hvf: arm: Support creating VMs with 64+GB of RAM on macOS 15+
v2 cover hvf: arm: Support creating VMs with 64+GB of RAM on macOS 15+
This patchsets focus is on lighting up the ability to create VMs with 64+GB of RAM through using some new APIs introduced in macOS 13. Due to the IPA sizes supported in macOS, the first version we can properly support this requirement is macOS 15 as (if the hardware supports it also) the kernel adds support for a 40b IPA size, which is the first supported ARM PARange value after 36b, so we can advertise this to the guest properly as well in id_aa64mmfr0_el1.
Today if you asked for a > 64GB VM you'd be met with a pretty unwieldy HV_BAD_ARGUMENT. On machines without 40b IPA support this patchset also improves this, and the message mirrors the kvm_type error you'd get on ARM:
"qemu-system-aarch64: -accel hvf: Addressing limited to 36 bits, but memory exceeds it by 18253611008 bytes"
Changes from V1 to V2 (Thanks Peter for review!):
- Added a new function pointer to MachineClass to be able to freeze the memory map and compute the highest guest physical address. We use this to inform VM creation on what IPA size we should ask the kernel for. This is very similar to what ARM's kvm_type() does.
- Fixed redundant loop in round_down_to_parange_bit_size
- Move the splitting up of hv_vm_create logic per platform to a separate patch. This is mostly for readability.
Danny Canter (3):
hw/boards: Add hvf_get_physical_address_range to MachineClass
hvf: Split up hv_vm_create logic per arch
hvf: arm: Allow creating VMs with 64+GB of RAM on macOS 15+
accel/hvf/hvf-accel-ops.c | 16 +++++++--- hw/arm/virt.c | 42 ++++++++++++++++++++++++- hw/i386/x86.c | 2 ++ include/hw/boards.h | 5 +++ include/sysemu/hvf_int.h | 1 + target/arm/hvf/hvf.c | 66 +++++++++++++++++++++++++++++++++++++++ target/arm/hvf_arm.h | 3 ++ target/arm/internals.h | 19 +++++++++++ target/arm/ptw.c | 15 +++++++++ target/i386/hvf/hvf.c | 5 +++ 10 files changed, 168 insertions(+), 6 deletions(-)
Signature · 1 line
2.39.5 (Apple Git-154)
PATCH v2 1/3 hw/boards: Add hvf_get_physical_address_range to MachineClass
This addition will be necessary for some HVF related work to follow. For HVF on ARM there exists a set of APIs in macOS 13 to be able to adjust the IPA size for a given VM. This is useful as by default HVF uses 36 bits as the IPA size, so to support guests with > 64GB of RAM we'll need to reach for this.
To have all the info necessary to carry this out however, we need some plumbing to be able to grab the memory map and compute the highest GPA prior to creating the VM. This is almost exactly like what kvm_type is used for on ARM today, and is also what this will be used for. We will compute the highest GPA and find what IPA size we'd need to satisfy this, and if it's valid (macOS today caps at 40b) we'll set this to be the IPA size in coming patches. This new method is only needed (today at least) on ARM, and obviously only for HVF/macOS, so admittedly it is much less generic than kvm_type today, but it seemed a somewhat sane way to get the information we need from the memmap at VM creation time.
--- hw/arm/virt.c | 9 ++++++++- hw/i386/x86.c | 2 ++ include/hw/boards.h | 5 +++++ 3 files changed, 15 insertions(+), 1 deletion(-)
hw/arm/virt.c +8 -1
@@ -2107,7 +2107,8 @@ static void machvirt_init(MachineState *machine)/* * In accelerated mode, the memory map is computed earlier in kvm_type() -* to create a VM with the right number of IPA bits. +* for Linux, or hvf_get_physical_address_range() for macOS to create a +* VM with the right number of IPA bits. */ if (!vms->memmap) {Object *cpuobj;
@@ -3027,6 +3028,11 @@ static int virt_kvm_type(MachineState *ms, const char *type_str)return fixed_ipa ?0 : requested_pa_size; } +static int virt_hvf_get_physical_address_range (MachineState *ms) +{ +return 0 ; +} +static void virt_machine_class_init (ObjectClass *oc,void *data) {MachineClass *mc =MACHINE_CLASS (oc);
@@ -3086,6 +3092,7 @@ static void virt_machine_class_init(ObjectClass *oc, void *data) mc->valid_cpu_types = valid_cpu_types; mc->get_default_cpu_node_id = virt_get_default_cpu_node_id; mc->kvm_type = virt_kvm_type; + mc->hvf_get_physical_address_range = virt_hvf_get_physical_address_range;assert (!mc->get_hotplug_handler); mc->get_hotplug_handler = virt_machine_get_hotplug_handler; hc->pre_plug = virt_machine_device_pre_plug_cb;
hw/i386/x86.c +2 -0
@@ -382,6 +382,8 @@ static void x86_machine_class_init(ObjectClass *oc, void *data) mc->get_default_cpu_node_id = x86_get_default_cpu_node_id; mc->possible_cpu_arch_ids = x86_possible_cpu_arch_ids; mc->kvm_type = x86_kvm_type; +/* Not needed for x86 */ + mc->hvf_get_physical_address_range =NULL ; x86mc->save_tsc_khz = true; x86mc->fwcfg_dma_enabled = true; nc->nmi_monitor_handler = x86_nmi;
include/hw/boards.h +5 -0
@@ -215,6 +215,10 @@ typedef struct {* Return the type of KVM corresponding to the kvm-type string option or * computed based on other criteria such as the host kernel capabilities. * kvm-type may be NULL if it is not needed. +* @hvf_get_physical_address_range: +* Returns the physical address range in bits to use for the HVF virtual +* machine based on the current boards memory map. This may be NULL if it +* is not needed. * @numa_mem_supported: * true if '--numa node.mem' option is supported and false otherwise * @hotplug_allowed:
@@ -256,6 +260,7 @@ struct MachineClass {void (*reset)(MachineState *state,ShutdownCause reason);void (*wakeup)(MachineState *state);int (*kvm_type)(MachineState *machine,const char *arg); +int (*hvf_get_physical_address_range)(MachineState *machine);BlockInterfaceType block_default_type;int units_per_default_bus;
Signature · 1 line
2.39.5 (Apple Git-154)
Re: [PATCH v2 1/3] hw/boards: Add hvf_get_physical_address_range to MachineClass
Quoted 25 lines · Danny Canter
This addition will be necessary for some HVF related work to follow. For HVF on ARM there exists a set of APIs in macOS 13 to be able to adjust the IPA size for a given VM. This is useful as by default HVF uses 36 bits as the IPA size, so to support guests with > 64GB of RAM we'll need to reach for this.
