RE: the x86 sysret_rip test fails on the Intel FRED architecture

From: "H. Peter Anvin" <hpa@zytor.com>
To: "Li, Xin3" <xin3.li@intel.com>, Brian Gerst <brgerst@gmail.com>,
	"Hansen, Dave" <dave.hansen@intel.com>
Cc: "tglx@linutronix.de" <tglx@linutronix.de>,
	"mingo@redhat.com" <mingo@redhat.com>,
	"bp@alien8.de" <bp@alien8.de>,
	"peterz@infradead.org" <peterz@infradead.org>,
	"dave.hansen@linux.intel.com" <dave.hansen@linux.intel.com>,
	"x86@kernel.org" <x86@kernel.org>,
	"linux-kernel@vger.kernel.org" <linux-kernel@vger.kernel.org>
Subject: RE: the x86 sysret_rip test fails on the Intel FRED architecture
Date: Sat, 21 Jan 2023 19:28:15 -0800	[thread overview]
Message-ID: <C3617458-3629-440D-A914-6792DBA03943@zytor.com> (raw)
In-Reply-To: <SA1PR11MB6734368CB5479E2D1C03C978A8CB9@SA1PR11MB6734.namprd11.prod.outlook.com>

On January 21, 2023 7:01:53 PM PST, "Li, Xin3" <xin3.li@intel.com> wrote:
>> > >> If not intentional, it might be something that can still be fixed.
>> > >> If it is intentional and is going to be with us for a while we have
>> > >> a few options.  If userspace is _really_ depending on this
>> > >> behavior, we could just clobber r11 ourselves in the FRED entry
>> > >> path.  If not, we can remove the assertion in the selftest.
>> > > We can't clobber it in the FRED entry path, since it is common for
>> > > all events, but we could do it in the syscall dispatch.
>> > >
>> > > However, it doesn't seem to make sense to do so to me. The current
>> > > behavior is much more of an artifact than desired behavior.
>> > I guess the SDM statements really are for the kernel's benefit and not
>> > for userspace.  Userspace _should_ be treating SYSCALL like a CALL and
>> > r11 like any old register that can be clobbered.  Right now, the
>> > kernel just happens to clobber it with RFLAGS.
>> >
>> > I do the the odds of anyone relying on this behavior are pretty small.
>> > Let's just zap the check from the selftest, document what we did in
>> > the FRED docs and changelog and move on.
>> 
>> Keep the selftest check, but also accept preserved RCX/R11.  What really matters is
>> that the kernel isn't leaking data.
>
>I feel it the same way, it looks to me that the check is to make sure
>R11 doesn't leak any kernel data because the Linux kernel deliberately
>overwrites R11 with the value of user level flags just before returning
>to user level.
>
>I wanted to zap the check, but as HPA said, this is an artifact to not leak
>any kernel data.  I guess it doesn't make a difference if the kernel sets
>R11 to 0.
>
>Maybe it's still reasonable to keep such a check for IDT.  However, it makes
>no sense for FRED systems, because all GP registers are saved/restored upon
>event delivery/return.
>
>Thanks!
>  Xin
>
>> 
>> --
>> Brian Gerst
>

The big thing is that the system calls that return with sysret v iret on IDT systems need to be consistent, in order to not leak kernel state.