On Sun, 2024-08-04 at 16:14 -0400, Olivier Langlois wrote: > I am trying to understand the result that the nvme driver has when it > calls pci_alloc_irq_vectors_affinity() from nvme_setup_irqs() > (drivers/nvme/host/pci.c) > > $ cat /proc/interrupts | grep nvme > 63: 9 0 0 0 PCI-MSIX- > 0000:00:04.0 > 0-edge nvme0q0 > 64: 0 0 0 237894 PCI-MSIX- > 0000:00:04.0 > 1-edge nvme0q1 > > $ cat /proc/irq/64/smp_affinity_list > 0-3 > > $ cat /proc/irq/64/effective_affinity_list > 3 > > I think that this happens somewhere below pci_msi_setup_msi_irqs() > (drivers/pci/msi/irqdomain.c) > but I am losing track of what is done precisely because I am not sure > of what is the irq_domain on my system. > > I have experimented by playing with the nvme io queues num that is > passed to pci_msi_setup_msi_irqs() > as the max_vectors params. > > The set irq effective affinity appears to always be the last cpu of > the > affinity mask. > > I would like to have some control on the selected effective_affinity > as > I am trying to use NOHZ_FULL effectively on my system. > > NOTE: > I am NOT using irqbalance > > thank you, > I have found arch/x86/kernel/apic/vector.c and kernel/irq/matrix.c I think that matrix_find_best_cpu() should consider if a CPU is NOHZ_FULL and NOT report it as the best cpu if there are other options... I'll try to play with the idea and report back if I get some success with it... Greetings,