To have all the info necessary to carry this out however, we need some plumbing to be able to grab the memory map and compute the highest GPA prior to creating the VM. This is almost exactly like what
kvm_typeis used for on ARM today, and is also what this will be used for. We will compute the highest GPA and find what IPA size we'd need to satisfy this, and if it's valid (macOS today caps at 40b) we'll set this to be the IPA size in coming patches. This new method is only needed (today at least) on ARM, and obviously only for HVF/macOS, so admittedly it is much less generic thankvm_typetoday, but it seemed a somewhat sane way to get the information we need from the memmap at VM creation time.Signed-off-by Danny Canter <[email protected]>hw/i386/x86.c +0
-0
hw/i386/x86.c static void x86_machine_class_init(ObjectClass *oc, void *data) · line 389
Quoted 8 lines
hw/i386/x86.c +2
-0@@ -382,6 +382,8 @@ static void x86_machine_class_init(ObjectClass *oc, void *data) mc->get_default_cpu_node_id = x86_get_default_cpu_node_id; mc->possible_cpu_arch_ids = x86_possible_cpu_arch_ids; mc->kvm_type = x86_kvm_type; +/* Not needed for x86 */ + mc->hvf_get_physical_address_range =NULL ; x86mc->save_tsc_khz = true; x86mc->fwcfg_dma_enabled = true; nc->nmi_monitor_handler = x86_nmi;
We guarantee that object and class structs are zero-initialized, so we don't need to explicitly set this field to NULL.
Otherwise
thanks
-- PMM
PATCH v2 2/3 hvf: Split up hv_vm_create logic per arch
In reply to Danny CanterThis is preliminary work to split up hv_vm_create logic per platform so we can support creating VMs with > 64GB of RAM on Apple Silicon machines. This is done via ARM HVF's hv_vm_config_create() (and other APIs that modify this config that will be coming in future patches). This should have no behavioral difference at all as hv_vm_config_create() just assigns the same default values as if you just passed NULL to the function.
--- accel/hvf/hvf-accel-ops.c | 6 +----- include/sysemu/hvf_int.h | 1 + target/arm/hvf/hvf.c | 9 +++++++++ target/i386/hvf/hvf.c | 5 +++++ 4 files changed, 16 insertions(+), 5 deletions(-)
accel/hvf/hvf-accel-ops.c +1 -5
@@ -61,10 +61,6 @@HVFState *hvf_state; -#ifdef __aarch64__ -#define HV_VM_DEFAULT NULL -#endif -/* Memory slots */ hvf_slot *hvf_find_overlap_slot (uint64_t start,uint64_t size)
@@ -324,7 +320,7 @@ static int hvf_accel_init(MachineState *ms)hv_return_t ret;HVFState *s; - ret =hv_vm_create (HV_VM_DEFAULT ); + ret =hvf_arch_vm_create (ms,0 );assert_hvf_ok (ret); s =g_new0 (HVFState,1 );
include/sysemu/hvf_int.h +1 -0
@@ -65,6 +65,7 @@ void assert_hvf_ok_impl(hv_return_t ret, const char *file, unsigned int line,#define assert_hvf_ok (EX ) assert_hvf_ok_impl((EX), __FILE__, __LINE__, #EX)const char *hvf_return_string (hv_return_t ret);int hvf_arch_init (void ); +hv_return_t hvf_arch_vm_create (MachineState *ms,uint32_t pa_range);int hvf_arch_init_vcpu (CPUState *cpu);void hvf_arch_vcpu_destroy (CPUState *cpu);int hvf_vcpu_exec (CPUState *);
target/arm/hvf/hvf.c +9 -0
@@ -929,6 +929,15 @@ void hvf_arch_vcpu_destroy(CPUState *cpu) { } +hv_return_t hvf_arch_vm_create (MachineState *ms,uint32_t pa_range) +{ +hv_vm_config_t config =hv_vm_config_create (); +hv_return_t ret =hv_vm_create (config); +os_release (config); + +return ret; +} +int hvf_arch_init_vcpu (CPUState *cpu) {ARMCPU *arm_cpu =ARM_CPU (cpu);
target/i386/hvf/hvf.c +5 -0
@@ -223,6 +223,11 @@ int hvf_arch_init(void)return 0 ; } +hv_return_t hvf_arch_vm_create (MachineState *ms,uint32_t pa_range) +{ +return hv_vm_create (HV_VM_DEFAULT ); +} +int hvf_arch_init_vcpu (CPUState *cpu) {X86CPU *x86cpu =X86_CPU (cpu);
Signature · 1 line
2.39.5 (Apple Git-154)
Re: [PATCH v2 2/3] hvf: Split up hv_vm_create logic per arch
Quoted 11 lines · Danny Canter
This is preliminary work to split up
hv_vm_createlogic per platform so we can support creating VMs with > 64GB of RAM on Apple Silicon machines. This is done via ARM HVF'shv_vm_config_create()(and other APIs that modify this config that will be coming in future patches). This should have no behavioral difference at all ashv_vm_config_create()just assigns the same default values as if you just passed NULL to the function.
