On Mon, Feb 24, 2025 at 07:59:28PM +0000, Michael Kelley wrote: > From: Hamza Mahfooz <[email protected]> Sent: Monday, February > 24, 2025 6:49 AM > > > > On Fri, Feb 21, 2025 at 11:01:09PM +0000, Michael Kelley wrote: > > > From: Hamza Mahfooz <[email protected]> Sent: Friday, > > > February > > 21, 2025 1:31 PM > > > > > > > > Since, the panic handlers may require certain cpus to be online to panic > > > > gracefully, we should call them before turning off SMP. Without this > > > > re-ordering, on Hyper-V hv_panic_vmbus_unload() times out, because the > > > > vmbus channel is bound to VMBUS_CONNECT_CPU and unless the crashing cpu > > > > is the same as VMBUS_CONNECT_CPU, VMBUS_CONNECT_CPU will be offlined by > > > > crash_smp_send_stop() before the vmbus channel can be deconstructed. > > > > > > Hamza -- what specifically is the problem with the way > > > vmbus_wait_for_unload() > > > works today? That code is aware of the problem that the unload response > > > comes > > > only on the VMBUS_CONNECT_CPU, and that cpu may not be able to handle > > > the interrupt. So the code polls the message page of each CPU to try to > > > get the > > > unload response message. Is there a scenario where that approach isn't > > > working? > > > > > > > It doesn't work on arm64 (if the crashing cpu isn't VMBUS_CONNECT_CPU), it > > always ends up at "VMBus UNLOAD did not complete" without fail. It seems > > like arm64's crash_smp_send_stop() is more aggressive than x86's. > > FWIW, I tested on a D16plds_v6 arm64 VM in Azure, running Ubuntu 20.04 with > a linux-next20252021 kernel. I caused a panic using "echo c > >/proc/sysrq-trigger" > using "taskset" to make sure the panic is triggered on a CPU other than CPU 0. > I didn't see any problem. The panic code path completely quickly, and there > were > no messages from vmbus_wait_for_unload(), including none of the periodic > "Waiting for unload" messages . I tried initiating the panic on several > different > CPUs (4, 7, and 15) with the same result. I tested with kdump disabled and > with > kdump enabled, both with no problems. > > So I think the current vmbus_wait_for_unload() code works on arm64, as least > in some ordinary scenarios. Any key differences in the configuration or test > environment when you see the "did not complete" message?
Can you try on a Standard_D16pls_v5 with the stock ubuntu image and kernel crash dump (i.e. linux-crashdump) installed and setup? > > Michael
