Re: [PATCH v29 23/32] x86/cet/shstk: Add user-mode shadow stack support

[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

 



First of all,

thanks a lot Dave for taking the time to communicate properly with me!

On Fri, Aug 27, 2021 at 01:25:29PM -0700, Dave Hansen wrote:
> I don't think this has anything to do with context-switching, really.
> 
> The code lands in shstk_setup() which wants to make sure that the new
> MSR values are set before the task goes out to userspace.  If
> TIF_NEED_FPU_LOAD was set, it could do that by going out to the XSAVE
> buffer and setting the MSR state in the buffer.  Before returning to
> userspace, it would be XRSTOR'd.  A WRMSR by itself would not be
> persistent because that XRSTOR would overwrite it.
> 
> But, if TIF_NEED_FPU_LOAD is *clear* it means the XSAVE buffer is
> out-of-date and the registers are live.  WRMSR can be used and there
> will be a XSAVE* to the task buffer during a context switch.
> 
> So, this code takes the coward's way out: it *forces* TIF_NEED_FPU_LOAD
> to be clear by making the registers live with fpregs_restore_userregs().
>  That lets it just use WRMSR instead of dealing with the XSAVE buffer
> directly.  If it didn't do this with the *WHOLE* set of user FPU state,
> we'd need more fine-granted "NEED_*_LOAD" tracking than our one FPU bit.
> 
> This is also *only* safe because the task is newly-exec()'d and the FPU
> state was just reset.  Otherwise, we might have had to worry that the
> non-PL3 SSPs have garbage or that non-SHSTK bits are set in MSR_IA32_U_CET.
> 
> That said, after staring at it, I *think* this code is functionally
> correct and OK performance-wise.

Right, except that that is being done in
setup_signal_shadow_stack()/restore_signal_shadow_stack() too, for the
restore token.

Which means, a potential XRSTOR each time just for a single MSR. That
means, twice per signal in the worst case.

Which means, shadow stack should be pretty noticeable in signal-heavy
benchmarks...

> I suspect that the (very blunt) XRSTOR inside of
> start_update_msrs()->fpregs_restore_userregs() is quite rare because
> TIF_NEED_FPU_LOAD will usually be clear due to the proximity to
> execve(). So, adding direct XSAVE buffer manipulation would probably
> only make it more error prone.

@Yu-cheng: please take Dave's explanation as is and stick it over
start_update_msrs() so that it is clear what that thing is doing.

Thx.

-- 
Regards/Gruss,
    Boris.

https://people.kernel.org/tglx/notes-about-netiquette



[Index of Archives]     [Kernel Newbies]     [Security]     [Netfilter]     [Bugtraq]     [Linux FS]     [Yosemite Forum]     [MIPS Linux]     [ARM Linux]     [Linux Security]     [Linux RAID]     [Samba]     [Video 4 Linux]     [Device Mapper]     [Linux Resources]

  Powered by Linux