On Thu, 2024-10-10 at 17:46 +0300, Andy Shevchenko wrote:

On Wed, Oct 09, 2024 at 10:35:08AM +0200, Philipp Stanner wrote:

quoted

pci_intx() is a hybrid function which can sometimes be managed
through
devres. To remove this hybrid nature from pci_intx(), it is
necessary to
port users to either an always-managed or a never-managed version.

hda_intel enables its PCI-Device with pcim_enable_device(). Thus,
it needs
the always-managed version.

Replace pci_intx() with pcim_intx().

...

quoted

 	bus->irq = chip->pci->irq;
 	chip->card->sync_irq = bus->irq;
-	pci_intx(chip->pci, !chip->msi);
+	pcim_intx(chip->pci, !chip->msi);
 	return 0;

I believe each driver needs an individual approach. Looking at the
above
I would first to understand why this one is being used and why we
can't
switch to pci{m}_alloc_irq_vectors(). (Yeah, managed
pci_alloc_irq_vectors()
is probably still missing, I don't remember if you introduced it or
not.

Alright alright – we touched it in the other mail briefly, but let me
point out another specific problem:

pci_alloc_irq_vectors() *uses* pci_intx(). And pci_intx() can be
managed sometimes.

See the problem? :(

So it's not just that I couldn't port the driver Alex is concerned
about, it's also that MSI itself is a user of pci_intx().

So a pcim_alloc_irq_vectors() might end up doing double-devres or God
knows what else. Only once pci_intx() is clean one can start thinking
about the code in pci/msi/

It's the biggest reason why I want to clean it up as suggested here,
and also why the only patch I'm really nervous about is number 8.


P.