Re: [PATCH kvmtool v7 2/3] aarch64: Add stolen time support

[Marc Zyngier <maz@kernel.org>]

Hi Sebastian,

On Wed, 02 Mar 2022 14:07:35 +0000,
Sebastian Ene <sebastianene@google.com> wrote:
> 
> This patch adds support for stolen time by sharing a memory region
> with the guest which will be used by the hypervisor to store the stolen
> time information. Reserve a 64kb MMIO memory region after the RTC peripheral
> to be used by pvtime. The exact format of the structure stored by the
> hypervisor is described in the ARM DEN0057A document.
> 
> Signed-off-by: Sebastian Ene <sebastianene@google.com>
> ---
>  Makefile                               |   1 +
>  arm/aarch64/arm-cpu.c                  |   2 +-
>  arm/aarch64/include/kvm/kvm-cpu-arch.h |   1 +
>  arm/aarch64/pvtime.c                   | 103 +++++++++++++++++++++++++
>  arm/include/arm-common/kvm-arch.h      |   6 +-
>  include/kvm/kvm-config.h               |   1 +
>  6 files changed, 112 insertions(+), 2 deletions(-)
>  create mode 100644 arm/aarch64/pvtime.c
> 
> diff --git a/Makefile b/Makefile
> index f251147..e9121dc 100644
> --- a/Makefile
> +++ b/Makefile
> @@ -182,6 +182,7 @@ ifeq ($(ARCH), arm64)
>  	OBJS		+= arm/aarch64/arm-cpu.o
>  	OBJS		+= arm/aarch64/kvm-cpu.o
>  	OBJS		+= arm/aarch64/kvm.o
> +	OBJS		+= arm/aarch64/pvtime.o
>  	ARCH_INCLUDE	:= $(HDRS_ARM_COMMON)
>  	ARCH_INCLUDE	+= -Iarm/aarch64/include
>  
> diff --git a/arm/aarch64/arm-cpu.c b/arm/aarch64/arm-cpu.c
> index d7572b7..7e4a3c1 100644
> --- a/arm/aarch64/arm-cpu.c
> +++ b/arm/aarch64/arm-cpu.c
> @@ -22,7 +22,7 @@ static void generate_fdt_nodes(void *fdt, struct kvm *kvm)
>  static int arm_cpu__vcpu_init(struct kvm_cpu *vcpu)
>  {
>  	vcpu->generate_fdt_nodes = generate_fdt_nodes;
> -	return 0;
> +	return kvm_cpu__setup_pvtime(vcpu);
>  }
>  
>  static struct kvm_arm_target target_generic_v8 = {
> diff --git a/arm/aarch64/include/kvm/kvm-cpu-arch.h b/arm/aarch64/include/kvm/kvm-cpu-arch.h
> index 8dfb82e..2b2c1ff 100644
> --- a/arm/aarch64/include/kvm/kvm-cpu-arch.h
> +++ b/arm/aarch64/include/kvm/kvm-cpu-arch.h
> @@ -19,5 +19,6 @@
>  
>  void kvm_cpu__select_features(struct kvm *kvm, struct kvm_vcpu_init *init);
>  int kvm_cpu__configure_features(struct kvm_cpu *vcpu);
> +int kvm_cpu__setup_pvtime(struct kvm_cpu *vcpu);
>  
>  #endif /* KVM__KVM_CPU_ARCH_H */
> diff --git a/arm/aarch64/pvtime.c b/arm/aarch64/pvtime.c
> new file mode 100644
> index 0000000..fdde683
> --- /dev/null
> +++ b/arm/aarch64/pvtime.c
> @@ -0,0 +1,103 @@
> +#include "kvm/kvm.h"
> +#include "kvm/kvm-cpu.h"
> +#include "kvm/util.h"
> +
> +#include <linux/byteorder.h>
> +#include <linux/types.h>
> +
> +#define ARM_PVTIME_STRUCT_SIZE		(64)
> +
> +struct pvtime_data_priv {
> +	bool	is_supported;
> +	char	*usr_mem;

Consider using void * for pointers that do not have any particular
semantics associated to them.

> +};
> +
> +static struct pvtime_data_priv pvtime_data = {
> +	.is_supported	= true,
> +	.usr_mem	= NULL
> +};
> +
> +static int pvtime__alloc_region(struct kvm *kvm)
> +{
> +	char *mem;
> +	int ret = 0;
> +
> +	mem = mmap(NULL, ARM_PVTIME_MMIO_SIZE, PROT_RW,

I sort of object to the 'MMIO' part of the name. The spec is quite
clear that this should be normal memory. That's purely cosmetic, but
still a bit confusing.

> +		   MAP_ANON_NORESERVE, -1, 0);
> +	if (mem == MAP_FAILED)
> +		return -errno;
> +
> +	ret = kvm__register_dev_mem(kvm, ARM_PVTIME_MMIO_BASE,
> +				    ARM_PVTIME_MMIO_SIZE, mem);

This, on the other side, is wrong. Since the pvtime pages are memory,
mapping them with device attributes will do the wrong thing (the
hypervisor will write to a cacheable mapping, and the guest will
bypass the cache due to the S2 override that you provide here).

kvm__register_ram() is more likely to lead to the behaviour you'd
expect.