Quoted 1 line
Signed-off-by Danny Canter <[email protected]>
thanks
-- PMM
PATCH v2 3/3 hvf: arm: Implement and use hvf_get_physical_address_range
In reply to Danny CanterThis patch's main focus is to use the previously added hvf_get_physical_address_range to inform VM creation about the IPA size we need for the VM, so we can extend the default 36b IPA size and support VMs with 64+GB of RAM. This is done by freezing the memory map, computing the highest GPA and then (depending on if the platform supports an IPA size that large) telling the kernel to use a size >= for the VM. In pursuit of this a couple of things related to how we handle the physical address range we expose to guests were altered, but for an explanation of what we were doing:
Today, to get the IPA size we were reading id_aa64mmfr0_el1's PARange field from a newly made vcpu. Unfortunately, HVF just returns the hosts PARange directly for the initial value and not the IPA size that will actually back the VM, so we believe we have much more address space than we actually do today it seems.
Starting in macOS 13.0 some APIs were introduced to be able to query the maximum IPA size the kernel supports, and to set the IPA size for a given VM. However, this still has a couple of issues on < macOS 15. Up until macOS 15 (and if the hardware supported it) the max IPA size was 39 bits which is not a valid PARange value, so we can't clamp down what we advertise in the vcpu's id_aa64mmfr0_el1 to our IPA size. Starting in macOS 15 however, the maximum IPA size is 40 bits (if it's supported in the hardware as well) which is also a valid PARange value so we can set our IPA size to the maximum as well as clamp down the PARange we advertise to the guest. This allows VMs with 64+ GB of RAM and should fix the oddness of the PARange situation as well.
--- accel/hvf/hvf-accel-ops.c | 12 ++++++++- hw/arm/virt.c | 31 +++++++++++++++++++++- target/arm/hvf/hvf.c | 56 ++++++++++++++++++++++++++++++++++++++- target/arm/hvf_arm.h | 19 +++++++++++++ target/arm/internals.h | 19 +++++++++++++ target/arm/ptw.c | 15 +++++++++++ 6 files changed, 149 insertions(+), 3 deletions(-)
accel/hvf/hvf-accel-ops.c +11 -1
@@ -53,6 +53,7 @@#include "exec/address-spaces.h" #include "exec/exec-all.h" #include "gdbstub/enums.h" +#include "hw/boards.h" #include "sysemu/cpus.h" #include "sysemu/hvf.h" #include "sysemu/hvf_int.h"
@@ -319,8 +320,17 @@ static int hvf_accel_init(MachineState *ms)int x;hv_return_t ret;HVFState *s; +int pa_range =36 ; +MachineClass *mc =MACHINE_GET_CLASS (ms); - ret =hvf_arch_vm_create (ms,0 ); +if (mc->hvf_get_physical_address_range) { + pa_range = mc->hvf_get_physical_address_range (ms); +if (pa_range <0 ) { +return -EINVAL ; + } + } + + ret =hvf_arch_vm_create (ms, (uint32_t )pa_range);assert_hvf_ok (ret); s =g_new0 (HVFState,1 );
hw/arm/virt.c +30 -1
@@ -66,6 +66,7 @@#include "hw/intc/arm_gicv3_its_common.h" #include "hw/irq.h" #include "kvm_arm.h" +#include "hvf_arm.h" #include "hw/firmware/smbios.h" #include "qapi/visitor.h" #include "qapi/qapi-visit-common.h"
@@ -3030,7 +3031,35 @@ static int virt_kvm_type(MachineState *ms, const char *type_str)static int virt_hvf_get_physical_address_range (MachineState *ms) { -return 0 ; +VirtMachineState *vms =VIRT_MACHINE (ms); + +int default_ipa_size =hvf_arm_get_default_ipa_bit_size (); +int max_ipa_size =hvf_arm_get_max_ipa_bit_size (); + +/* We freeze the memory map to compute the highest gpa */ +virt_set_memmap (vms, max_ipa_size); + +int requested_ipa_size =64 -clz64 (vms->highest_gpa); + +/* +* If we're <= the default IPA size just use the default. +* If we're above the default but below the maximum, round up to +* the maximum. hvf_arm_get_max_ipa_bit_size() conveniently only +* returns values that are valid ARM PARange values. +*/ +if (requested_ipa_size <= default_ipa_size) { + requested_ipa_size = default_ipa_size; + }else if (requested_ipa_size <= max_ipa_size) { + requested_ipa_size = max_ipa_size; + }else { +error_report ("-m and ,maxmem option values " +"require an IPA range (%d bits) larger than " +"the one supported by the host (%d bits)" , + requested_ipa_size, max_ipa_size); +return -1 ; + } + +return requested_ipa_size; }static void virt_machine_class_init (ObjectClass *oc,void *data)
target/arm/hvf/hvf.c +55 -1
99 lines not on this page · raw message
target/arm/hvf_arm.h +19 -0
@@ -22,4 +22,23 @@ void hvf_arm_init_debug(void);void hvf_arm_set_cpu_features_from_host (ARMCPU *cpu); +#ifdef CONFIG_HVF + +uint32_t hvf_arm_get_default_ipa_bit_size (void ); +uint32_t hvf_arm_get_max_ipa_bit_size (void ); + +#else + +static inline uint32_t hvf_arm_get_default_ipa_bit_size (void ) +{ +return 0 ; +} + +static inline uint32_t hvf_arm_get_max_ipa_bit_size (void ) +{ +return 0 ; +} + +#endif +#endif
target/arm/internals.h +19 -0
@@ -450,6 +450,25 @@ static inline void update_spsel(CPUARMState *env, uint32_t imm)*/ unsigned int arm_pamax (ARMCPU *cpu); +/* +* round_down_to_parange_index +* @bit_size: uint8_t +* +* Rounds down the bit_size supplied to the first supported ARM physical +* address range and returns the index for this. The index is intended to +* be used to set ID_AA64MMFR0_EL1's PARANGE bits. +*/ +uint8_t round_down_to_parange_index (uint8_t bit_size); + +/* +* round_down_to_parange_bit_size +* @bit_size: uint8_t +* +* Rounds down the bit_size supplied to the first supported ARM physical +* address range bit size and returns this. +*/ +uint8_t round_down_to_parange_bit_size (uint8_t bit_size); +/* Return true if extended addresses are enabled. * This is always the case if our translation regime is 64 bit, * but depends on TTBCR.EAE for 32 bit.
