Re: [PATCH v3] i386/cpu: fixup number of addressable IDs for logical processors in the physical package

[Igor Mammedov <imammedo@redhat.com>]

On Fri, 20 Sep 2024 02:29:46 +0800
Zhao Liu <zhao1.liu@intel.com> wrote:

> Hi Chuang and Igor,
> 
> Sorry for late reply,
> 
> On Wed, Sep 18, 2024 at 09:18:15PM +0800, Chuang Xu wrote:
> > Date: Wed, 18 Sep 2024 21:18:15 +0800
> > From: Chuang Xu <xuchuangxclwt@bytedance.com>
> > Subject: [PATCH v3] i386/cpu: fixup number of addressable IDs for logical
> >  processors in the physical package
> > X-Mailer: git-send-email 2.24.3 (Apple Git-128)
> > 
> > When QEMU is started with:
> > -cpu host,migratable=on,host-cache-info=on,l3-cache=off
> > -smp 180,sockets=2,dies=1,cores=45,threads=2
> > 
> > Try to execute "cpuid -1 -l 1 -r" in guest, we'll obtain a value of 90 for
> > CPUID.01H.EBX[23:16], while the expected value is 128. And Try to
> > execute "cpuid -1 -l 4 -r" in guest, we'll obtain a value of 63 for
> > CPUID.04H.EAX[31:26] as expected.
> > 
> > As (1+CPUID.04H.EAX[31:26]) round up to the nearest power-of-2 integer,
> > we'd beter round up CPUID.01H.EBX[23:16] to the nearest power-of-2
> > integer too. Otherwise we may encounter unexpected results in guest.
> > 
> > For example, when QEMU is started with CLI above and xtopology is disabled,
> > guest kernel 5.15.120 uses CPUID.01H.EBX[23:16]/(1+CPUID.04H.EAX[31:26]) to
> > calculate threads-per-core in detect_ht(). Then guest will get "90/(1+63)=1"
> > as the result, even though theads-per-core should actually be 2.
> > 
> > So let us round up CPUID.01H.EBX[23:16] to the nearest power-of-2 integer
> > to solve the unexpected result.
> > 
> > Signed-off-by: Guixiong Wei <weiguixiong@bytedance.com>
> > Signed-off-by: Yipeng Yin <yinyipeng@bytedance.com>
> > Signed-off-by: Chuang Xu <xuchuangxclwt@bytedance.com>
> > ---
> >  target/i386/cpu.c | 2 +-
> >  1 file changed, 1 insertion(+), 1 deletion(-)
> > 
> > diff --git a/target/i386/cpu.c b/target/i386/cpu.c
> > index 4c2e6f3a71..3710ae5283 100644
> > --- a/target/i386/cpu.c
> > +++ b/target/i386/cpu.c
> > @@ -6417,7 +6417,7 @@ void cpu_x86_cpuid(CPUX86State *env, uint32_t index, uint32_t count,
> >          }
> >          *edx = env->features[FEAT_1_EDX];
> >          if (threads_per_pkg > 1) {
> > -            *ebx |= threads_per_pkg << 16;
> > +            *ebx |= pow2ceil(threads_per_pkg) << 16;  
> 
> Yes, the fix is right.
> 
> About the "Maximum number of addressable IDs", the commit 88dd4ca06c83
> ("i386/cpu: Use APIC ID info to encode cache topo in CPUID[4]")
> introduced the new way to calculate.
> 
> The pow2ceil() works for current SMP topology, but may be wrong on
> hybrid topology, as the reason I listed in the commit message:
> 
> > The nearest power-of-2 integer can be calculated by pow2ceil() or by
> > using APIC ID offset/width (like L3 topology using 1 << die_offset [3]).  
> 
> > But in fact, CPUID.04H:EAX[bits 25:14] and CPUID.04H:EAX[bits 31:26]
> > are associated with APIC ID. For example, in linux kernel, the field
> > "num_threads_sharing" (Bits 25 - 14) is parsed with APIC ID. And for
> > another example, on Alder Lake P, the CPUID.04H:EAX[bits 31:26] is not
> > matched with actual core numbers and it's calculated by:
> > "(1 << (pkg_offset - core_offset)) - 1".  
> 
> Using APIC ID offset to calculate is the hardware's approach, so I tried
> to use APIC ID instead of pow2ceil() and replaced all pow2ceil() case.

Well, hybrid case needs some more explanation then.

'pow2ceil(threads_per_pkg) << 16' - does exactly what SDM says for CPUID.01H.EBX[23:16]

Can you point to a spec that confirms that above is wrong and
explain in more details how hybrid case is supposed to work
and where it's documented?
 


> 
> Hi Igor, do you agree? :-)
> 
> Best Regards,
> Zhao
>