> +	if (ret) {
> +		munmap(mem, ARM_PVTIME_MMIO_SIZE);
> +		return ret;
> +	}
> +
> +	pvtime_data.usr_mem = mem;
> +	return ret;
> +}
> +
> +static int pvtime__teardown_region(struct kvm *kvm)
> +{
> +	if (pvtime_data.usr_mem == NULL)
> +		return 0;
> +
> +	kvm__destroy_mem(kvm, ARM_PVTIME_MMIO_BASE,
> +			 ARM_PVTIME_MMIO_SIZE, pvtime_data.usr_mem);
> +	munmap(pvtime_data.usr_mem, ARM_PVTIME_MMIO_SIZE);
> +	pvtime_data.usr_mem = NULL;
> +	return 0;
> +}
> +
> +dev_exit(pvtime__teardown_region);
> +
> +int kvm_cpu__setup_pvtime(struct kvm_cpu *vcpu)
> +{
> +	int ret;
> +	bool has_stolen_time;
> +	u64 pvtime_guest_addr = ARM_PVTIME_MMIO_BASE + vcpu->cpu_id *
> +		ARM_PVTIME_STRUCT_SIZE;
> +	struct kvm_config *kvm_cfg = NULL;
> +	struct kvm_device_attr pvtime_attr = (struct kvm_device_attr) {
> +		.group	= KVM_ARM_VCPU_PVTIME_CTRL,
> +		.addr	= KVM_ARM_VCPU_PVTIME_IPA
> +	};
> +
> +	kvm_cfg = &vcpu->kvm->cfg;
> +	if (kvm_cfg->no_pvtime)
> +		return 0;
> +
> +	if (!pvtime_data.is_supported)
> +		return -ENOTSUP;

It is a bit odd to have this hard failure if running on a system that
doesn't have pvtime. It forces the user to alter their command-line,
which is a bit annoying. I'd rather have a soft-fail here.

> +
> +	has_stolen_time = kvm__supports_extension(vcpu->kvm,
> +						  KVM_CAP_STEAL_TIME);
> +	if (!has_stolen_time)
> +		return 0;
> +
> +	ret = ioctl(vcpu->vcpu_fd, KVM_HAS_DEVICE_ATTR, &pvtime_attr);
> +	if (ret) {
> +		perror("KVM_HAS_DEVICE_ATTR failed\n");
> +		goto out_err;
> +	}
> +
> +	if (!pvtime_data.usr_mem) {
> +		ret = pvtime__alloc_region(vcpu->kvm);
> +		if (ret) {
> +			perror("Failed allocating pvtime region\n");
> +			goto out_err;
> +		}
> +	}
> +
> +	pvtime_attr.addr = (u64)&pvtime_guest_addr;
> +	ret = ioctl(vcpu->vcpu_fd, KVM_SET_DEVICE_ATTR, &pvtime_attr);
> +	if (!ret)
> +		return 0;
> +
> +	perror("KVM_SET_DEVICE_ATTR failed\n");
> +	pvtime__teardown_region(vcpu->kvm);
> +out_err:
> +	pvtime_data.is_supported = false;
> +	return ret;
> +}
> diff --git a/arm/include/arm-common/kvm-arch.h b/arm/include/arm-common/kvm-arch.h
> index c645ac0..3f82663 100644
> --- a/arm/include/arm-common/kvm-arch.h
> +++ b/arm/include/arm-common/kvm-arch.h
> @@ -15,7 +15,8 @@
>   * |  PCI  |////| plat  |       |        |     |         |
>   * |  I/O  |////| MMIO: | Flash | virtio | GIC |   PCI   |  DRAM
>   * | space |////| UART, |       |  MMIO  |     |  (AXI)  |
> - * |       |////| RTC   |       |        |     |         |
> + * |       |////| RTC,  |       |        |     |         |
> + * |       |////| PVTIME|       |        |     |         |
>   * +-------+----+-------+-------+--------+-----+---------+---......
>   */
>  
> @@ -34,6 +35,9 @@
>  #define ARM_RTC_MMIO_BASE	(ARM_UART_MMIO_BASE + ARM_UART_MMIO_SIZE)
>  #define ARM_RTC_MMIO_SIZE	0x10000
>  
> +#define ARM_PVTIME_MMIO_BASE	(ARM_RTC_MMIO_BASE + ARM_RTC_MMIO_SIZE)
> +#define ARM_PVTIME_MMIO_SIZE	SZ_64K
> +
>  #define KVM_FLASH_MMIO_BASE	(ARM_MMIO_AREA + 0x1000000)
>  #define KVM_FLASH_MAX_SIZE	0x1000000
>  
> diff --git a/include/kvm/kvm-config.h b/include/kvm/kvm-config.h
> index 6a5720c..48adf27 100644
> --- a/include/kvm/kvm-config.h
> +++ b/include/kvm/kvm-config.h
> @@ -62,6 +62,7 @@ struct kvm_config {
>  	bool no_dhcp;
>  	bool ioport_debug;
>  	bool mmio_debug;
> +	bool no_pvtime;
>  };
>  
>  #endif

Thanks,

	M.

-- 
Without deviation from the norm, progress is not possible.
_______________________________________________
kvmarm mailing list
kvmarm@lists.cs.columbia.edu
https://lists.cs.columbia.edu/mailman/listinfo/kvmarm