Re: [Qemu-devel] [PULL 11/13] target-i386: forward CPUID cache leaves when -cpu host is used

[Peter Lieven]

On 19.11.2013 13:14, Paolo Bonzini wrote:
> Il 19/11/2013 13:03, Peter Lieven ha scritto:
> > > Can you test which of these two work?  But I agree it's best to disable
> > > cache-leaf forwarding.
> > The first does make windows boot again and it calculates a
> > correct combination of cpus, threads, cores and sockets. But
> > I think the reason it boots is because cores=threads=1.
> >
> > As its more intuitive (I think) I would prefer your "cores over threads
> > over socket ".
> > The last thing I would think of is emulating more than 1 socket. -smp N
> > would then mean, N cores, no hyper-threading, 1 socket.
> After looking more at the docs, I think I found the bug.  Can you test this?
>
> diff --git a/target-i386/cpu.c b/target-i386/cpu.c
> index 864c80e..16d4db1 100644
> --- a/target-i386/cpu.c
> +++ b/target-i386/cpu.c
> @@ -2086,14 +2086,10 @@ void cpu_x86_cpuid(CPUX86State *env, uint32_t index, 
> uint32_t count,
>           /* cache info: needed for Core compatibility */
>           if (cpu->cache_info_passthrough) {
>               host_cpuid(index, count, eax, ebx, ecx, edx);
> -            break;
> -        }
> -        if (cs->nr_cores > 1) {
> -            *eax = (cs->nr_cores - 1) << 26;
> +            *eax &= ~0xFC000000;
>           } else {
>               *eax = 0;
> -        }
> -        switch (count) {
> +            switch (count) {
>               case 0: /* L1 dcache info */
>                   *eax |= CPUID_4_TYPE_DCACHE | \
>                           CPUID_4_LEVEL(1) | \
> @@ -2118,9 +2114,6 @@ void cpu_x86_cpuid(CPUX86State *env, uint32_t index, 
> uint32_t count,
>                   *eax |= CPUID_4_TYPE_UNIFIED | \
>                           CPUID_4_LEVEL(2) | \
>                           CPUID_4_SELF_INIT_LEVEL;
> -                if (cs->nr_threads > 1) {
> -                    *eax |= (cs->nr_threads - 1) << 14;
> -                }
>                   *ebx = (L2_LINE_SIZE - 1) | \
>                          ((L2_PARTITIONS - 1) << 12) | \
>                          ((L2_ASSOCIATIVITY - 1) << 22);
> @@ -2133,6 +2126,12 @@ void cpu_x86_cpuid(CPUX86State *env, uint32_t index, 
> uint32_t count,
>                   *ecx = 0;
>                   *edx = 0;
>                   break;
> +            }
> +        }
> +
> +        /* We give out APIC IDs ourselves, so force bits 31..26 even for "-cpu 
> host".  */
> +        if (cs->nr_cores > 1) {
> +            *eax |= (cs->nr_cores - 1) << 26;
>           }
>           break;
>       case 5:
>
> Paolo
I already tried exactly this fix. Its reading index 0x0000004 for increasing 
indexes until qemu aborts:

~/git/qemu$ x86_64-softmmu/qemu-system-x86_64 -m 2048 -drive if=virtio,file=iscsi://172.21.200.45/iqn.2001-05.com.equallogic:0-8a0906-9d95c510a-344001d54795289f-2012-r2-1-7-0/0,format=raw,cache=writeback,aio=native -cpu host -monitor stdio -vnc :1 
-enable-kvm -usb -usbdevice tablet -vga cirrus -global virtio-blk-pci.scsi=off -smp 4,cores=4,threads=1,sockets=1  -serial null  -parallel null -boot c

(qemu) cpuid_data is full, no space for cpuid(eax:0x4,ecx:0x5d)
Abgebrochen (Speicherabzug geschrieben)

If you really want to have this feature:

a) fix smp_parse and leave it at "prefer sockets over cores over threads", but 
use your new code:

        if (cpus == 0) {
            sockets = sockets > 0 ? sockets : 1;
            cores = cores > 0 ? cores : 1;
            threads = threads > 0 ? threads : 1;
            cpus = cores * threads * sockets;
        } else if (sockets == 0) {
            cores = cores > 0 ? cores : 1;
            threads = threads > 0 ? threads : 1;
            sockets = cpus / (cores * threads);
        } else if (cores == 0) {
            threads = threads > 0 ? threads : 1;
            cores = cpus / (sockets * threads);
        } else {
            threads = cpus / (sockets * cores);
        }


b) disable cache leaf pass-thru as soon as threads*cores > 1. It seems to work 
as long as there is only one core with one thread per socket.

Peter