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: <829a3198-1b18-c566-20ee-3e4f23ddbffe@acm.org>
Date:   Fri, 16 Jun 2023 07:41:25 -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] sched/cputime: Make IRQ time accounting configurable at
 boot time
Content-Language: en-US
To:     Peter Zijlstra <peterz@infradead.org>
Cc:     linux-kernel@vger.kernel.org, Ingo Molnar <mingo@redhat.com>,
        Juri Lelli <juri.lelli@redhat.com>,
        Vincent Guittot <vincent.guittot@linaro.org>,
        Dietmar Eggemann <dietmar.eggemann@arm.com>,
        Steven Rostedt <rostedt@goodmis.org>
References: <20230615203745.1259-1-bvanassche@acm.org>
 <20230616074518.GW4253@hirez.programming.kicks-ass.net>
From:   Bart Van Assche <bvanassche@acm.org>
In-Reply-To: <20230616074518.GW4253@hirez.programming.kicks-ass.net>
Content-Type: text/plain; charset=UTF-8; format=flowed
Content-Transfer-Encoding: 7bit
Precedence: bulk

On 6/16/23 00:45, Peter Zijlstra wrote:
> On Thu, Jun 15, 2023 at 01:37:26PM -0700, Bart Van Assche wrote:
>> Some producers of Android devices want IRQ time accounting enabled while
>> others want IRQ time accounting disabled. Hence, make IRQ time accounting
>> configurable at boot time.
> 
> Why would they want this disabled? IRQ time accounting avoids a number
> of issues under high irq/softirq pressure.
> 
> Disabling this makes no sense.

This is why disabling IRQ time accounting makes a ton of sense to me:
* If disabling IRQ time accounting would not be useful, there wouldn't
   be a kernel configuration option that controls whether it is enabled
   or disabled - it would be enabled all the time.
* If enabling IRQ time accounting would be essential, all Linux
   distributors would enable it. In the x86 kernels I checked, IRQ time
   accounting is disabled (Debian and openSUSE).
* For x86 there is already a kernel parameter for disabling IRQ time
   accounting (tsc=noirqtime).
* Modern hardware, e.g. UFSHCI 4.0 controllers, supports sending the
   completion interrupt to the CPU core that submitted the I/O. With such
   hardware IRQ overload (100% spent in IRQ handlers and 0% outside IRQ
   handlers) is impossible because the submitter is slowed down by the
   completion interrupts.
* The performance overhead of CONFIG_IRQ_TIME_ACCOUNTING is
   unacceptable. A quick test in an x86 VM shows that enabling
   CONFIG_IRQ_TIME_ACCOUNTING reduces IOPS by 10% (220K -> 200K). On an
   Android setup I measured an IOPS reduction of 40% (100K -> 60K) due
   to CONFIG_IRQ_TIME_ACCOUNTING. This is not acceptable.

Thanks,

Bart.