From mboxrd@z Thu Jan 1 00:00:00 1970 Received: from eggs.gnu.org ([2001:4830:134:3::10]:47701) by lists.gnu.org with esmtp (Exim 4.71) (envelope-from ) id 1VikTl-0000cG-6p for qemu-devel@nongnu.org; Tue, 19 Nov 2013 07:32:37 -0500 Received: from Debian-exim by eggs.gnu.org with spam-scanned (Exim 4.71) (envelope-from ) id 1VikTc-0003nX-8f for qemu-devel@nongnu.org; Tue, 19 Nov 2013 07:32:29 -0500 Received: from mx.ipv6.kamp.de ([2a02:248:0:51::16]:45136 helo=mx01.kamp.de) by eggs.gnu.org with esmtp (Exim 4.71) (envelope-from ) id 1VikTb-0003nF-KP for qemu-devel@nongnu.org; Tue, 19 Nov 2013 07:32:20 -0500 Message-ID: <528B5A5B.1090705@kamp.de> Date: Tue, 19 Nov 2013 13:32:27 +0100 From: Peter Lieven MIME-Version: 1.0 References: <1379694292-1601-1-git-send-email-pbonzini@redhat.com> <1379694292-1601-12-git-send-email-pbonzini@redhat.com> <528A310A.60607@dlhnet.de> <528A3422.1030701@kamp.de> <528A3C4A.4090001@redhat.com> <528B3C86.3030309@kamp.de> <528B41A5.1080504@redhat.com> <528B53A6.3030204@kamp.de> <528B561C.9070200@redhat.com> In-Reply-To: <528B561C.9070200@redhat.com> Content-Type: text/plain; charset=UTF-8; format=flowed Content-Transfer-Encoding: 7bit Subject: Re: [Qemu-devel] [PULL 11/13] target-i386: forward CPUID cache leaves when -cpu host is used List-Id: List-Unsubscribe: , List-Archive: List-Post: List-Help: List-Subscribe: , To: Paolo Bonzini Cc: qemu-devel@nongnu.org, Anthony Liguori , =?UTF-8?B?QmVub8OudCBDYW5ldA==?= 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