Re: [Qemu-devel] [PATCH] tcg/i386: Do not display HT warning for TCG

[Pranith Kumar <bobby.prani@gmail.com>]

On Wed, Apr 19, 2017 at 10:26 PM, Eduardo Habkost <ehabkost@redhat.com> wrote:
> On Wed, Apr 19, 2017 at 06:03:01PM -0400, Pranith Kumar wrote:
>> On Wed, Apr 19, 2017 at 5:33 PM, Eduardo Habkost <ehabkost@redhat.com> wrote:
>> > On Wed, Apr 19, 2017 at 05:25:23PM -0400, Pranith Kumar wrote:
>> >> On Wed, Apr 19, 2017 at 4:57 PM, Eduardo Habkost <ehabkost@redhat.com> wrote:
>> >> > On Wed, Apr 19, 2017 at 04:16:53PM -0400, Pranith Kumar wrote:
>> >> >> On Wed, Apr 19, 2017 at 4:13 PM, Eduardo Habkost <ehabkost@redhat.com> wrote:
>> >> >> > On Wed, Apr 19, 2017 at 04:00:49PM -0400, Pranith Kumar wrote:
>> >> >> >> On Wed, Apr 19, 2017 at 3:54 PM, Pranith Kumar <bobby.prani@gmail.com> wrote:
>> >> >> >> > When we enable hyperthreading (using threads smp argument), we warn
>> >> >> >> > the user if the cpu is an AMD cpu. This does not make sense on TCG and
>> >> >> >> > is also obsolete now that AMD Ryzen support hyperthreading.
>> >> >> >> >
>> >> >> >> > Fix this by adding CPUID_HT bit to the TCG features and explicitly
>> >> >> >> > checking this bit in the cpuid.
>> >> >> >> >
>> >> >> >> > Signed-off-by: Pranith Kumar <bobby.prani@gmail.com>
>> >> >> >> > ---
>> >> >> >> >  target/i386/cpu.c | 10 +++++-----
>> >> >> >> >  1 file changed, 5 insertions(+), 5 deletions(-)
>> >> >> >> >
>> >> >> >> > diff --git a/target/i386/cpu.c b/target/i386/cpu.c
>> >> >> >> > index 13c0985f11..f34bb5ead7 100644
>> >> >> >> > --- a/target/i386/cpu.c
>> >> >> >> > +++ b/target/i386/cpu.c
>> >> >> >> > @@ -202,12 +202,12 @@ static void x86_cpu_vendor_words2str(char *dst, uint32_t vendor1,
>> >> >> >> >  #define TCG_FEATURES (CPUID_FP87 | CPUID_PSE | CPUID_TSC | CPUID_MSR | \
>> >> >> >> >            CPUID_PAE | CPUID_MCE | CPUID_CX8 | CPUID_APIC | CPUID_SEP | \
>> >> >> >> >            CPUID_MTRR | CPUID_PGE | CPUID_MCA | CPUID_CMOV | CPUID_PAT | \
>> >> >> >> > -          CPUID_PSE36 | CPUID_CLFLUSH | CPUID_ACPI | CPUID_MMX | \
>> >> >> >> > +          CPUID_PSE36 | CPUID_CLFLUSH | CPUID_ACPI | CPUID_MMX | CPUID_HT | \
>> >> >> >> >            CPUID_FXSR | CPUID_SSE | CPUID_SSE2 | CPUID_SS | CPUID_DE)
>> >> >> >> >            /* partly implemented:
>> >> >> >> >            CPUID_MTRR, CPUID_MCA, CPUID_CLFLUSH (needed for Win64) */
>> >> >> >> >            /* missing:
>> >> >> >> > -          CPUID_VME, CPUID_DTS, CPUID_SS, CPUID_HT, CPUID_TM, CPUID_PBE */
>> >> >> >> > +          CPUID_VME, CPUID_DTS, CPUID_SS, CPUID_TM, CPUID_PBE */
>> >> >> >> >  #define TCG_EXT_FEATURES (CPUID_EXT_SSE3 | CPUID_EXT_PCLMULQDQ | \
>> >> >> >> >            CPUID_EXT_MONITOR | CPUID_EXT_SSSE3 | CPUID_EXT_CX16 | \
>> >> >> >> >            CPUID_EXT_SSE41 | CPUID_EXT_SSE42 | CPUID_EXT_POPCNT | \
>> >> >> >> > @@ -3643,7 +3643,7 @@ static void x86_cpu_realizefn(DeviceState *dev, Error **errp)
>> >> >> >> >
>> >> >> >> >      qemu_init_vcpu(cs);
>> >> >> >> >
>> >> >> >> > -    /* Only Intel CPUs support hyperthreading. Even though QEMU fixes this
>> >> >> >> > +    /* Only some CPUs support hyperthreading. Even though QEMU fixes this
>> >> >> >> >       * issue by adjusting CPUID_0000_0001_EBX and CPUID_8000_0008_ECX
>> >> >> >> >       * based on inputs (sockets,cores,threads), it is still better to gives
>> >> >> >> >       * users a warning.
>> >> >> >> > @@ -3651,8 +3651,8 @@ static void x86_cpu_realizefn(DeviceState *dev, Error **errp)
>> >> >> >> >       * NOTE: the following code has to follow qemu_init_vcpu(). Otherwise
>> >> >> >> >       * cs->nr_threads hasn't be populated yet and the checking is incorrect.
>> >> >> >> >       */
>> >> >> >> > -    if (!IS_INTEL_CPU(env) && cs->nr_threads > 1 && !ht_warned) {
>> >> >> >> > -        error_report("AMD CPU doesn't support hyperthreading. Please configure"
>> >> >> >> > +    if ((env->features[FEAT_1_EDX] & CPUID_HT) && (cs->nr_threads > 1) && !ht_warned) {
>> >> >> >>
>> >> >> >> I missed a '!' here. We need to warn if CPUID_HT is not set. But I see
>> >> >> >> that it is not being set even on HT enabled processors (i7-3770). How
>> >> >> >> do I test for it?
>> >> >> >
>> >> >> > CPUID_HT is automatically set on the CPU if
>> >> >> > (nr_threads * nr_cores) > 1. See the "case 1:" block in
>> >> >> > cpu_x86_cpuid().
>> >> >> >
>> >> >> > AFAIK, the point of the warning is to let the user know that the
>> >> >> > guest OS is likely to ignore thread topology information if CPUID
>> >> >> > vendor is not Intel, and has nothing to do with TCG or KVM
>> >> >> > capabilities. Maybe the warning is obsolete today and guests
>> >> >> > won't be confused by a HT AMD CPU, but we need to confirm that.
>> >> >>
>> >> >> We assume an AMD cpu for x86 TCG and it prints this warning when an
>> >> >> smp guest is run with (cores=2,threads=2) argument. The main point of
>> >> >> this patch is to remove that warning when using TCG.
>> >> >
>> >> > What is different under TCG? If guests were confused with
>> >> > threads=2 + vendor=AMD using KVM, what exactly would make them
>> >> > not confused by threads=2 + vendor=AMD when using TCG?
>> >>
>> >> I am not sure why we chose AMD to be the vendor for TCG since, as you
>> >> note, it does not really matter if it's AMD or Intel in TCG mode.
>> >
>> > I don't know why the default on TCG is AMD, either.
>> >
>> >>
>> >> >
>> >> > I just ran a Fedora 25 guest using "-machine accel=tcg
>> >> > -smp 2,threads=2", and it still thinks it's running on a 2-core
>> >> > CPU instead of a 2-thread CPU when CPU vendor is AMD.
>> >>
>> >> This all started with using Windows XP. It has an arbitrary limit on
>> >> the number of cores it can support to 2 cores. So using -smp 4 will
>> >> still only show 2 cores. However, we can overcome this by using -smp
>> >> cores=2,threads=2 and 4 cores show up in the guest. But we see the
>> >> ugly warning that HT is not supported on AMD cpus.
>> >
>> > Is Windows XP able to detect 2 threads * 2 cores even if CPUID
>> > vendor is AMD?
>> >
>>
>> Yep. I don't think it cares for vendor id.
>
> What happens if you use -smp 4,cores=4 ?

