Andy Lutomirski <luto@xxxxxxxxxx> writes:
>> On Jul 23, 2020, at 1:22 PM, Thomas Gleixner <tglx@xxxxxxxxxxxxx> wrote:
>> Andy Lutomirski <luto@xxxxxxxxxxxxxx> writes:
>>> My suggestion is to enlarge pt_regs. The save and restore logic can
>>> probably be in C, but pt_regs is the logical place to put a register
>>> that is saved and restored across all entries.
>>
>> Kinda, but that still sucks because schedule from #PF will get it wrong
>> unless you do extra nasties.
>
> This seems like we’re reinventing the wheel. PKRS is not
> fundamentally different from, say, RSP. If we want to save it across
> exceptions, we save it on entry and context-switch-out and restore it
> on exit and context-switch-in.
It's fundamentally different from RSP because it has state (refcount)
attached, which RSP clearly has not. If you get rid of the state then
yes.
>>> Whoever does this work will have the delightful job of figuring out
>>> whether BPF thinks that the layout of pt_regs is ABI and, if so,
>>> fixing the resulting mess.
>>>
>>> The fact the new fields will go at the beginning of pt_regs will make
>>> this an entertaining prospect.
>>
>> Good luck with all of that.
>
> We can always cheat like this:
>
> struct real_pt_regs {
> unsigned long pkrs;
> struct pt_regs regs;
> };
>
> and pass a pointer to regs around. What BPF doesn't know about can't hurt it.
Yes, but that's the easy part of the problem :)
Thanks,
tglx
