On Tue, Dec 03, 2024 at 10:12:36PM +0100, Thomas Gleixner wrote: > On Tue, Dec 03 2024 at 15:36, Frank Li wrote: > > The follow steps trigger kernel dump warning and > > platform_device_msi_init_and_alloc_irqs() return false. > > > > 1: platform_device_msi_init_and_alloc_irqs(); > > 2: platform_device_msi_free_irqs_all(); > > 3: platform_device_msi_init_and_alloc_irqs(); > > > > Do below two things in platform_device_msi_init_and_alloc_irqs(). > > - msi_create_device_irq_domain() > > - msi_domain_alloc_irqs_range() > > > > But only call msi_domain_free_irqs_all() in > > platform_device_msi_free_irqs_all(), which missed call > > msi_remove_device_irq_domain(). > > It's not a missed call. It's intentional as all existing users remove > the device afterwards. > > > This cause above kernel dump when call > > platform_device_msi_init_and_alloc_irqs() again. > > Sure, but that's not a fix and not required for stable because no > existing driver is affected by this unless I'm missing something. > > What's the actual use case for this? You describe in great length what > fails, which is nice, but I'm missing the larger picture here. PCI host send a door bell to PCI endpoint, which use platform msi to trigger a IRQ. PCI Host side PCI endpoint side Send "enable" command -> call platform_device_msi_init_and_alloc_irqs() Get doorbell address <- send back MSI address by shared memory Write data to doorbell -> MSI irq handler triggered. Send "Disable" command -> call platform_device_msi_free_irqs_all() At endpoint side, need dymatic response "enable/disable" commands. Of course, I can call msi_remove_device_irq_domain() in my disable function. But I think it should be symetic in alloc/free pair functions. Previous version, alloc/free in bind/unbind function. I missed to do unbind/bind test. the principle should be the same. Frank > > Thanks, > > tglx