target/arm/ptw.c +15 -0
@@ -96,6 +96,21 @@ static const uint8_t pamax_map[] = { [6] = 52, }; +uint8_t round_down_to_parange_index (uint8_t bit_size) +{ +for (int i =ARRAY_SIZE (pamax_map) -1 ; i >=0 ; i--) { +if (pamax_map[i] <= bit_size) { +return i; + } + } +g_assert_not_reached (); +} + +uint8_t round_down_to_parange_bit_size (uint8_t bit_size) +{ +return pamax_map[round_down_to_parange_index (bit_size)]; +} +/* * The cpu-specific constant value of PAMax; also used by hw/arm/virt. * Note that machvirt_init calls this on a CPU that is inited but not realized!
Signature · 1 line
2.39.5 (Apple Git-154)
Re: [PATCH v2 3/3] hvf: arm: Implement and use hvf_get_physical_address_range
Quoted 32 lines · Danny Canter
This patch's main focus is to use the previously added
hvf_get_physical_address_rangeto inform VM creation about the IPA size we need for the VM, so we can extend the default 36b IPA size and support VMs with 64+GB of RAM. This is done by freezing the memory map, computing the highest GPA and then (depending on if the platform supports an IPA size that large) telling the kernel to use a size >= for the VM. In pursuit of this a couple of things related to how we handle the physical address range we expose to guests were altered, but for an explanation of what we were doing:Today, to get the IPA size we were reading id_aa64mmfr0_el1's PARange field from a newly made vcpu. Unfortunately, HVF just returns the hosts PARange directly for the initial value and not the IPA size that will actually back the VM, so we believe we have much more address space than we actually do today it seems.
Starting in macOS 13.0 some APIs were introduced to be able to query the maximum IPA size the kernel supports, and to set the IPA size for a given VM. However, this still has a couple of issues on < macOS 15. Up until macOS 15 (and if the hardware supported it) the max IPA size was 39 bits which is not a valid PARange value, so we can't clamp down what we advertise in the vcpu's
id_aa64mmfr0_el1to our IPA size. Starting in macOS 15 however, the maximum IPA size is 40 bits (if it's supported in the hardware as well) which is also a valid PARange value so we can set our IPA size to the maximum as well as clamp down the PARange we advertise to the guest. This allows VMs with 64+ GB of RAM and should fix the oddness of the PARange situation as well.
Quoted 1 line
Signed-off-by Danny Canter <[email protected]>
thanks
-- PMM
Re: [PATCH v2 3/3] hvf: arm: Implement and use hvf_get_physical_address_range
In reply to Danny CanterHi Danny,
Quoted 18 lines · Danny Canter
This patch's main focus is to use the previously added
hvf_get_physical_address_rangeto inform VM creation about the IPA size we need for the VM, so we can extend the default 36b IPA size and support VMs with 64+GB of RAM. This is done by freezing the memory map, computing the highest GPA and then (depending on if the platform supports an IPA size that large) telling the kernel to use a size >= for the VM. In pursuit of this a couple of things related to how we handle the physical address range we expose to guests were altered, but for an explanation of what we were doing:Today, to get the IPA size we were reading id_aa64mmfr0_el1's PARange field from a newly made vcpu. Unfortunately, HVF just returns the hosts PARange directly for the initial value and not the IPA size that will actually back the VM, so we believe we have much more address space than we actually do today it seems.
Quoted 12 lines
Starting in macOS 13.0 some APIs were introduced to be able to query the maximum IPA size the kernel supports, and to set the IPA size for a given VM. However, this still has a couple of issues on < macOS 15. Up until macOS 15 (and if the hardware supported it) the max IPA size was 39 bits which is not a valid PARange value, so we can't clamp down what we advertise in the vcpu's
id_aa64mmfr0_el1to our IPA size. Starting in macOS 15 however, the maximum IPA size is 40 bits (if it's supported in the hardware as well) which is also a valid PARange value so we can set our IPA size to the maximum as well as clamp down the PARange we advertise to the guest. This allows VMs with 64+ GB of RAM and should fix the oddness of the PARange situation as well.
Could you have a look at the following issue related to your patch?
