>>> address space. As interim step until the world is healed and everyone
>>> embraces correct code, this patch allows to only expose 47 bits of
>>> virtual address space to user space.
>>>
>> Is this a code generation/toolchain issue?
>
>
> mozjs uses a single 64bit value to combine doubles, ints and pointers into a
> single variable. It is very smart and uses the upper 17 bits for metadata
> such as "which type of variable is this". Coincidentally those bits happen
> to overlap the "double is an infinite number" bits, so that you can also
> express a NaN with it. When using such a value, the upper 17 bits get masked
> out.
>
> That one was fixed upstream by force allocating the javascript heap starting
> at a fixed location which is below 47 bits.
>
> js-1.8.5 has the same as above, but also uses pointers to .rodata as
> javascript pointers, so it doesn't only use the heap, it also uses pointers
> to the library itself, which gets mapped high up the address space. I don't
> have a solution for that one yet.
>
> IcedTea for java-1.7 had a bug where it incorrectly caused an overflow when
> trying to calculating a relative adrp offset from <address high up> to
> <address really low>, so that the resulting pointer had the upper bits set
> as 1s. That one is long fixed upstream, we only ran into it because we used
> an ancient IcedTea snapshot.
>
> My main concern however is with code that I do not know is broken today.
>

I really think papering over it like this is not the way to go.

Thanks,
Ard.