Received-SPF: pass (google.com: best guess record for domain of linux-kernel-owner@vger.kernel.org designates 209.132.180.67 as permitted sender) client-ip=209.132.180.67;
Content-Type: text/plain; charset=utf-8
Content-Transfer-Encoding: quoted-printable
From:   Andy Lutomirski <luto@amacapital.net>
Mime-Version: 1.0 (1.0)
Subject: Re: [PATCH v10 6/6] x86/split_lock: Enable split lock detection by kernel parameter
Date:   Thu, 21 Nov 2019 11:01:39 -0800
Message-Id: <4BB1CB74-887B-4F40-B3B2-F0147B264C34@amacapital.net>
References: <20191121185303.GB199273@romley-ivt3.sc.intel.com>
Cc:     Andy Lutomirski <luto@kernel.org>,
        David Laight <David.Laight@aculab.com>,
        Ingo Molnar <mingo@kernel.org>,
        Peter Zijlstra <peterz@infradead.org>,
        Thomas Gleixner <tglx@linutronix.de>,
        Ingo Molnar <mingo@redhat.com>, Borislav Petkov <bp@alien8.de>,
        H Peter Anvin <hpa@zytor.com>,
        Tony Luck <tony.luck@intel.com>,
        Ashok Raj <ashok.raj@intel.com>,
        Ravi V Shankar <ravi.v.shankar@intel.com>,
        linux-kernel <linux-kernel@vger.kernel.org>, x86 <x86@kernel.org>
In-Reply-To: <20191121185303.GB199273@romley-ivt3.sc.intel.com>
To:     Fenghua Yu <fenghua.yu@intel.com>
Sender: linux-kernel-owner@vger.kernel.org
Precedence: bulk


> On Nov 21, 2019, at 10:40 AM, Fenghua Yu <fenghua.yu@intel.com> wrote:
>=20
> =EF=BB=BFOn Thu, Nov 21, 2019 at 09:51:03AM -0800, Andy Lutomirski wrote:
>>> On Thu, Nov 21, 2019 at 9:43 AM David Laight <David.Laight@aculab.com> w=
rote:
>>>=20
>>> From: Ingo Molnar
>>>> Sent: 21 November 2019 17:12
>>>> * Peter Zijlstra <peterz@infradead.org> wrote:
>>> ...
>>>>> This feature MUST be default enabled, otherwise everything will
>>>>> be/remain broken and we'll end up in the situation where you can't use=

>>>>> it even if you wanted to.
>>>>=20
>>>> Agreed.
>>>=20
>>> Before it can be enabled by default someone needs to go through the
>>> kernel and fix all the code that abuses the 'bit' functions by using the=
m
>>> on int[] instead of long[].
>>>=20
>>> I've only seen one fix go through for one use case of one piece of code
>>> that repeatedly uses potentially misaligned int[] arrays for bitmasks.
>>>=20
>>=20
>> Can we really not just change the lock asm to use 32-bit accesses for
>> set_bit(), etc?  Sure, it will fail if the bit index is greater than
>> 2^32, but that seems nuts.
>>=20
>> (Why the *hell* do the bitops use long anyway?  They're *bit masks*
>> for crying out loud.  As in, users generally want to operate on fixed
>> numbers of bits.)
>=20
> We are working on a separate patch set to fix all split lock issues
> in atomic bitops. Per Peter Anvin and Tony Luck suggestions:
> 1. Still keep the byte optimization if nr is constant. No split lock.
> 2. If type of *addr is unsigned long, do quadword atomic instruction
>   on addr. No split lock.
> 3. If type of *addr is unsigned int, do word atomic instruction
>   on addr. No split lock.
> 4. Otherwise, re-calculate addr to point the 32-bit address which contains=

>   the bit and operate on the bit. No split lock.
>=20
> Only small percentage of atomic bitops calls are in case 4 (e.g. 3%
> for set_bit()) which need a few extra instructions to re-calculate
> address but can avoid big split lock overhead.
>=20
> To get real type of *addr instead of type cast type "unsigned long",
> the atomic bitops APIs are changed to macros from functions. This change
> need to touch all architectures.
>=20

Isn=E2=80=99t the kernel full of casts to long* to match the signature?  Doi=
ng this based on type seems silly to me. I think it=E2=80=99s better to just=
 to a 32-bit operation unconditionally and to try to optimize it using b*l w=
hen safe.=