https://gitlab.com/qemu-project/qemu/-/issues/2800
Quoted 285 lines
Signed-off-by Danny Canter <[email protected]>--- accel/hvf/hvf-accel-ops.c | 12 ++++++++- hw/arm/virt.c | 31 +++++++++++++++++++++- target/arm/hvf/hvf.c | 56 ++++++++++++++++++++++++++++++++++++++- target/arm/hvf_arm.h | 19 +++++++++++++ target/arm/internals.h | 19 +++++++++++++ target/arm/ptw.c | 15 +++++++++++ 6 files changed, 149 insertions(+), 3 deletions(-)accel/hvf/hvf-accel-ops.c +11
-1@@ -53,6 +53,7 @@#include "exec/address-spaces.h" #include "exec/exec-all.h" #include "gdbstub/enums.h" +#include "hw/boards.h" #include "sysemu/cpus.h" #include "sysemu/hvf.h" #include "sysemu/hvf_int.h" @@ -319,8 +320,17 @@ static int hvf_accel_init(MachineState *ms)int x;hv_return_t ret;HVFState *s; +int pa_range =36 ; +MachineClass *mc =MACHINE_GET_CLASS (ms); - ret =hvf_arch_vm_create (ms,0 ); +if (mc->hvf_get_physical_address_range) { + pa_range = mc->hvf_get_physical_address_range (ms); +if (pa_range <0 ) { +return -EINVAL ; + } + } + + ret =hvf_arch_vm_create (ms, (uint32_t )pa_range);assert_hvf_ok (ret); s =g_new0 (HVFState,1 );hw/arm/virt.c +30
-1@@ -66,6 +66,7 @@#include "hw/intc/arm_gicv3_its_common.h" #include "hw/irq.h" #include "kvm_arm.h" +#include "hvf_arm.h" #include "hw/firmware/smbios.h" #include "qapi/visitor.h" #include "qapi/qapi-visit-common.h" @@ -3030,7 +3031,35 @@ static int virt_kvm_type(MachineState *ms, const char *type_str)static int virt_hvf_get_physical_address_range (MachineState *ms) { -return 0 ; +VirtMachineState *vms =VIRT_MACHINE (ms); + +int default_ipa_size =hvf_arm_get_default_ipa_bit_size (); +int max_ipa_size =hvf_arm_get_max_ipa_bit_size (); + +/* We freeze the memory map to compute the highest gpa */ +virt_set_memmap (vms, max_ipa_size); + +int requested_ipa_size =64 -clz64 (vms->highest_gpa); + +/* +* If we're <= the default IPA size just use the default. +* If we're above the default but below the maximum, round up to +* the maximum. hvf_arm_get_max_ipa_bit_size() conveniently only +* returns values that are valid ARM PARange values. +*/ +if (requested_ipa_size <= default_ipa_size) { + requested_ipa_size = default_ipa_size; + }else if (requested_ipa_size <= max_ipa_size) { + requested_ipa_size = max_ipa_size; + }else { +error_report ("-m and ,maxmem option values " +"require an IPA range (%d bits) larger than " +"the one supported by the host (%d bits)" , + requested_ipa_size, max_ipa_size); +return -1 ; + } + +return requested_ipa_size; }static void virt_machine_class_init (ObjectClass *oc,void *data)target/arm/hvf/hvf.c +55
-1@@ -22,6 +22,7 @@#include <mach/mach_time.h> #include "exec/address-spaces.h" +#include "hw/boards.h" #include "hw/irq.h" #include "qemu/main-loop.h" #include "sysemu/cpus.h" @@ -297,6 +298,8 @@ void hvf_arm_init_debug(void)static void hvf_wfi (CPUState *cpu); +static uint32_t chosen_ipa_bit_size; + typedef struct HVFVTimer { /* Vtimer value during migration and paused state */ uint64_t vtimer_val;@@ -839,6 +842,16 @@ static uint64_t hvf_get_reg(CPUState *cpu, int rt)return val; } +static void clamp_id_aa64mmfr0_parange_to_ipa_size (uint64_t *id_aa64mmfr0) +{ +uint32_t ipa_size = chosen_ipa_bit_size ? + chosen_ipa_bit_size :hvf_arm_get_max_ipa_bit_size (); + +/* Clamp down the PARange to the IPA size the kernel supports. */ +uint8_t index =round_down_to_parange_index (ipa_size); + *id_aa64mmfr0 = (*id_aa64mmfr0 & ~R_ID_AA64MMFR0_PARANGE_MASK ) | index; +} +static bool hvf_arm_get_host_cpu_features (ARMHostCPUFeatures *ahcf) {ARMISARegisters host_isar = {};@@ -882,6 +895,8 @@ static bool hvf_arm_get_host_cpu_features(ARMHostCPUFeatures *ahcf) r |=hv_vcpu_get_sys_reg (fd,HV_SYS_REG_MIDR_EL1 , &ahcf->midr); r |=hv_vcpu_destroy (fd); +clamp_id_aa64mmfr0_parange_to_ipa_size (&host_isar.id_aa64mmfr0); + ahcf->isar = host_isar;/* @@ -904,6 +919,30 @@ static bool hvf_arm_get_host_cpu_features(ARMHostCPUFeatures *ahcf)return r ==HV_SUCCESS ; } +uint32_t hvf_arm_get_default_ipa_bit_size (void ) +{ +uint32_t default_ipa_size; +hv_return_t ret =hv_vm_config_get_default_ipa_size (&default_ipa_size); +assert_hvf_ok (ret); + +return default_ipa_size; +} + +uint32_t hvf_arm_get_max_ipa_bit_size (void ) +{ +uint32_t max_ipa_size; +hv_return_t ret =hv_vm_config_get_max_ipa_size (&max_ipa_size); +assert_hvf_ok (ret); + +/* +* We clamp any IPA size we want to back the VM with to a valid PARange +* value so the guest doesn't try and map memory outside of the valid range. +* This logic just clamps the passed in IPA bit size to the first valid +* PARange value <= to it. +*/ +return round_down_to_parange_bit_size (max_ipa_size); +} +void hvf_arm_set_cpu_features_from_host (ARMCPU *cpu) {if (!arm_host_cpu_features.dtb_compatible) {@@ -931,8 +970,18 @@ void hvf_arch_vcpu_destroy(CPUState *cpu)hv_return_t hvf_arch_vm_create (MachineState *ms,uint32_t pa_range) { +hv_return_t ret;hv_vm_config_t config =hv_vm_config_create (); -hv_return_t ret =hv_vm_create (config); + + ret =hv_vm_config_set_ipa_size (config, pa_range); +if (ret !