Re: [PATCH net-next v3 2/3] net: stmmac: Add glue driver for Motorcomm YT6801 ethernet controller

[Philipp Stanner]

On Mon, 2025-12-08 at 10:53 +0000, Russell King (Oracle) wrote:
> On Mon, Dec 08, 2025 at 11:47:23AM +0100, Philipp Stanner wrote:
> > On Mon, 2025-12-08 at 10:15 +0000, Russell King (Oracle) wrote:
> > > On Mon, Dec 08, 2025 at 10:54:36AM +0100, Philipp Stanner wrote:
> > > > The bad news is that it's not super trivial to remove. I looked into it
> > > > about two times and decided I can't invest that time currently. You
> > > > need to go over all drivers again to see who uses pcim_enable_device(),
> > > > then add free_irq_vecs() for them all and so on…
> > > 
> > > So that I can confirm, you're saying that all drivers that call
> > > pci_alloc_irq_vectors() should call pci_free_irq_vectors() in their
> > > ->remove() method and not rely on the devres behaviour that
> > > pcim_enable_device() will permit.
> > 
> > "permit" is kind of a generous word. This behavior is dangerous and
> > there were bugs because of that in the past, because it confused
> > programmers. See:
> > 
> > f00059b4c1b0 drm/vboxvideo: fix mapping leaks
> > 
> > 
> > pcim_enable_device() used to switch all sorts of functions into managed
> > mode. As far as I could figure out through git, back in 2009 it was
> > intended that ALL pci functions are switched into managed mode that
> > way. That's also how it was documented.
> > 
> > The ecosystem then fractured, however. Some functions were always
> > managed (pcim_), some never, and some sometimes.
> > 
> > I removed all "sometimes managed" functions since 2024. The last
> > remainder is MSI.
> > 
> > If we want to remove that, we need to:
> >    1. Find all drivers that rely on pci_free_irq_vectors() being run
> >       automatically. IOW those that use pcim_enable_device() + wrappers
> >       around pci_setup_msi_context().
> >    2. Port those drivers to do the free_irq_vecs manually, if it's not
> >       a problem if it's called twice. If that were a problem, those
> >       drivers would also need to replace pcim_enable_device() with
> >       pci_enable_device().
> >    3. Once all drivers are ported, remove the devres code from msi.c
> >    4. Do associated cleanup work in PCI.
> > 
> > > 
> > > In terms of whether it's safe to call this twice, pci_free_irq_vectors()
> > > calls pci_disable_msix() and pci_disable_msi().
> > > 
> > > pci_disable_msix() checks:
> > > 
> > >         if (!pci_msi_enabled() || !dev || !dev->msix_enabled)
> > >                 return;
> > > 
> > > which will set dev->msix_enabled to 0 via pci_msix_shutdown().
> > > 
> > > pci_disable_msi() does a similar check:
> > > 
> > >         if (!pci_msi_enabled() || !dev || !dev->msi_enabled)
> > >                 return;
> > > 
> > > and similarly pci_msi_shutdown() sets dev->msi_enabled to 0.
> > > 
> > > So my conclusion is it's safe to call pci_free_irq_vectors() twice for
> > > the same device.
> > > 
> > 
> > Hm. Looks good.
> 
> So, what do you want to see for new drivers such as the one at the top
> of this thread? Should they explicitly call pci_free_irq_vectors() even
> though they call pcim_enable_device() ?


Yes, I think that's the right thing to do. It makes removing that
feature from MSI easier, since there will not be even more drivers to
port.

P.