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;
From:   Valentin Schneider <vschneid@redhat.com>
To:     Tariq Toukan <ttoukan.linux@gmail.com>,
        Tariq Toukan <tariqt@nvidia.com>,
        "David S. Miller" <davem@davemloft.net>,
        Saeed Mahameed <saeedm@nvidia.com>,
        Jakub Kicinski <kuba@kernel.org>,
        Ingo Molnar <mingo@redhat.com>,
        Peter Zijlstra <peterz@infradead.org>,
        Juri Lelli <juri.lelli@redhat.com>
Cc:     Eric Dumazet <edumazet@google.com>,
        Paolo Abeni <pabeni@redhat.com>, netdev@vger.kernel.org,
        Gal Pressman <gal@nvidia.com>,
        Vincent Guittot <vincent.guittot@linaro.org>,
        linux-kernel@vger.kernel.org
Subject: Re: [PATCH net-next V4 1/3] sched/topology: Add NUMA-based CPUs
 spread API
In-Reply-To: <69829c71-d51c-b25f-2d74-5fdd231ed9e4@gmail.com>
References: <20220728191203.4055-1-tariqt@nvidia.com>
 <20220728191203.4055-2-tariqt@nvidia.com>
 <xhsmhedxvdikz.mognet@vschneid.remote.csb>
 <df8b684d-ede6-7412-423d-51d57365e065@gmail.com>
 <xhsmh35e5d9b4.mognet@vschneid.remote.csb>
 <12fd25f9-96fb-d0e0-14ec-3f08c01a5a4b@gmail.com>
 <xhsmhzggdbmv6.mognet@vschneid.remote.csb>
 <69829c71-d51c-b25f-2d74-5fdd231ed9e4@gmail.com>
Date:   Tue, 09 Aug 2022 18:36:59 +0100
Message-ID: <xhsmhwnbhb9ok.mognet@vschneid.remote.csb>
MIME-Version: 1.0
Content-Type: text/plain
Precedence: bulk

On 09/08/22 17:04, Tariq Toukan wrote:
> On 8/9/2022 3:52 PM, Valentin Schneider wrote:
>> On 09/08/22 13:18, Tariq Toukan wrote:
>>> On 8/9/2022 1:02 PM, Valentin Schneider wrote:
>>>>
>>>> Are there cases where we can't figure this out in advance? From what I grok
>>>> out of the two callsites you patched, all vectors will be used unless some
>>>> error happens, so compressing the CPUs in a single cpumask seemed
>>>> sufficient.
>>>>
>>>
>>> All vectors will be initialized to support the maximum number of traffic
>>> rings. However, the actual number of traffic rings can be controlled and
>>> set to a lower number N_actual < N. In this case, we'll be using only
>>> N_actual instances and we want them to be the first/closest.
>>
>> Ok, that makes sense, thank you.
>>
>> In that case I wonder if we'd want a public-facing iterator for
>> sched_domains_numa_masks[%i][node], rather than copy a portion of
>> it. Something like the below (naming and implementation haven't been
>> thought about too much).
>>
>>    const struct cpumask *sched_numa_level_mask(int node, int level)
>>    {
>>            struct cpumask ***masks = rcu_dereference(sched_domains_numa_masks);
>>
>>            if (node >= nr_node_ids || level >= sched_domains_numa_levels)
>>                    return NULL;
>>
>>            if (!masks)
>>                    return NULL;
>>
>>            return masks[level][node];
>>    }
>>    EXPORT_SYMBOL_GPL(sched_numa_level_mask);
>>
>
> The above can be kept static, and expose only the foo() function below,
> similar to my sched_cpus_set_spread().
>

So what I was thinking with this was to only have to export the
sched_numa_level_mask() thing and the iterator, and then consumers would be
free to use whatever storage form they want (cpumask, array, list...).

Right now I believe sched_domains_numa_masks has the right shape for the
interface (for a given node, you a cpumask per distance level) and I
don't want to impose an interface that uses just an array, but perhaps I'm
being silly and the array isn't so bad and simpler to use - that said we
could always build an array-based helper on top of the array of cpumasks
thing.

Clearly I need to scratch my head a bit longer :-)

> LGTM.
> How do you suggest to proceed?
> You want to formalize it? Or should I take it from here?
>

I need to have a think (feel free to ponder and share your thoughts as
well) - I'm happy to push something if I get a brain-wave, but don't let
that hold you back either.