On Thu, 09 Apr 2020 12:58:11 -0000 Philipp Eppelt <1871...@bugs.launchpad.net> wrote:
> Public bug reported: > > Setup: > CPU: AMD EPYC-v2 or host's EPYC cpu > Linux 64-bit fedora host; Kernel version 5.5.15-200.fc31 > qemu version: self build > git-head: f3bac27cc1e303e1860cc55b9b6889ba39dee587 > config: Configured with: '../configure' > '--target-list=x86_64-softmmu,mips64el-softmmu,mips64-softmmu,mipsel-softmmu,mips-softmmu,i386-softmmu,aarch64-softmmu,arm-softmmu' > '--prefix=/opt/qemu-master' > > Cmdline: > qemu-system-x86_64 -kernel > /home/peppelt/code/l4/internal/.build-x86_64/bin/amd64_gen/bootstrap -append > "" -initrd "./fiasco/.build-x86_64/fiasco , ... " -serial stdio -nographic > -monitor none -nographic -monitor none -cpu EPYC-v2 -m 4G -smp 4 > > Issue: > We are developing an microkernel operating system called L4Re. We recently > got an AMD EPYC server for testing and we couldn't execute SMP tests of our > system when running Linux + qemu + VM w/ L4Re. > In fact, the kernel did not recognize any APs at all. On AMD CPUs the kernel > checks for the number of cores reported in CPUID leaf 0x8000_0008.ECX[NC] or > [ApicIdSize]. [0][1] > > The physical machine reports for leaf 0x8000_0008: EAX: 0x3030 EBX: > 0x18cf757 ECX: 0x703f EDX: 0x1000 > The lower four bits of ECX are the [NC] field and all set. > > When querying inside qemu with -enable-kvm -cpu host -smp 4 (basically as > replacement and addition to the above cmdline) the CPUID leaf shows: EAX: > 0x3024, EBX: 0x1001000, ECX: 0x0, EDX: 0x0 > Note, ECX is zero. Indicating that this is no SMP capabale CPU. > > I'm debugging it using my local machine and the QEMU provided EPYC-v2 > CPU model and it is reproducible there as well and reports: EAX: > 0x3028, EBX: 0x0, ECX: 0x0, EDX: 0x0 > > I checked other AMD based CPU models (phenom, opteron_g3/g5) and they behave > the same. [2] shows the CPUID 0x8000'0008 handling in the QEMU source. > I believe that behavior here is wrong as ECX[NC] should report the number of > cores per processor, as stated in the AMD manual [2] p.584. In my > understanding -smp 4 should then lead to ECX[NC] = 0x3. > > The following table shows my findings with the -smp option: > Option | Qemu guest observed ECX value > -smp 4 | 0x0 > -smp 4,cores=4 | 0x3 > -smp 4,cores=2,thread=2 | 0x3 > -smp 4,cores=4,threads=2 | QEMU boot error: topology false. > > Now, I'm asking myself how the terminology of the AMD manual maps to QEMU's > -smp option. > Obviously, nr_cores and nr_threads correspond to the cores and threads > options on the cmdline and cores * threads <= 4 (in this example), but what > corresponds the X in -smp X to? I'd say X corresponds to number of logical CPUs. Depending on presence of other options these are distributed among them in magical manner (see pc_smp_parse() for starters) > Querying 0x8000'0008 on the physical processor results in different > reports than quering QEMU's model as does it with -enable-kvm -cpu host. > > Furthermore, the ACPI.MADT shows 4 local APICs to be present while the > CPU leave reports a single core processor. it matches -smp X as it should be. > > This leads me to the conclusion that CPUID 0x8000'0008.ECX reports the > wrong number. CCed author of recent epyc patches who might know how AMD should work better than me. > > Please let me know, if you need more information from my side. > > > [0] > https://github.com/kernkonzept/fiasco/blob/522ccc5f29ab120213cf02d71328e2b879cbbd19/src/kern/ia32/kernel_thread-ia32.cpp#L109 > [1] > https://github.com/kernkonzept/fiasco/blob/522ccc5f29ab120213cf02d71328e2b879cbbd19/src/kern/ia32/cpu-ia32.cpp#L1120 > [2] > https://github.com/qemu/qemu/blob/f2a8261110c32c4dccd84e774d8dd7a0524e00fb/target/i386/cpu.c#L5835 > [3] https://www.amd.com/system/files/TechDocs/24594.pdf > > ** Affects: qemu > Importance: Undecided > Status: New >
