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: <1e71ad64-f5bc-2581-87b8-9bd0932a7875@gmail.com>
Date:   Mon, 25 Jul 2022 13:18:56 -0700
MIME-Version: 1.0
User-Agent: Mozilla/5.0 (X11; Linux x86_64; rv:91.0) Gecko/20100101
 Thunderbird/91.11.0
Subject: Re: Why set .suppress_bind_attrs even though .remove() implemented?
Content-Language: en-US
To:     Marc Zyngier <maz@kernel.org>, Bjorn Helgaas <helgaas@kernel.org>
Cc:     =?UTF-8?Q?Pali_Roh=c3=a1r?= <pali@kernel.org>,
        Johan Hovold <johan+linaro@kernel.org>,
        Kishon Vijay Abraham I <kishon@ti.com>,
        Xiaowei Song <songxiaowei@hisilicon.com>,
        Binghui Wang <wangbinghui@hisilicon.com>,
        Thierry Reding <thierry.reding@gmail.com>,
        Ryder Lee <ryder.lee@mediatek.com>,
        Jianjun Wang <jianjun.wang@mediatek.com>,
        linux-pci@vger.kernel.org,
        =?UTF-8?Q?Krzysztof_Wilczy=c5=84ski?= <kw@linux.com>,
        Ley Foon Tan <ley.foon.tan@intel.com>,
        linux-kernel@vger.kernel.org
References: <87k085xekg.wl-maz@kernel.org> <20220722171706.GA1911557@bhelgaas>
 <877d42df5t.wl-maz@kernel.org>
From:   Florian Fainelli <f.fainelli@gmail.com>
In-Reply-To: <877d42df5t.wl-maz@kernel.org>
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 7bit
Precedence: bulk

On 7/24/22 02:38, Marc Zyngier wrote:
> On Fri, 22 Jul 2022 18:17:06 +0100,
> Bjorn Helgaas <helgaas@kernel.org> wrote:
>>
>> On Fri, Jul 22, 2022 at 06:06:07PM +0100, Marc Zyngier wrote:
>>> On Fri, 22 Jul 2022 15:39:05 +0100,
>>> Bjorn Helgaas <helgaas@kernel.org> wrote:
>>>>
>>>> [+cc Marc, can you clarify when we need irq_dispose_mapping()?]
>>>
>>> In general, interrupt controllers should not have to discard mappings
>>> themselves, just like they rarely create mappings themselves. That's
>>> usually a different layer that has created it (DT, for example).
>>>
>>> The problem is that these mappings persist even if the interrupt has
>>> been released by the driver (it called free_irq()), and the IRQ number
>>> can be further reused. The client driver could dispose of the mapping
>>> after having released the IRQ, but nobody does that in practice.
>>>
>>> From the point of view of the controller, there is no simple way to
>>> tell when an interrupt is "unused". And even if a driver was
>>> overzealous and called irq_dispose_mapping() on all the possible
>>> mappings (and made sure no mapping could be created in parallel), this
>>> could result in a bunch of dangling pointers should a client driver
>>> still have the interrupt requested.
>>>
>>> Fixing this is pretty hard, as IRQ descriptors are leaky (you can
>>> either have a pointer to one, or just an IRQ number -- they are
>>> strictly equivalent). So in general, being able to remove an interrupt
>>> controller driver is at best fragile, and I'm trying not to get more
>>> of this in the tree.
>>
>> Thank you!
>>
>> How do we identify an interrupt controller driver?  Apparently some of
>> these PCIe controller drivers also include an interrupt controller
>> driver, but I don't know what to look for to find them.
> 
> If you see a struct irq_chip somewhere, this is an interrupt
> controller.  And yes, most of the PCIe RC drivers will have some sort
> of interrupt controller driver for INTx support, as well as MSI when
> the RC doesn't/cannot rely on the platform providing one.
> 
> It means that these PCIe RC drivers probably shouldn't be removable if
> built as modules. Which I don't think is a big problem. You want
> modularity to reduce the size of the kernel image and only load the
> drivers the platform actually requires, saving memory in the process.
> And for something as fundamental as an interrupt controller (and PCIe
> in general), you probably want to keep it around for the lifetime of
> the machine.

No disagreement, however being able to fully remove and load the module again ensures that you bring the hardware and software in a sane state every time, i.e.: it does help find actual bugs in either implementations. It's also a faster turn around time if you are working on that specific subsystem in avoid rebooting the kernel needlessly (that puts a lot of faith into not crashing the kernel, still).
-- 
Florian