Re: enabling same PPI interrupt to all CPU in ARM64 SMP
Date: Fri, 12 May 2023 17:09:40 UTC
On Fri, May 12, 2023 at 9:51 AM Souradeep Chakrabarti <schakrabarti@microsoft.com> wrote: > > > > > >-----Original Message----- > >From: Souradeep Chakrabarti > >Sent: Monday, May 8, 2023 6:39 PM > >To: Kyle Evans <kevans@freebsd.org> > >Cc: Wei Hu <weh@microsoft.com>; freebsd-hackers@FreeBSD.org > >Subject: enabling same PPI interrupt to all CPU in ARM64 SMP > > > >Hi , > > > >While using SMP in ARM64 Hyper-V we are getting stuck in boot if there is a > >interrupt for VMBus coming to CPU1 and VMBus interrupt handler is not getting > >that interrupt. > > > >In ARM64 Hyper-V we are using IRQ18 for VMBus and it is a PPI interrupt. > > > >But Hypev-V host sends interrupt to this IRQ 18 for both CPU0 and CPU1 in 2CPU > >system. > >This is based on the corresponding VMBus channel which assigned with the CPU. > > > >Now VMBus ISR is getting the interrupt in CPU0 but not getting from CPU1. > >Any idea, how we can use the same PPI 18 for all the CPU cores? > > > >Any help will be appreciated, as this is blocking the enablement of FreeBSD in Azure > >ARM64. > [Souradeep] > Can someone please help me it. > Looking at least at the GIC implementation, it looks like this is a known limitation: 875 /* 876 * XXX - In case that per CPU interrupt is going to be enabled in time 877 * when SMP is already started, we need some IPI call which 878 * enables it on others CPUs. Further, it's more complicated as 879 * pic_enable_source() and pic_disable_source() should act on 880 * per CPU basis only. Thus, it should be solved here somehow. 881 */ 882 if (isrc->isrc_flags & INTR_ISRCF_PPI) 883 CPU_SET(PCPU_GET(cpuid), &isrc->isrc_cpu); I think we need something /like/ this: https://people.freebsd.org/~kevans/ppi.diff, though it still has the caveat that PPIs effectively cannot be fully setup before SI_SUB_SMP. So, it's likely almost a NOP for existing platforms (will emit a warning with bootverbose for armv8 timers) but might do the trick for you. Thanks, Kyle Evans