=HV_SUCCESS ) { + goto cleanup; + } + chosen_ipa_bit_size = pa_range; + + ret =hv_vm_create (config); + +cleanup:os_release (config);return ret;@@ -1004,6 +1053,11 @@ int hvf_arch_init_vcpu(CPUState *cpu) &arm_cpu->isar.id_aa64mmfr0);assert_hvf_ok (ret); +clamp_id_aa64mmfr0_parange_to_ipa_size (&arm_cpu->isar.id_aa64mmfr0); + ret =hv_vcpu_set_sys_reg (cpu->accel->fd,HV_SYS_REG_ID_AA64MMFR0_EL1 , + arm_cpu->isar.id_aa64mmfr0); +assert_hvf_ok (ret); +return 0 ; }target/arm/hvf_arm.h +19
-0@@ -22,4 +22,23 @@ void hvf_arm_init_debug(void);void hvf_arm_set_cpu_features_from_host (ARMCPU *cpu); +#ifdef CONFIG_HVF + +uint32_t hvf_arm_get_default_ipa_bit_size (void ); +uint32_t hvf_arm_get_max_ipa_bit_size (void ); + +#else + +static inline uint32_t hvf_arm_get_default_ipa_bit_size (void ) +{ +return 0 ; +} + +static inline uint32_t hvf_arm_get_max_ipa_bit_size (void ) +{ +return 0 ; +} + +#endif +#endif target/arm/internals.h +19
-0@@ -450,6 +450,25 @@ static inline void update_spsel(CPUARMState *env, uint32_t imm)*/ unsigned int arm_pamax (ARMCPU *cpu); +/* +* round_down_to_parange_index +* @bit_size: uint8_t +* +* Rounds down the bit_size supplied to the first supported ARM physical +* address range and returns the index for this. The index is intended to +* be used to set ID_AA64MMFR0_EL1's PARANGE bits. +*/ +uint8_t round_down_to_parange_index (uint8_t bit_size); + +/* +* round_down_to_parange_bit_size +* @bit_size: uint8_t +* +* Rounds down the bit_size supplied to the first supported ARM physical +* address range bit size and returns this. +*/ +uint8_t round_down_to_parange_bit_size (uint8_t bit_size); +/* Return true if extended addresses are enabled. * This is always the case if our translation regime is 64 bit, * but depends on TTBCR.EAE for 32 bit. target/arm/ptw.c +15
-0@@ -96,6 +96,21 @@ static const uint8_t pamax_map[] = { [6] = 52, }; +uint8_t round_down_to_parange_index (uint8_t bit_size) +{ +for (int i =ARRAY_SIZE (pamax_map) -1 ; i >=0 ; i--) { +if (pamax_map[i] <= bit_size) { +return i; + } + } +g_assert_not_reached (); +} + +uint8_t round_down_to_parange_bit_size (uint8_t bit_size) +{ +return pamax_map[round_down_to_parange_index (bit_size)]; +} +/* * The cpu-specific constant value of PAMax; also used by hw/arm/virt. * Note that machvirt_init calls this on a CPU that is inited but not realized!
Re: [PATCH v2 3/3] hvf: arm: Implement and use hvf_get_physical_address_range
Will do. I’ll reach out if I need extra info. The issue appears to be closed though, was this fixed/no-repro already though?
-Danny
Quoted 303 lines · Philippe Mathieu-Daudé
Hi Danny,
Quoted 20 lines · Danny Canter
This patch's main focus is to use the previously added
hvf_get_physical_address_rangeto inform VM creation about the IPA size we need for the VM, so we can extend the default 36b IPA size and support VMs with 64+GB of RAM. This is done by freezing the memory map, computing the highest GPA and then (depending on if the platform supports an IPA size that large) telling the kernel to use a size >= for the VM. In pursuit of this a couple of things related to how we handle the physical address range we expose to guests were altered, but for an explanation of what we were doing:
Today, to get the IPA size we were reading id_aa64mmfr0_el1's PARange field from a newly made vcpu. Unfortunately, HVF just returns the hosts PARange directly for the initial value and not the IPA size that will actually back the VM, so we believe we have much more address space than we actually do today it seems. Starting in macOS 13.0 some APIs were introduced to be able to query the maximum IPA size the kernel supports, and to set the IPA size for a given VM. However, this still has a couple of issues on < macOS 15. Up until macOS 15 (and if the hardware supportedQuoted 8 lines
it) the max IPA size was 39 bits which is not a valid PARange value, so we can't clamp down what we advertise in the vcpu's
id_aa64mmfr0_el1to our IPA size. Starting in macOS 15 however, the maximum IPA size is 40 bits (if it's supported in the hardware as well) which is also a valid PARange value so we can set our IPA size to the maximum as well as clamp down the PARange we advertise to the guest. This allows VMs with 64+ GB of RAM and should fix the oddness of the PARange situation as well.Could you have a look at the following issue related to your patch?
