Re: [PATCH v8 0/6] Introduce CET supervisor state support

[Chang S. Bae]

On 6/2/2025 11:22 PM, Chao Gao wrote:
> Aside from the splat, task #2 could execute AMX instructions without
> requesting permissions, but its AMX state would be discarded during the
> next FPU switch, as RFBM[18] is cleared when executing XSAVES. And, in the

Right, AMX instructions can be executed when XFD is disarmed. But in 
this case, it's inside a signal handler. On sigreturn, XTILE_DATA will 
be reloaded with the init state, since fpstate::user_xfeatures[18] is zero.

> "flipped" scenario you mentioned, task #2 might receive an extra #NM, after
> which its fpstate would be re-allocated (although the size won't increase
> further).

Yes.

> So, for well-behaved tasks that never use AMX, there is no impact; tasks
> that use AMX may receive extra #NM. There won't be any unexpected #PF,
> memory corruption, or kernel panic.

A signal handler is expected to stay within the bounds of 
async-signal-safe functions, so using AMX in that context is highly 
unlikely in practice. While the issue has existed, its real-world impact 
appears quite minimal in my view.

Thanks,
Chang