On Wed, May 26, 2021, at 2:14 PM, Babu Moger wrote:
>
>
> On 5/25/21 7:36 PM, Andy Lutomirski wrote:
> > On 5/25/21 2:37 PM, Babu Moger wrote:
> >
> >> Investigation so far.
> >> 1. The test fails on AMD(Milan) systems. Test appears to pass on Intel
> >> systems. Tested on old Intel system available here.
> >
> > Ten cents says that you have discovered that AMD returns a synthesized
> > value for XINUSE[PKRU] and that write_pkru() is buggy.
> >
> > Any volunteers to fix the if (!pk) case in write_pkru() and see if the
> > problem goes away? (Or at least to print something for the if (!pk)
> > case and seee if it gets printed?)
>
> I inserted this patch to test this path.
>
>
> diff --git a/arch/x86/include/asm/pgtable.h b/arch/x86/include/asm/pgtable.h
> index b1099f2d9800..a0bca917f864 100644
> --- a/arch/x86/include/asm/pgtable.h
> +++ b/arch/x86/include/asm/pgtable.h
> @@ -151,6 +160,8 @@ static inline void write_pkru(u32 pkru)
> fpregs_lock();
> if (pk)
> pk->pkru = pkru;
> + else
> + printk_once("%s read_pkru 0x%x\n",__FUNCTION__, read_pkru());
> __write_pkru(pkru);
> fpregs_unlock();
> }
>
>
> Yes. The pk value is sometimes NULL and the value in the pkru register is
> 0 at that time. Here is the dmesg output.
>
> [ 1.700889] write_pkru read_pkru 0x0
>
Sweet, maybe I get my hypothetical dime back! Both bugs are real.
Dave and I (mostly Dave — I’m supplying the snark) are working on an omnibus patch to fix this whole mess.
Thanks,
Andy
