Paolo, On 4/25/20 4:52 PM, Paolo Bonzini wrote:
On 24/04/20 07:08, Suravee Suthikulpanit wrote:commit f458d039db7e ("kvm: ioapic: Lazy update IOAPIC EOI") introduces the following regression on Intel VMX APICv. BUG: stack guard page was hit at 000000008f595917 \ (stack is 00000000bdefe5a4..00000000ae2b06f5) kernel stack overflow (double-fault): 0000 [#1] SMP NOPTI RIP: 0010:kvm_set_irq+0x51/0x160 [kvm] Call Trace: irqfd_resampler_ack+0x32/0x90 [kvm] kvm_notify_acked_irq+0x62/0xd0 [kvm] kvm_ioapic_update_eoi_one.isra.0+0x30/0x120 [kvm] ioapic_set_irq+0x20e/0x240 [kvm] kvm_ioapic_set_irq+0x5c/0x80 [kvm] kvm_set_irq+0xbb/0x160 [kvm] ? kvm_hv_set_sint+0x20/0x20 [kvm] irqfd_resampler_ack+0x32/0x90 [kvm] kvm_notify_acked_irq+0x62/0xd0 [kvm] kvm_ioapic_update_eoi_one.isra.0+0x30/0x120 [kvm] ioapic_set_irq+0x20e/0x240 [kvm] kvm_ioapic_set_irq+0x5c/0x80 [kvm] kvm_set_irq+0xbb/0x160 [kvm] ? kvm_hv_set_sint+0x20/0x20 [kvm] .... This is due to the logic always force IOAPIC lazy update EOI mechanism when APICv is activated, which is only needed by AMD SVM AVIC. Fixes by introducing struct kvm_arch.use_lazy_eoi variable to specify whether the architecture needs lazy update EOI support.You are not explaining why the same infinite loop cannot happen on AMD. It seems to me that it is also fixed by adding a check for re-entrancy in ioapic_lazy_update_eoi. It's easy to add one since ioapic_lazy_update_eoi is called with the ioapic->lock taken. Paolo
I finally reproduced on AMD system as well. I'll send out a new patch for this based on your suggestion. Suravee
