Received-SPF: pass (google.com: domain of linux-kernel+bounces-132175-linux.lists.archive=gmail.com@vger.kernel.org designates 147.75.48.161 as permitted sender) client-ip=147.75.48.161;
Date: Thu, 4 Apr 2024 14:55:20 -0700
In-Reply-To: <8ae52273-7b79-46b7-9cd5-2a0c401311c3@intel.com>
Precedence: bulk
Mime-Version: 1.0
References: <20240401185451.6897208962d3fc47a12ffb5a@otheo.eu>
 <20240401185712.a80deeb2fddeded0ad42cc04@otheo.eu> <8ee463af-fdbf-4514-bb6e-bf2fd61fbc06@intel.com>
 <20240402192314.a9b4e05637444314f47557e4@otheo.eu> <d7e89d23-b692-4e70-baae-5df5b3984620@intel.com>
 <20240403175407.f81ebc5cc3300ffa0c39f597@otheo.eu> <7db8d0a8-c668-44be-a348-58120a97fc2b@intel.com>
 <Zg7GpY0_U_7yhe6P@google.com> <8ae52273-7b79-46b7-9cd5-2a0c401311c3@intel.com>
Message-ID: <Zg8hyG__NJZrB-ju@google.com>
Subject: Re: [PATCH 1/1] x86/mm/pae: Align up pteval_t, pmdval_t and pudval_t
 to avoid split locks
From: Sean Christopherson <seanjc@google.com>
To: Dave Hansen <dave.hansen@intel.com>
Cc: Javier Pello <devel@otheo.eu>, linux-kernel@vger.kernel.org, x86@kernel.org, 
	Thomas Gleixner <tglx@linutronix.de>, Ingo Molnar <mingo@redhat.com>, Borislav Petkov <bp@alien8.de>, 
	Dave Hansen <dave.hansen@linux.intel.com>, "H. Peter Anvin" <hpa@zytor.com>
Content-Type: text/plain; charset="us-ascii"

On Thu, Apr 04, 2024, Dave Hansen wrote:
> On 4/4/24 08:26, Sean Christopherson wrote:
> > In other words, practically speaking this isn't about supporting a new hardware
> > feature on 32-bit kernels, it's about preserving performance in real world
> > scenarios when running 32-bit kernels on new hardware.
> 
> Realistically, most of the 32-bit kernels in the world are going to be
> *OLD* distros, right?  Old CentOS/RHEL/SLES kernels from before the
> kernel had split lock detection, or split lock fixes.  Those trip over
> VMM split lock detection now, and presumably will forever.
> 
> I suspect new CentOS/RHEL/SLES kernels that have split lock detection
> all happened after 32-bit support was dropped from those distros.
> 
> I think that basically leaves Debian.  Someone would need to:
> 
>  1. Make a *new* 32-bit Debian distro install (or one of the other
>     less common distros that still do 32-bit)
>  2. Run it on hardware with split lock detection
>  3. On a VMM that enables split lock detection
>  4. Stay close enough to mainline to get split lock fixes (like from
>     this thread)
>  5. Care about performance, despite *ACTIVELY* choosing a 32-bit distro
>     on 64-bit hardware in 2024
>
> Those steps are certainly possible.  I'm just not sure how much trouble
> we want to go to in 2024 to support people that choose new 32-bit
> distros and desire performance.

I'm worried about a scenario where the throttling is so bad that it's not a perf
issue, but a functional issue ("performance" was a bad choice of word).

I do agree that the probability of this being a real problem is super low, but at
the same time it doesn't seem too onerous to clean up.  And in the unlikely case
that this does cause a problem, the pain on our end can be quite high.

> It feels to me to be approaching "I want a pony" territory.
> 
> Or am I just lacking empathy today? :)

Nah, I'm probably asking for a pony, but AFAICT it's a pretty cheap pony.