MIME-Version: 1.0
References: <20230307023946.14516-1-xin3.li@intel.com> <20230307023946.14516-29-xin3.li@intel.com>
 <CAJhGHyC6LgCwdDTkiy2TaQ8wzBQQfrx8ni7fY8vH-bUT2kR8pg@mail.gmail.com> <ed318bd6-25b2-efcf-0cc4-c57699f6654a@citrix.com>
In-Reply-To: <ed318bd6-25b2-efcf-0cc4-c57699f6654a@citrix.com>
From:   Lai Jiangshan <jiangshanlai@gmail.com>
Date:   Fri, 17 Mar 2023 21:02:52 +0800
Message-ID: <CAJhGHyAdyqWp_W3yHMn0euYG9TMctpSmNnqr_e+=FR3rN5UzpA@mail.gmail.com>
Subject: Re: [PATCH v5 28/34] x86/fred: fixup fault on ERETU by jumping to fred_entrypoint_user
To:     andrew.cooper3@citrix.com
Cc:     Xin Li <xin3.li@intel.com>, linux-kernel@vger.kernel.org,
        x86@kernel.org, kvm@vger.kernel.org, tglx@linutronix.de,
        mingo@redhat.com, bp@alien8.de, dave.hansen@linux.intel.com,
        hpa@zytor.com, peterz@infradead.org, seanjc@google.com,
        pbonzini@redhat.com, ravi.v.shankar@intel.com
Content-Type: text/plain; charset="UTF-8"
Content-Transfer-Encoding: quoted-printable
Precedence: bulk

On Fri, Mar 17, 2023 at 5:56=E2=80=AFPM <andrew.cooper3@citrix.com> wrote:
>
> On 17/03/2023 9:39 am, Lai Jiangshan wrote:
> >> +#ifdef CONFIG_X86_FRED
> >> +static bool ex_handler_eretu(const struct exception_table_entry *fixu=
p,
> >> +                            struct pt_regs *regs, unsigned long error=
_code)
> >> +{
> >> +       struct pt_regs *uregs =3D (struct pt_regs *)(regs->sp - offset=
of(struct pt_regs, ip));
> >> +       unsigned short ss =3D uregs->ss;
> >> +       unsigned short cs =3D uregs->cs;
> >> +
> >> +       fred_info(uregs)->edata =3D fred_event_data(regs);
> >> +       uregs->ssx =3D regs->ssx;
> >> +       uregs->ss =3D ss;
> >> +       uregs->csx =3D regs->csx;
> >> +       uregs->current_stack_level =3D 0;
> >> +       uregs->cs =3D cs;
> > Hello
> >
> > If the ERETU instruction had tried to return from NMI to ring3 and just=
 faulted,
> > is NMI still blocked?
> >
> > We know that IRET unconditionally enables NMI, but I can't find any clu=
e in the
> > FRED's manual.
> >
> > In the pseudocode of ERETU in the manual, it seems that NMI is only ena=
bled when
> > ERETU succeeds with bit28 in csx set.  If so, this code will fail to re=
enable
> > NMI if bit28 is not explicitly re-set in csx.
>
> IRET clearing NMI blocking is the source of an immense amount of grief,
> and ultimately the reason why Linux and others can't use supervisor
> shadow stacks at the moment.
>
> Changing this property, so NMIs only get unblocked on successful
> execution of an ERET{S,U}, was a key demand of the FRED spec.
>
> i.e. until you have successfully ERET*'d, you're still logically in the
> NMI handler and NMIs need to remain blocked even when handling the #GP
> from a bad ERET.
>

Handling of the #GP for a bad ERETU can be rescheduled. It is not
OK to reschedule with NMI blocked.

I think "regs->nmi =3D 1;" (not uregs->nmi) can fix the problem.