| Message ID | 1326066759-8210-2-git-send-email-agraf@suse.de |
|---|---|
| State | New |
| Headers | show |
* Alexander Graf <agraf@suse.de> [2012-01-08 17:53]: > When running QEMU without -cpu parameter, the user usually wants a sane > default. So far, we're using the qemu64/qemu32 CPU type, which basically > means "the maximum TCG can emulate". it also means we all maximum possible migration targets. Have you given any thought to migration with -cpu best? > > That's a really good default when using TCG, but when running with KVM > we much rather want a default saying "the maximum performance I can get". > > Fortunately we just added an option that gives us the best performance > while still staying safe on the testability side of things: -cpu best. > So all we need to do is make -cpu best the default when the user doesn't > define any. > > This fixes a lot of subtile breakage in the GNU toolchain (libgmp) which > hicks up on QEMU's non-existent CPU models. > > This patch also adds a new pc-1.1 machine type to keep backwards compatible > with older versions of QEMU. > > Signed-off-by: Alexander Graf <agraf@suse.de> > --- > hw/pc_piix.c | 42 ++++++++++++++++++++++++++++++++++-------- > 1 files changed, 34 insertions(+), 8 deletions(-) > > diff --git a/hw/pc_piix.c b/hw/pc_piix.c > index 00f525e..3d78ccb 100644 > --- a/hw/pc_piix.c > +++ b/hw/pc_piix.c > @@ -79,7 +79,8 @@ static void pc_init1(MemoryRegion *system_memory, > const char *initrd_filename, > const char *cpu_model, > int pci_enabled, > - int kvmclock_enabled) > + int kvmclock_enabled, > + int may_cpu_best) > { > int i; > ram_addr_t below_4g_mem_size, above_4g_mem_size; > @@ -102,6 +103,9 @@ static void pc_init1(MemoryRegion *system_memory, > MemoryRegion *rom_memory; > DeviceState *dev; > > + if (!cpu_model && kvm_enabled() && may_cpu_best) { > + cpu_model = "best"; > + } > pc_cpus_init(cpu_model); > > if (kvmclock_enabled) { > @@ -263,7 +267,21 @@ static void pc_init_pci(ram_addr_t ram_size, > get_system_io(), > ram_size, boot_device, > kernel_filename, kernel_cmdline, > - initrd_filename, cpu_model, 1, 1); > + initrd_filename, cpu_model, 1, 1, 1); > +} > + > +static void pc_init_pci_oldcpu(ram_addr_t ram_size, > + const char *boot_device, > + const char *kernel_filename, > + const char *kernel_cmdline, > + const char *initrd_filename, > + const char *cpu_model) > +{ > + pc_init1(get_system_memory(), > + get_system_io(), > + ram_size, boot_device, > + kernel_filename, kernel_cmdline, > + initrd_filename, cpu_model, 1, 1, 0); > } > > static void pc_init_pci_no_kvmclock(ram_addr_t ram_size, > @@ -277,7 +295,7 @@ static void pc_init_pci_no_kvmclock(ram_addr_t ram_size, > get_system_io(), > ram_size, boot_device, > kernel_filename, kernel_cmdline, > - initrd_filename, cpu_model, 1, 0); > + initrd_filename, cpu_model, 1, 0, 0); > } > > static void pc_init_isa(ram_addr_t ram_size, > @@ -293,7 +311,7 @@ static void pc_init_isa(ram_addr_t ram_size, > get_system_io(), > ram_size, boot_device, > kernel_filename, kernel_cmdline, > - initrd_filename, cpu_model, 0, 1); > + initrd_filename, cpu_model, 0, 1, 0); > } > > #ifdef CONFIG_XEN > @@ -314,8 +332,8 @@ static void pc_xen_hvm_init(ram_addr_t ram_size, > } > #endif > > -static QEMUMachine pc_machine_v1_0 = { > - .name = "pc-1.0", > +static QEMUMachine pc_machine_v1_1 = { > + .name = "pc-1.1", > .alias = "pc", > .desc = "Standard PC", > .init = pc_init_pci, > @@ -323,17 +341,24 @@ static QEMUMachine pc_machine_v1_0 = { > .is_default = 1, > }; > > +static QEMUMachine pc_machine_v1_0 = { > + .name = "pc-1.0", > + .desc = "Standard PC", > + .init = pc_init_pci_oldcpu, > + .max_cpus = 255, > +}; > + > static QEMUMachine pc_machine_v0_15 = { > .name = "pc-0.15", > .desc = "Standard PC", > - .init = pc_init_pci, > + .init = pc_init_pci_oldcpu, > .max_cpus = 255, > }; > > static QEMUMachine pc_machine_v0_14 = { > .name = "pc-0.14", > .desc = "Standard PC", > - .init = pc_init_pci, > + .init = pc_init_pci_oldcpu, > .max_cpus = 255, > .compat_props = (GlobalProperty[]) { > { > @@ -612,6 +637,7 @@ static QEMUMachine xenfv_machine = { > > static void pc_machine_init(void) > { > + qemu_register_machine(&pc_machine_v1_1); > qemu_register_machine(&pc_machine_v1_0); > qemu_register_machine(&pc_machine_v0_15); > qemu_register_machine(&pc_machine_v0_14); > -- > 1.6.0.2 >