https://gitlab.com/qemu-project/qemu/-/issues/2800Quoted 264 lines
Signed-off-by Danny Canter <[email protected]>--- accel/hvf/hvf-accel-ops.c | 12 ++++++++- hw/arm/virt.c | 31 +++++++++++++++++++++- target/arm/hvf/hvf.c | 56 ++++++++++++++++++++++++++++++++++++++- target/arm/hvf_arm.h | 19 +++++++++++++ target/arm/internals.h | 19 +++++++++++++ target/arm/ptw.c | 15 +++++++++++ 6 files changed, 149 insertions(+), 3 deletions(-)accel/hvf/hvf-accel-ops.c +11
-0@@ -53,6 +53,7 @@#include "exec/address-spaces.h" #include "exec/exec-all.h" #include "gdbstub/enums.h" +#include "hw/boards.h" #include "sysemu/cpus.h" #include "sysemu/hvf.h" #include "sysemu/hvf_int.h" @@ -319,8 +320,17 @@ static int hvf_accel_init(MachineState *ms)int x;hv_return_t ret;HVFState *s; +int pa_range =36 ; +MachineClass *mc =MACHINE_GET_CLASS (ms); - ret =hvf_arch_vm_create (ms,0 ); +if (mc->hvf_get_physical_address_range) { + pa_range = mc->hvf_get_physical_address_range (ms); +if (pa_range <0 ) { +return -EINVAL ; + } + } + + ret =hvf_arch_vm_create (ms, (uint32_t )pa_range);assert_hvf_ok (ret); s =g_new0 (HVFState,1 );hw/arm/virt.c +30
-1@@ -66,6 +66,7 @@#include "hw/intc/arm_gicv3_its_common.h" #include "hw/irq.h" #include "kvm_arm.h" +#include "hvf_arm.h" #include "hw/firmware/smbios.h" #include "qapi/visitor.h" #include "qapi/qapi-visit-common.h" @@ -3030,7 +3031,35 @@ static int virt_kvm_type(MachineState *ms, const char *type_str)static int virt_hvf_get_physical_address_range (MachineState *ms) { -return 0 ; +VirtMachineState *vms =VIRT_MACHINE (ms); + +int default_ipa_size =hvf_arm_get_default_ipa_bit_size (); +int max_ipa_size =hvf_arm_get_max_ipa_bit_size (); + +/* We freeze the memory map to compute the highest gpa */ +virt_set_memmap (vms, max_ipa_size); + +int requested_ipa_size =64 -clz64 (vms->highest_gpa); + +/* +* If we're <= the default IPA size just use the default. +* If we're above the default but below the maximum, round up to +* the maximum. hvf_arm_get_max_ipa_bit_size() conveniently only +* returns values that are valid ARM PARange values. +*/ +if (requested_ipa_size <= default_ipa_size) { + requested_ipa_size = default_ipa_size; + }else if (requested_ipa_size <= max_ipa_size) { + requested_ipa_size = max_ipa_size; + }else { +error_report ("-m and ,maxmem option values " +"require an IPA range (%d bits) larger than " +"the one supported by the host (%d bits)" , + requested_ipa_size, max_ipa_size); +return -1 ; + } + +return requested_ipa_size; }static void virt_machine_class_init (ObjectClass *oc,void *data)target/arm/hvf/hvf.c +50
-1@@ -22,6 +22,7 @@#include <mach/mach_time.h> #include "exec/address-spaces.h" +#include "hw/boards.h" #include "hw/irq.h" #include "qemu/main-loop.h" #include "sysemu/cpus.h" @@ -297,6 +298,8 @@ void hvf_arm_init_debug(void)static void hvf_wfi (CPUState *cpu); +static uint32_t chosen_ipa_bit_size; + typedef struct HVFVTimer { /* Vtimer value during migration and paused state */ uint64_t vtimer_val;@@ -839,6 +842,16 @@ static uint64_t hvf_get_reg(CPUState *cpu, int rt)return val; } +static void clamp_id_aa64mmfr0_parange_to_ipa_size(uint64_t *id_aa64mmfr0) +{ +uint32_t ipa_size = chosen_ipa_bit_size ? + chosen_ipa_bit_size :hvf_arm_get_max_ipa_bit_size (); + +/* Clamp down the PARange to the IPA size the kernel supports. */ +uint8_t index =round_down_to_parange_index (ipa_size); + *id_aa64mmfr0 = (*id_aa64mmfr0 & ~R_ID_AA64MMFR0_PARANGE_MASK ) | index; +} +static bool hvf_arm_get_host_cpu_features (ARMHostCPUFeatures *ahcf) {ARMISARegisters host_isar = {};@@ -882,6 +895,8 @@ static bool hvf_arm_get_host_cpu_features(ARMHostCPUFeatures *ahcf) r |=hv_vcpu_get_sys_reg (fd,HV_SYS_REG_MIDR_EL1 , &ahcf->midr); r |=hv_vcpu_destroy (fd); +clamp_id_aa64mmfr0_parange_to_ipa_size (&host_isar.id_aa64mmfr0); + ahcf->isar = host_isar;/* @@ -904,6 +919,30 @@ static bool hvf_arm_get_host_cpu_features(ARMHostCPUFeatures *ahcf)return r ==HV_SUCCESS ; } +uint32_t hvf_arm_get_default_ipa_bit_size(void) +{ +uint32_t default_ipa_size; +hv_return_t ret =hv_vm_config_get_default_ipa_size (&default_ipa_size); +assert_hvf_ok (ret); + +return default_ipa_size; +} + +uint32_t hvf_arm_get_max_ipa_bit_size (void ) +{ +uint32_t max_ipa_size; +hv_return_t ret =hv_vm_config_get_max_ipa_size (&max_ipa_size); +assert_hvf_ok (ret); + +/* +* We clamp any IPA size we want to back the VM with to a valid PARange +* value so the guest doesn't try and map memory outside of the valid range. +* This logic just clamps the passed in IPA bit size to the first valid +* PARange value <= to it. +*/ +return round_down_to_parange_bit_size (max_ipa_size); +} +void hvf_arm_set_cpu_features_from_host (ARMCPU *cpu) {if (!arm_host_cpu_features.dtb_compatible) {@@ -931,8 +970,18 @@ void hvf_arch_vcpu_destroy(CPUState *cpu)hv_return_t hvf_arch_vm_create (MachineState *ms,uint32_t pa_range) { +hv_return_t ret;hv_vm_config_t config =hv_vm_config_create (); -hv_return_t ret =hv_vm_create (config); + + ret =hv_vm_config_set_ipa_size (config, pa_range); +if (ret !=HV_SUCCESS ) { + goto cleanup; + } + chosen_ipa_bit_size = pa_range; + + ret =hv_vm_create (config); + +cleanup:os_release (config);return ret;@@ -1004,6 +1053,11 @@ int hvf_arch_init_vcpu(CPUState *cpu) &arm_cpu->isar.id_aa64mmfr0);assert_hvf_ok (ret); +clamp_id_aa64mmfr0_parange_to_ipa_size (&arm_cpu->isar.id_aa64mmfr0); + ret =hv_vcpu_set_sys_reg (cpu->accel->fd,HV_SYS_REG_ID_AA64MMFR0_EL1 , + arm_cpu->isar.id_aa64mmfr0); +assert_hvf_ok (ret); +return 0 ; } diff --git a/target/arm/hvf_arm.h b/target/arm/hvf_arm.hindex
