On Fri, Jan 25, 2019 at 05:53:42PM +0800, Ming Lei wrote:

Hi,

The current support for allocating interrupt sets requires that same 'max_vec'
and 'min_vec' is passed to pci_alloc_irq_vectors_affinity(), then driver has to
try to allocate again and again until it succeeds.

This patch introduces .setup_affinity callback, and we can use it to
re-caculate interrupt sets and build affinity for each set after
irq vectors are allocated.

Turns out both genirq/affinity and nvme code get simplified a lot.

Please review and comment!

Ming Lei (5):
  genirq/affinity: move allocation of 'node_to_cpumask' to
    irq_build_affinity_masks
  genirq/affinity: allow driver to setup managed IRQ's affinity
  genirq/affinity: introduce irq_build_affinity()
  nvme-pci: simplify nvme_setup_irqs() via .setup_affinity callback
  genirq/affinity: remove support for allocating interrupt sets

 drivers/nvme/host/pci.c   |  97 +++++++++++++++++++++--------------------
 drivers/pci/msi.c         |  14 ------
 include/linux/interrupt.h |  42 ++++++++++++------
 kernel/irq/affinity.c     | 108 ++++++++++++++++++++++++----------------------
 4 files changed, 133 insertions(+), 128 deletions(-)

oops, forget to CC Keith.

Thanks,
Ming