On 16.01.2012, at 20:30, Ryan Harper wrote: > * Alexander Graf <agraf@suse.de> [2012-01-08 17:53]: >> When running QEMU without -cpu parameter, the user usually wants a sane >> default. So far, we're using the qemu64/qemu32 CPU type, which basically >> means "the maximum TCG can emulate". > > it also means we all maximum possible migration targets. Have you > given any thought to migration with -cpu best? If you have the same boxes in your cluster, migration just works. If you don't, you usually use a specific CPU model that is the least dominator between your boxes either way. The current kvm64 type is broken. Libgmp just abort()s when we pass it in. So anything is better than what we do today on AMD hosts :). Alex
* Alexander Graf <agraf@suse.de> [2012-01-16 13:37]: > > On 16.01.2012, at 20:30, Ryan Harper wrote: > > > * Alexander Graf <agraf@suse.de> [2012-01-08 17:53]: > >> When running QEMU without -cpu parameter, the user usually wants a sane > >> default. So far, we're using the qemu64/qemu32 CPU type, which basically > >> means "the maximum TCG can emulate". > > > > it also means we all maximum possible migration targets. Have you > > given any thought to migration with -cpu best? > > If you have the same boxes in your cluster, migration just works. If > you don't, you usually use a specific CPU model that is the least > dominator between your boxes either way. Sure, but the idea behind -cpu best is to not have to figure that out; you had suggested that the qemu64/qemu32 were just related to TCG, and what I'm suggesting is that it's also the most compatible w.r.t migration. it sounds like if migration is a requirement, then -cpu best probably isn't something that would be used. I suppose I'm OK with that, or at least I don't have a better suggestion on how to carefully push up the capabilities without at some point breaking migration. > > The current kvm64 type is broken. Libgmp just abort()s when we pass it > in. So anything is better than what we do today on AMD hosts :). I wonder if it breaks with Cyris cpus... other tools tend to do runtime detection (mplayer). > > > Alex
On 16.01.2012, at 20:46, Ryan Harper wrote: > * Alexander Graf <agraf@suse.de> [2012-01-16 13:37]: >> >> On 16.01.2012, at 20:30, Ryan Harper wrote: >> >>> * Alexander Graf <agraf@suse.de> [2012-01-08 17:53]: >>>> When running QEMU without -cpu parameter, the user usually wants a sane >>>> default. So far, we're using the qemu64/qemu32 CPU type, which basically >>>> means "the maximum TCG can emulate". >>> >>> it also means we all maximum possible migration targets. Have you >>> given any thought to migration with -cpu best? >> >> If you have the same boxes in your cluster, migration just works. If >> you don't, you usually use a specific CPU model that is the least >> dominator between your boxes either way. > > Sure, but the idea behind -cpu best is to not have to figure that out; > you had suggested that the qemu64/qemu32 were just related to TCG, and > what I'm suggesting is that it's also the most compatible w.r.t > migration. The, the most compatible wrt migration is -cpu kvm64 / kvm32. > it sounds like if migration is a requirement, then -cpu best probably > isn't something that would be used. I suppose I'm OK with that, or at > least I don't have a better suggestion on how to carefully push up the > capabilities without at some point breaking migration. Yes, if you're interested in migration, then you're almost guaranteed to have a toolstack on top that has knowledge of your whole cluster and can do the least dominator detection over all of your nodes. On the QEMU level we don't know anything about other machines. > >> >> The current kvm64 type is broken. Libgmp just abort()s when we pass it >> in. So anything is better than what we do today on AMD hosts :). > > I wonder if it breaks with Cyris cpus... other tools tend to do runtime > detection (mplayer). It probably does :). But then again those don't do KVM, do they? Alex
