Re: [PATCH] Fix resume on x86-32 machines

[Linus Torvalds]

On Mon, Dec 11, 2017 at 6:22 AM, Rafael J. Wysocki <rjw@xxxxxxxxxxxxx> wrote:
> On Sunday, December 10, 2017 10:58:23 PM CET Andy Lutomirski wrote:
>>
>> I'm guessing that the real issue is that 32-bit needs %fs restored early for TLS.
>
> I *think* you are right.
>
> Anyway, that should be easy enough to verify.
>
> Pavel, can you please check if the below change works too?

So Jarkko confirmed this works for him, but the more I look at this
crap, the less I like it.

Why do we save fs/ds/es/ss at all on x86-32? Don't they all have fixed
values in the kernel, with %fs being __KERNEL_PERCPU, and the others
being __USER_DS?

Nothing else can possibly be valid, as far as I can tell.

I think we actually leave the user-space percpu segment in %gs (or the
stack canary base), so that one we should actually save/restore, but
I'm getting the feeling that we should just reset the other segment
registers to known values on 32-bit.

Also, why does the 32-bit code do

        loadsegment(es, ctxt->es);

but the 64-bit code does

        asm volatile ("movw %0, %%es" :: "r" (ctxt->es));

And look at that confusion between MSR_GS_BASE and MSR_KERNEL_GS_BASE
all within the 64-bit case.

In particular, note how we reload the %gs segment in between the two -
wouldn't that mess with the currently active gs base if %gs can be
non-zero?

Christ, what a mess.

So I think that whole sequence is garbage. It has been written as some
kind of "save and restore registers", but that's not what it really
then does - or what it should do.

It should make sure to restore a sane kernel state, not some random
register state.

And the 32-bit and 64-bit code really should strive to be at least
_sanely_ different, not this randomly and insanely different mess.

But yes, Rafael's patch looks like the minimal one-liner. But I think
we should do the %gs load early too for the 32-bit stack canary case,
kind of like we need to do %fs for percpu base.

                  Linus