[PATCH v2] PCI: rockchip: Avoid accessing PCIe registers with clocks gated

[Bjorn Helgaas]

On Tue, Jun 29, 2021 at 11:52:44AM +0100, Robin Murphy wrote:
> On 2021-06-29 07:17, Javier Martinez Canillas wrote:
> > On 6/29/21 2:38 AM, Bjorn Helgaas wrote:
> > > On Thu, Jun 24, 2021 at 05:40:40PM -0500, Bjorn Helgaas wrote:
> > 
> > [snip]
> > 
> > > > > 
> > > > > So let's just move all the IRQ init before the pci_host_probe() call, that
> > > > > will prevent issues like this and seems to be the correct thing to do too.
> > > > 
> > > > Previously we registered rockchip_pcie_subsys_irq_handler() and
> > > > rockchip_pcie_client_irq_handler() before the PCIe clocks were
> > > > enabled.  That's a problem because they depend on those clocks being
> > > > enabled, and your patch fixes that.
> > > > 
> > > > rockchip_pcie_legacy_int_handler() depends on rockchip->irq_domain,
> > > > which isn't initialized until rockchip_pcie_init_irq_domain().
> > > > Previously we registered rockchip_pcie_legacy_int_handler() as the
> > > > handler for the "legacy" IRQ before rockchip_pcie_init_irq_domain().
> > > > 
> > > > I think your patch *also* fixes that problem, right?
> > > 
> > > The lack of consistency in how we use
> > > irq_set_chained_handler_and_data() really bugs me.
> > > 
> > > Your patch fixes the ordering issue where we installed
> > > rockchip_pcie_legacy_int_handler() before initializing data
> > > (rockchip->irq_domain) that it depends on.
> > > 
> > > But AFAICT, rockchip still has the problem that we don't *unregister*
> > > rockchip_pcie_legacy_int_handler() when the rockchip-pcie module is
> > > removed.  Doesn't this mean that if we unload the module, then receive
> > > an interrupt from the device, we'll try to call a function that is no
> > > longer present?
> > 
> > Good question, I don't to be honest. I'll have to dig deeper on this but
> > my experience is that the module removal (and device unbind) is not that
> > well tested on ARM device drivers in general.
> 
> Well, it does use devm_request_irq() so the handler should be unregistered
> by devres *after* ->remove has finished, however that does still leave a
> potential race window in which a pending IRQ could be taken during the later
> part of rockchip_pcie_remove() after it has started turning off critical
> things. Unless the clocks and regulators can also be delegated to devres, it
> might be more robust to explicitly manage the IRQs as well. Mixing the two
> schemes can be problematic when the exact order of both setup and teardown
> matters.

I don't understand the devm_request_irq() connection.  I'm looking at
this irq_set_chained_handler_and_data() call [1]:

  static int rockchip_pcie_setup_irq(struct rockchip_pcie *rockchip)
  {
    ...
    irq = platform_get_irq_byname(pdev, "legacy");
    irq_set_chained_handler_and_data(irq,
				     rockchip_pcie_legacy_int_handler,
				     rockchip);

    irq = platform_get_irq_byname(pdev, "client");
    ...

We look up "irq", pass it to irq_set_chained_handler_and_data(), and
throw it away without saving it anywhere.  How would anything know how
to unregister rockchip_pcie_legacy_int_handler()?

I could imagine irq_set_chained_handler_and_data() saving what's
needed for unregistration, but I would think that would require a
device pointer, which we don't give it.

I'm IRQ-illiterate, so please educate me!

Bjorn

[1] https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/tree/drivers/pci/controller/pcie-rockchip-host.c?id=v5.13#n562