* Alexander Graf <agraf@suse.de> [2012-01-16 13:52]: > > On 16.01.2012, at 20:46, Ryan Harper wrote: > > > * Alexander Graf <agraf@suse.de> [2012-01-16 13:37]: > >> > >> On 16.01.2012, at 20:30, Ryan Harper wrote: > >> > >>> * Alexander Graf <agraf@suse.de> [2012-01-08 17:53]: > >>>> When running QEMU without -cpu parameter, the user usually wants a sane > >>>> default. So far, we're using the qemu64/qemu32 CPU type, which basically > >>>> means "the maximum TCG can emulate". > >>> > >>> it also means we all maximum possible migration targets. Have you > >>> given any thought to migration with -cpu best? > >> > >> If you have the same boxes in your cluster, migration just works. If > >> you don't, you usually use a specific CPU model that is the least > >> dominator between your boxes either way. > > > > Sure, but the idea behind -cpu best is to not have to figure that out; > > you had suggested that the qemu64/qemu32 were just related to TCG, and > > what I'm suggesting is that it's also the most compatible w.r.t > > migration. > > The, the most compatible wrt migration is -cpu kvm64 / kvm32. > > > it sounds like if migration is a requirement, then -cpu best probably > > isn't something that would be used. I suppose I'm OK with that, or at > > least I don't have a better suggestion on how to carefully push up the > > capabilities without at some point breaking migration. > > Yes, if you're interested in migration, then you're almost guaranteed to have a toolstack on top that has knowledge of your whole cluster and can do the least dominator detection over all of your nodes. On the QEMU level we don't know anything about other machines. > > > > >> > >> The current kvm64 type is broken. Libgmp just abort()s when we pass it > >> in. So anything is better than what we do today on AMD hosts :). > > > > I wonder if it breaks with Cyris cpus... other tools tend to do runtime > > detection (mplayer). > > It probably does :). But then again those don't do KVM, do they? not following; mplayer issues SSE2, 3 and 4 instructions to see what works to figure out how to optimize; it doesn't care if the cpu is called QEMU64 or Cyrus or AMD. I'm not saying that we can't do better than qemu64 w.r.t best cpu to select by default, but there are plenty of applications that want to optimize their code based on what's available, but this is done via code execution instead of string comparision. > > > Alex >
On 16.01.2012, at 21:13, Ryan Harper wrote: > * Alexander Graf <agraf@suse.de> [2012-01-16 13:52]: >> >> On 16.01.2012, at 20:46, Ryan Harper wrote: >> >>> * Alexander Graf <agraf@suse.de> [2012-01-16 13:37]: >>>> >>>> On 16.01.2012, at 20:30, Ryan Harper wrote: >>>> >>>>> * Alexander Graf <agraf@suse.de> [2012-01-08 17:53]: >>>>>> When running QEMU without -cpu parameter, the user usually wants a sane >>>>>> default. So far, we're using the qemu64/qemu32 CPU type, which basically >>>>>> means "the maximum TCG can emulate". >>>>> >>>>> it also means we all maximum possible migration targets. Have you >>>>> given any thought to migration with -cpu best? >>>> >>>> If you have the same boxes in your cluster, migration just works. If >>>> you don't, you usually use a specific CPU model that is the least >>>> dominator between your boxes either way. >>> >>> Sure, but the idea behind -cpu best is to not have to figure that out; >>> you had suggested that the qemu64/qemu32 were just related to TCG, and >>> what I'm suggesting is that it's also the most compatible w.r.t >>> migration. >> >> The, the most compatible wrt migration is -cpu kvm64 / kvm32. >> >>> it sounds like if migration is a requirement, then -cpu best probably >>> isn't something that would be used. I suppose I'm OK with that, or at >>> least I don't have a better suggestion on how to carefully push up the >>> capabilities without at some point breaking migration. >> >> Yes, if you're interested in migration, then you're almost guaranteed to have a toolstack on top that has knowledge of your whole cluster and can do the least dominator detection over all of your nodes. On the QEMU level we don't know anything about other machines. >> >>> >>>> >>>> The current kvm64 type is broken. Libgmp just abort()s when we pass it >>>> in. So anything is better than what we