Hi Vijaya, On Tue, Jul 05, 2016 at 08:41:31PM +0530, vijay.kilari@xxxxxxxxx wrote: > From: Vijaya Kumar K <vijayak@xxxxxxxxxxxxxxxxxx> > > The dist and redist regions are created and registered in > vgic_register_dist_iodevs() and vgic_v3_init_redist_iodev() > calls for distributor and redistributor respectively when > vgic_v3_map_resources() is called. This mapping of resources > is done when vcpu run ioctl is called. > > Below is the call stack of the same. > > [<ffff0000080b2e8c>] vgic_register_redist_iodevs+0x94/0x27c > [<ffff0000080b0fd8>] vgic_v3_map_resources+0x138/0x188 > [<ffff0000080affa4>] kvm_vgic_map_resources+0xb0/0xb8 > [<ffff0000080a3888>] kvm_arch_vcpu_ioctl_run+0x4a8/0x550 > [<ffff00000809c218>] kvm_vcpu_ioctl+0x304/0x74c > [<ffff000008232cac>] do_vfs_ioctl+0xc0/0x754 > [<ffff0000082333d0>] SyS_ioctl+0x90/0xa4 > [<ffff000008084af0>] el0_svc_naked+0x24/0x28 > > During live migration, the destination VM first restores the all the > GIC registers(dist, rdist and cpuif registers) using ioctl's before > resuming the VM.So no vcpu run ioctl is called untill complete > GIC context is restored. > > Hence, In case of live migration, when ioctls are called to write > dist/rdist registers the ioctls fails as > vcpu->kvm->arch.vgic.dist_iodev andkvm->arch.vgic.redist_iodevs > are NULL. > Is this for GICv3 only? Can you give me a more specific pointer to the place where this fails or a traceback? Definitely the approach of registering things early doesn't work, because it breaks all sorts of other things. So the trick is to make sure userspace accesses can work without a registers kvm iodev. I will have a look at this, but some more conceret info as requested above will be helpful. Thanks, -Christoffer _______________________________________________ kvmarm mailing list kvmarm@xxxxxxxxxxxxxxxxxxxxx https://lists.cs.columbia.edu/mailman/listinfo/kvmarm
