On Fri, Feb 14, 2025 at 12:45:52PM +0530, Manivannan Sadhasivam wrote: > On Tue, Feb 11, 2025 at 04:23:53PM +0800, Tsai Sung-Fu wrote: > > >Because you cannot set affinity for chained MSIs as these MSIs are muxed to > > >another parent interrupt. Since the IRQ affinity is all about changing which CPU > > >gets the IRQ, affinity setting is only possible for the MSI parent. > > > > So if we can find the MSI parent by making use of chained > > relationships (32 MSI vectors muxed to 1 parent), > > is it possible that we can add that implementation back ? > > We have another patch that would like to add the > > dw_pci_msi_set_affinity feature. > > Would it be a possible try from your perspective ? > > > > This question was brought up plenty of times and the concern from the irqchip > maintainer Marc was that if you change the affinity of the parent when the child > MSI affinity changes, it tends to break the userspace ABI of the parent. > > See below links: > > https://lore.kernel.org/all/87mtg0i8m8.wl-maz@xxxxxxxxxx/ > https://lore.kernel.org/all/874k0bf7f7.wl-maz@xxxxxxxxxx/ It's hard to meaningfully talk about a patch that hasn't been posted yet, but the implementation we have at least attempts to make *some* kind of resolution to those ABI questions. For one, it rejects affinity changes that are incompatible (by some definition) with affinities requested by other virqs shared on the same parent line. It also updates their effective affinities upon changes. Those replies seem to over-focus on dynamic, user-space initiated changes too. But how about for "managed-affinity" interrupts? Those are requested by drivers internally to the kernel (a la pci_alloc_irq_vectors_affinity()), and can't be changed by user space afterward. It seems like there'd be room for supporting that, provided we don't allow conflicting/non-overlapping configurations. I do see that Marc sketched out a complex sysfs/hierarchy API in some of his replies. I'm not sure that would provide too much value to the managed-affinity cases we're interested in, but I get the appeal for user-managed affinity. Brian