do today on AMD hosts :). >>> >>> I wonder if it breaks with Cyris cpus... other tools tend to do runtime >>> detection (mplayer). >> >> It probably does :). But then again those don't do KVM, do they? > > not following; mplayer issues SSE2, 3 and 4 instructions to see what > works to figure out how to optimize; it doesn't care if the cpu is > called QEMU64 or Cyrus or AMD. I'm not saying that we can't do better > than qemu64 w.r.t best cpu to select by default, but there are plenty of > applications that want to optimize their code based on what's available, > but this is done via code execution instead of string comparision. The problem with -cpu kvm64 is that we choose a family/model that doesn't exist in the real world, and then glue AuthenticAMD or GenuineIntel in the vendor string. Libgmp checks for existing CPUs, finds that this CPU doesn't match any real world IDs and abort()s. The problem is that there is not a single CPU on this planet in silicon that has the same model+family numbers, but exists in AuthenticAMD _and_ GenuineIntel flavors. We need to pass the host vendor in though, because the guest uses it to detect if it should execute SYSCALL or SYSENTER, because intel and amd screwed up heavily on that one. It's not about feature flags which is what mplayer uses. Those are fine :). Alex
* Alexander Graf <agraf@suse.de> [2012-01-16 14:52]: > > On 16.01.2012, at 21:13, Ryan Harper wrote: > > > * Alexander Graf <agraf@suse.de> [2012-01-16 13:52]: > >> > >> On 16.01.2012, at 20:46, Ryan Harper wrote: > >> > >>> * Alexander Graf <agraf@suse.de> [2012-01-16 13:37]: > >>>> > >>>> On 16.01.2012, at 20:30, Ryan Harper wrote: > >>>> > >>>>> * Alexander Graf <agraf@suse.de> [2012-01-08 17:53]: > >>>>>> When running QEMU without -cpu parameter, the user usually wants a sane > >>>>>> default. So far, we're using the qemu64/qemu32 CPU type, which basically > >>>>>> means "the maximum TCG can emulate". > >>>>> > >>>>> it also means we all maximum possible migration targets. Have you > >>>>> given any thought to migration with -cpu best? > >>>> > >>>> If you have the same boxes in your cluster, migration just works. If > >>>> you don't, you usually use a specific CPU model that is the least > >>>> dominator between your boxes either way. > >>> > >>> Sure, but the idea behind -cpu best is to not have to figure that out; > >>> you had suggested that the qemu64/qemu32 were just related to TCG, and > >>> what I'm suggesting is that it's also the most compatible w.r.t > >>> migration. > >> > >> The, the most compatible wrt migration is -cpu kvm64 / kvm32. > >> > >>> it sounds like if migration is a requirement, then -cpu best probably > >>> isn't something that would be used. I suppose I'm OK with that, or at > >>> least I don't have a better suggestion on how to carefully push up the > >>> capabilities without at some point breaking migration. > >> > >> Yes, if you're interested in migration, then you're almost guaranteed to have a toolstack on top that has knowledge of your whole cluster and can do the least dominator detection over all of your nodes. On the QEMU level we don't know anything about other machines. > >> > >>> > >>>> > >>>> The current kvm64 type is broken. Libgmp just abort()s when we pass it > >>>> in. So anything is better than what we do today on AMD hosts :). > >>> > >>> I wonder if it breaks with Cyris cpus... other tools tend to do runtime > >>> detection (mplayer). > >> > >> It probably does :). But then again those don't do KVM, do they? > > > > not following; mplayer issues SSE2, 3 and 4 instructions to see what > > works to figure out how to optimize; it doesn't care if the cpu is > > called QEMU64 or Cyrus or AMD. I'm not saying that we can't do better > > than qemu64 w.r.t best cpu to select by default, but there are plenty of > > applications that want to optimize their code based on what's available, > > but this is done via code execution instead of string comparision. > > The problem with -cpu kvm64 is that we choose a family/model that > doesn't exist in the real world, and then glue AuthenticAMD or > GenuineIntel in the vendor string. Libgmp checks for existing CPUs, > finds that this CPU doesn't match any real world IDs and abort()s. > > The problem is that there is not a single CPU on this planet in > silicon that has the same model+family numbers, but exists in > AuthenticAMD _and_ GenuineIntel flavors. We need to pass the host > vendor in though, because the guest uses it to detect if it should > execute SYSCALL or SYSENTER, because intel and amd screwed up heavily > on that one. I forgot about this one. =(
