Received-SPF: pass (google.com: domain of linux-kernel-owner@vger.kernel.org designates 2620:137:e000::1:20 as permitted sender) client-ip=2620:137:e000::1:20;
Message-ID: <7422da56-a5a1-d6b2-a2a0-fd4ee5dde5e0@intel.com>
Date:   Tue, 11 Jul 2023 16:07:55 -0700
MIME-Version: 1.0
User-Agent: Mozilla/5.0 (X11; Linux x86_64; rv:102.0) Gecko/20100101
 Thunderbird/102.11.0
Subject: Re: [PATCH] x86/platform/uv: Abort UV initialization when reduced
 nr_cpus requires it
Content-Language: en-US
To:     Steve Wahl <steve.wahl@hpe.com>,
        Dimitri Sivanich <dimitri.sivanich@hpe.com>,
        Russ Anderson <russ.anderson@hpe.com>,
        Thomas Gleixner <tglx@linutronix.de>,
        Ingo Molnar <mingo@redhat.com>, Borislav Petkov <bp@alien8.de>,
        Dave Hansen <dave.hansen@linux.intel.com>, x86@kernel.org,
        "H. Peter Anvin" <hpa@zytor.com>, linux-kernel@vger.kernel.org
References: <20230711202618.85562-1-steve.wahl@hpe.com>
From:   Dave Hansen <dave.hansen@intel.com>
In-Reply-To: <20230711202618.85562-1-steve.wahl@hpe.com>
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 7bit
Precedence: bulk

On 7/11/23 13:26, Steve Wahl wrote:
> When nr_cpus is set to a smaller number than actually present, there
> is some node-to-socket mapping info we won't get access to in

First of all, no "we's" in commit messages.

> https://www.kernel.org/doc/html/next/process/maintainer-tip.html#changelog

> build_socket_tables().  This could later result in using a -1 value
> for some array indexing, and eventual kernel page faults.
> 
> To avoid this, if any unfilled table entries are found, print a
> warning message, and resume initializing, acting as if this is not a
> UV system.  UV features will be unavailable, but we will not cause
> kernel dumps.
> 
> This is a condition we expect only in platform debugging situations,
> not in day-to-day operation.

This seems like a hack.

The real problem is that you've got an online Linux NUMA node with no
CPUs.  uv_system_init_hub() (probably) goes and does:

>         for_each_node(nodeid)
>                 __uv_hub_info_list[nodeid] = uv_hub_info_list_blade[uv_node_to_blade_id(nodeid)];

But the node=>blade lookup uses socket numbers.  No CPUs means no socket
numbers.  You _have_ the blade information _somewhere_.  Is there really
no other way to map it to a NUMA node than using the CPU apicid?