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: <458ac4d2-bb8a-0359-f198-dd53f4c84bd3@redhat.com>
Date:   Fri, 4 Aug 2023 12:51:47 -0400
MIME-Version: 1.0
User-Agent: Mozilla/5.0 (X11; Linux x86_64; rv:102.0) Gecko/20100101
 Thunderbird/102.7.1
Subject: Re: [PATCH v3] perf/arm-dmc620: Fix
 dmc620_pmu_irqs_lock/cpu_hotplug_lock circular lock dependency
Content-Language: en-US
To:     Will Deacon <will@kernel.org>
Cc:     Mark Rutland <mark.rutland@arm.com>,
        linux-arm-kernel@lists.infradead.org, linux-kernel@vger.kernel.org,
        Robin Murphy <robin.murphy@arm.com>
References: <20230722031729.3913953-1-longman@redhat.com>
 <20230728150614.GF21718@willie-the-truck>
 <62d4b353-0237-9ec6-a63e-8a7a6764aba5@redhat.com>
 <20230804162812.GC30679@willie-the-truck>
From:   Waiman Long <longman@redhat.com>
In-Reply-To: <20230804162812.GC30679@willie-the-truck>
Content-Type: text/plain; charset=UTF-8; format=flowed
Content-Transfer-Encoding: 8bit
Precedence: bulk

On 8/4/23 12:28, Will Deacon wrote:
>>>>    struct dmc620_pmu {
>>>> @@ -423,9 +424,14 @@ static struct dmc620_pmu_irq *__dmc620_pmu_get_irq(int irq_num)
>>>>    	struct dmc620_pmu_irq *irq;
>>>>    	int ret;
>>>> -	list_for_each_entry(irq, &dmc620_pmu_irqs, irqs_node)
>>>> -		if (irq->irq_num == irq_num && refcount_inc_not_zero(&irq->refcount))
>>>> +	list_for_each_entry(irq, &dmc620_pmu_irqs, irqs_node) {
>>>> +		if (irq->irq_num != irq_num)
>>>> +			continue;
>>>> +		if (!irq->valid)
>>>> +			return ERR_PTR(-EAGAIN);	/* Try again later */
>>> It looks like this can bubble up to the probe() routine. Does the driver
>>> core handle -EAGAIN coming back from a probe routine?
>> Right, I should add code to handle this error condition. I think it can be
>> handled in dmc620_pmu_get_irq(). The important thing is to release the
>> mutex, wait a few ms and try again. What do you think?
> I don't really follow, but waiting a few ms and trying again sounds like
> a really nasty hack for something which doesn't appear to be constrained
> by broken hardware. In other words, we got ourselves into this mess, so
> we should be able to resolve it properly.

 From my point of view, the proper way to solve the problem is to 
reverse the locking order. Since you don't to add a EXPORT statement to 
the core kernel code, we will have to find a way around it by not 
holding the dmc620_pmu_irqs_lock when cpuhp_state_add_instance_nocalls() 
is called. Another alternative that I can think of is to add one more 
mutex that we will hold just for the entirety of  __dmc620_pmu_get_irq() 
and take dmc620_pmu_irqs_lock only when the linked list is being 
modified. That will eliminate the need to introduce arbitrary wait as 
other caller of __dmc620_pmu_get_irq() will wait in the new mutex. Will 
this work for you?

Cheers,
Longman