e848c1d27d..26c717b382100644target/arm/hvf_arm.h +18
-0@@ -22,4 +22,23 @@ void hvf_arm_init_debug(void);void hvf_arm_set_cpu_features_from_host (ARMCPU *cpu); +#ifdef CONFIG_HVF + +uint32_t hvf_arm_get_default_ipa_bit_size (void ); +uint32_t hvf_arm_get_max_ipa_bit_size (void ); + +#else + +static inline uint32_t hvf_arm_get_default_ipa_bit_size (void ) +{ +return 0 ; +} + +static inline uint32_t hvf_arm_get_max_ipa_bit_size (void ) +{ +return 0 ; +} + +#endif +#endif target/arm/internals.h +18
-0@@ -450,6 +450,25 @@ static inline void update_spsel(CPUARMState *env, uint32_t imm)*/ unsigned int arm_pamax (ARMCPU *cpu); +/* +* round_down_to_parange_index +* @bit_size: uint8_t +* +* Rounds down the bit_size supplied to the first supported ARM physical +* address range and returns the index for this. The index is intended to +* be used to set ID_AA64MMFR0_EL1's PARANGE bits. +*/ +uint8_t round_down_to_parange_index(uint8_t bit_size); + +/* +* round_down_to_parange_bit_size +* @bit_size: uint8_t +* +* Rounds down the bit_size supplied to the first supported ARM physical +* address range bit size and returns this. +*/ +uint8_t round_down_to_parange_bit_size (uint8_t bit_size); +/* Return true if extended addresses are enabled. * This is always the case if our translation regime is 64 bit, * but depends on TTBCR.EAE for 32 bit. target/arm/ptw.c +14
-0@@ -96,6 +96,21 @@ static const uint8_t pamax_map[] = { [6 ] =52 , }; +uint8_t round_down_to_parange_index(uint8_t bit_size) +{ +for (int i =ARRAY_SIZE (pamax_map) -1 ; i >=0 ; i--) { +if (pamax_map[i] <= bit_size) { +return i; + } + } +g_assert_not_reached (); +} + +uint8_t round_down_to_parange_bit_size (uint8_t bit_size) +{ +return pamax_map[round_down_to_parange_index (bit_size)]; +} +/* * The cpu-specific constant value of PAMax; also used by hw/arm/virt. * Note that machvirt_init calls this on a CPU that is inited but not realized!
Re: [PATCH v2 3/3] hvf: arm: Implement and use hvf_get_physical_address_range
Quoted 2 lines · Danny Canter
Will do. I’ll reach out if I need extra info. The issue appears to be closed though, was this fixed/no-repro already though?
Whoops, no, that must have been a mis-click on my part.
While you're looking at address-space related bugs,
https://gitlab.com/qemu-project/qemu/-/issues/2713
is another recent one -- user reports that QEMU says they're limited to 32 bits even though their mac/macos has a 40-bit IVA space.
thanks
-- PMM
Re: [PATCH v2 3/3] hvf: arm: Implement and use hvf_get_physical_address_range
No worries, will get a machine on 15.2 at least as it seems both reports are on that or higher. I’ll likely have time mid to late week to debug this. Thanks!
-Danny
Quoted 16 lines · Peter Maydell
Quoted 2 lines · Danny Canter
Will do. I’ll reach out if I need extra info. The issue appears to be closed though, was this fixed/no-repro already though?
Whoops, no, that must have been a mis-click on my part.
While you're looking at address-space related bugs,
https://gitlab.com/qemu-project/qemu/-/issues/2713
is another recent one -- user reports that QEMU says they're limited to 32 bits even though their mac/macos has a 40-bit IVA space.thanks
-- PMM
Quoted 22 lines · Danny Canter
This patchsets focus is on lighting up the ability to create VMs with 64+GB of RAM through using some new APIs introduced in macOS 13. Due to the IPA sizes supported in macOS, the first version we can properly support this requirement is macOS 15 as (if the hardware supports it also) the kernel adds support for a 40b IPA size, which is the first supported ARM PARange value after 36b, so we can advertise this to the guest properly as well in
id_aa64mmfr0_el1.Today if you asked for a > 64GB VM you'd be met with a pretty unwieldy
HV_BAD_ARGUMENT. On machines without 40b IPA support this patchset also improves this, and the message mirrors thekvm_typeerror you'd get on ARM:"qemu-system-aarch64: -accel hvf: Addressing limited to 36 bits, but memory exceeds it by 18253611008 bytes"
Changes from V1 to V2 (Thanks Peter for review!):
- Added a new function pointer to MachineClass to be able to freeze the memory map and compute the highest guest physical address. We use this to inform VM creation on what IPA size we should ask the kernel for. This is very similar to what ARM's
kvm_type()does.
Quoted 4 lines
- Fixed redundant loop in
round_down_to_parange_bit_size- Move the splitting up of
hv_vm_createlogic per platform to a separate patch. This is mostly for readability.
I only had one minor comment on patch 1, so I've applied the series to target-arm.next and made that tweak there.
thanks
-- PMM