Received-SPF: pass (google.com: domain of linux-kernel-owner@vger.kernel.org designates 23.128.96.18 as permitted sender) client-ip=23.128.96.18;
Content-Type: text/plain; charset=utf-8
Content-Transfer-Encoding: quoted-printable
From:   Andy Lutomirski <luto@amacapital.net>
Mime-Version: 1.0 (1.0)
Date:   Sat, 18 Apr 2020 13:30:05 -0700
Subject: Re: [PATCH] x86/memcpy: Introduce memcpy_mcsafe_fast
Message-Id: <67FF611B-D10E-4BAF-92EE-684C83C9107E@amacapital.net>
Cc:     Dan Williams <dan.j.williams@intel.com>,
        Thomas Gleixner <tglx@linutronix.de>,
        Ingo Molnar <mingo@redhat.com>, X86 ML <x86@kernel.org>,
        stable <stable@vger.kernel.org>, Borislav Petkov <bp@alien8.de>,
        "H. Peter Anvin" <hpa@zytor.com>,
        Peter Zijlstra <peterz@infradead.org>,
        Tony Luck <tony.luck@intel.com>,
        Erwin Tsaur <erwin.tsaur@intel.com>,
        Linux Kernel Mailing List <linux-kernel@vger.kernel.org>,
        linux-nvdimm <linux-nvdimm@lists.01.org>
To:     Linus Torvalds <torvalds@linux-foundation.org>
Sender: linux-kernel-owner@vger.kernel.org
Precedence: bulk


--Andy

> On Apr 18, 2020, at 12:42 PM, Linus Torvalds <torvalds@linux-foundation.or=
g> wrote:
>=20
>>> On Fri, Apr 17, 2020 at 5:12 PM Dan Williams <dan.j.williams@intel.com> w=
rote:
>>>=20
>>> @@ -106,12 +108,10 @@ static __always_inline __must_check unsigned long
>>> memcpy_mcsafe(void *dst, const void *src, size_t cnt)
>>> {
>>> #ifdef CONFIG_X86_MCE
>>> -       i(static_branch_unlikely(&mcsafe_key))
>>> -               return __memcpy_mcsafe(dst, src, cnt);
>>> -       else
>>> +       if (static_branch_unlikely(&mcsafe_slow_key))
>>> +               return memcpy_mcsafe_slow(dst, src, cnt);
>>> #endif
>>> -               memcpy(dst, src, cnt);
>>> -       return 0;
>>> +       return memcpy_mcsafe_fast(dst, src, cnt);
>>> }
>=20
> It strikes me that I see no advantages to making this an inline function a=
t all.
>=20
> Even for the good case - where it turns into just a memcpy because MCE
> is entirely disabled - it doesn't seem to matter.
>=20
> The only case that really helps is when the memcpy can be turned into
> a single access. Which - and I checked - does exist, with people doing
>=20
>        r =3D memcpy_mcsafe(&sb_seq_count, &sb(wc)->seq_count, sizeof(uint6=
4_t));
>=20
> to read a single 64-bit field which looks aligned to me.
>=20
> But that code is incredible garbage anyway, since even on a broken
> machine, there's no actual reason to use the slow variant for that
> whole access that I can tell. The macs-safe copy routines do not do
> anything worthwhile for a single access.

Maybe I=E2=80=99m missing something obvious, but what=E2=80=99s the alternat=
ive?  The _mcsafe variants don=E2=80=99t just avoid the REP mess =E2=80=94 t=
hey also tell the kernel that this particular access is recoverable via exta=
ble. With a regular memory access, the CPU may not explode, but do_machine_c=
heck() will, at very best, OOPS, and even that requires a certain degree of o=
ptimism.  A panic is more likely.