diff --git a/hw/pc_piix.c b/hw/pc_piix.c index 00f525e..3d78ccb 100644 --- a/hw/pc_piix.c +++ b/hw/pc_piix.c @@ -79,7 +79,8 @@ static void pc_init1(MemoryRegion *system_memory, const char *initrd_filename, const char *cpu_model, int pci_enabled, - int kvmclock_enabled) + int kvmclock_enabled, + int may_cpu_best) { int i; ram_addr_t below_4g_mem_size, above_4g_mem_size; @@ -102,6 +103,9 @@ static void pc_init1(MemoryRegion *system_memory, MemoryRegion *rom_memory; DeviceState *dev; + if (!cpu_model && kvm_enabled() && may_cpu_best) { + cpu_model = "best"; + } pc_cpus_init(cpu_model); if (kvmclock_enabled) { @@ -263,7 +267,21 @@ static void pc_init_pci(ram_addr_t ram_size, get_system_io(), ram_size, boot_device, kernel_filename, kernel_cmdline, - initrd_filename, cpu_model, 1, 1); + initrd_filename, cpu_model, 1, 1, 1); +} + +static void pc_init_pci_oldcpu(ram_addr_t ram_size, + const char *boot_device, + const char *kernel_filename, + const char *kernel_cmdline, + const char *initrd_filename, + const char *cpu_model) +{ + pc_init1(get_system_memory(), + get_system_io(), + ram_size, boot_device, + kernel_filename, kernel_cmdline, + initrd_filename, cpu_model, 1, 1, 0); } static void pc_init_pci_no_kvmclock(ram_addr_t ram_size, @@ -277,7 +295,7 @@ static void pc_init_pci_no_kvmclock(ram_addr_t ram_size, get_system_io(), ram_size, boot_device, kernel_filename, kernel_cmdline, - initrd_filename, cpu_model, 1, 0); + initrd_filename, cpu_model, 1, 0, 0); } static void pc_init_isa(ram_addr_t ram_size, @@ -293,7 +311,7 @@ static void pc_init_isa(ram_addr_t ram_size, get_system_io(), ram_size, boot_device, kernel_filename, kernel_cmdline, - initrd_filename, cpu_model, 0, 1); + initrd_filename, cpu_model, 0, 1, 0); } #ifdef CONFIG_XEN @@ -314,8 +332,8 @@ static void pc_xen_hvm_init(ram_addr_t ram_size, } #endif -static QEMUMachine pc_machine_v1_0 = { - .name = "pc-1.0", +static QEMUMachine pc_machine_v1_1 = { + .name = "pc-1.1", .alias = "pc", .desc = "Standard PC", .init = pc_init_pci, @@ -323,17 +341,24 @@ static QEMUMachine pc_machine_v1_0 = { .is_default = 1, }; +static QEMUMachine pc_machine_v1_0 = { + .name = "pc-1.0", + .desc = "Standard PC", + .init = pc_init_pci_oldcpu, + .max_cpus = 255, +}; + static QEMUMachine pc_machine_v0_15 = { .name = "pc-0.15", .desc = "Standard PC", - .init = pc_init_pci, + .init = pc_init_pci_oldcpu, .max_cpus = 255, }; static QEMUMachine pc_machine_v0_14 = { .name = "pc-0.14", .desc = "Standard PC", - .init = pc_init_pci, + .init = pc_init_pci_oldcpu, .max_cpus = 255, .compat_props = (GlobalProperty[]) { { @@ -612,6 +637,7 @@ static QEMUMachine xenfv_machine = { static void pc_machine_init(void) { + qemu_register_machine(&pc_machine_v1_1); qemu_register_machine(&pc_machine_v1_0); qemu_register_machine(&pc_machine_v0_15); qemu_register_machine(&pc_machine_v0_14);
When running QEMU without -cpu parameter, the user usually wants a sane default. So far, we're using the qemu64/qemu32 CPU type, which basically means "the maximum TCG can emulate". That's a really good default when using TCG, but when running with KVM we much rather want a default saying "the maximum performance I can get". Fortunately we just added an option that gives us the best performance while still staying safe on the testability side of things: -cpu best. So all we need to do is make -cpu best the default when the user doesn't define any. This fixes a lot of subtile breakage in the GNU toolchain (libgmp) which hicks up on QEMU's non-existent CPU models. This patch also adds a new pc-1.1 machine type to keep backwards compatible with older versions of QEMU. Signed-off-by: Alexander Graf <agraf@suse.de> --- hw/pc_piix.c | 42 ++++++++++++++++++++++++++++++++++-------- 1 files changed, 34 insertions(+), 8 deletions(-)