This is interesting. The above shows 4 cores in the guest. I then
tried various combinations:

-smp 4 shows 2 cores
-smp 4,cores=4 shows 4 cores
-smp 4,threads=4 shows 4 cores
-smp 4,sockets=1 shows 4 cores

I am a bit confused why smp 4 is showing only 2 cores? and why the
rest are showing 4 cores

>
>>
>> >> So we can either:
>> >>
>> >> * Change the vendor to Intel for TCG, or
>> >> * Warn based on the HT bit being set
>> >
>> > Could you clarify what you mean by the second item? The HT bit is
>> > always set when nr_cores*nr_threads > 1.
>>
>> Hmm, I see it now. We just warn the user if the host processor does
>> not support HT.
>
> We don't look at the host processor at all: the warn depends only
> on the CPUID data we're showing to the guest.
>
>>
>> So how about setting the HT bit only if the host processor supports
>> it? Or since it apparently does not matter anyway, remove the warning
>> entirely?
>
> The host processor doesn't matter. What matters is the CPUID
> data seen by the guest, and it doesn't matter if we're using TCG
> or KVM.
>
> Linux guests seem to ignore the hyperthreading information if
> vendor ID is AMD, and that seems to be the original reason for
> the warning. I would like to remove the warning, too, but it
> looks like this fact hasn't changed.

-- 
Pranith