Re: [PATCH v5 07/22] KVM: arm64: vgic-its: Implement vgic_its_has_attr_regs and attr_regs_access

[Auger Eric <eric.auger@redhat.com>]

Hi Christoffer,

On 27/04/2017 13:00, Christoffer Dall wrote:
> On Fri, Apr 14, 2017 at 12:15:19PM +0200, Eric Auger wrote:
>> This patch implements vgic_its_has_attr_regs and vgic_its_attr_regs_access
>> upon the MMIO framework. VGIC ITS KVM device KVM_DEV_ARM_VGIC_GRP_ITS_REGS
>> group becomes functional.
>>
>> At least GITS_CREADR and GITS_IIDR require to differentiate a guest write
>> action from a user access. As such let's introduce a new uaccess_its_write
>> vgic_register_region callback.
>>
>> Signed-off-by: Eric Auger <eric.auger@redhat.com>
>>
>> ---
>> v4 -> v5:
>> - use GITS_TYPER instead of offset 0x8
>> - uaccess_its_write now can return an error
>>
>> v3 -> v4:
>> - remove changes to the REGISTER_ITS_DESC macro. This will be handled in
>>   subsequent patch with the introduction of a new REGISTER_ITS_DESC_UACCESS
>>   macro
>> - fix IIDR access and add a comment wrt full length access
>> - handle endianness
>> - add kvm lock and vcpus lock
>> ---
>>  virt/kvm/arm/vgic/vgic-its.c  | 79 +++++++++++++++++++++++++++++++++++++++++--
>>  virt/kvm/arm/vgic/vgic-mmio.h |  9 +++--
>>  2 files changed, 84 insertions(+), 4 deletions(-)
>>
>> diff --git a/virt/kvm/arm/vgic/vgic-its.c b/virt/kvm/arm/vgic/vgic-its.c
>> index f687e91..a9a2c12 100644
>> --- a/virt/kvm/arm/vgic/vgic-its.c
>> +++ b/virt/kvm/arm/vgic/vgic-its.c
>> @@ -1469,14 +1469,89 @@ static void vgic_its_destroy(struct kvm_device *kvm_dev)
>>  int vgic_its_has_attr_regs(struct kvm_device *dev,
>>  			   struct kvm_device_attr *attr)
>>  {
>> -	return -ENXIO;
>> +	const struct vgic_register_region *region;
>> +	struct vgic_io_device iodev = {
>> +		.regions = its_registers,
>> +		.nr_regions = ARRAY_SIZE(its_registers),
>> +	};
>> +	gpa_t offset = attr->attr;
>> +
>> +	region = vgic_find_mmio_region(iodev.regions,
>> +				       iodev.nr_regions,
>> +				       offset);
> 
> why do you need to define the iodev here?  Can't you just pass
> its_registers and ARRAY_SIZE(its_registers) directly?
Yes makes sense.
> 
>> +	if (!region)
>> +		return -ENXIO;
>> +
>> +	return 0;
>>  }
>>  
>>  int vgic_its_attr_regs_access(struct kvm_device *dev,
>>  			      struct kvm_device_attr *attr,
>>  			      u64 *reg, bool is_write)
>>  {
>> -	return -ENXIO;
>> +	const struct vgic_register_region *region;
>> +	struct vgic_io_device iodev = {
>> +		.regions = its_registers,
>> +		.nr_regions = ARRAY_SIZE(its_registers),
>> +	};
>> +	struct vgic_its *its = dev->private;
>> +	gpa_t addr, offset = attr->attr;
>> +	unsigned int len;
>> +	unsigned long data = 0;
>> +	int ret = 0;
> 
> nit: you could structure this whole declaration block more nicely by
> having a separate line for the declaration and initialization of offset,
> and by moving the iodev declaration/initialization to the end.
> 
> It might also be cleaner to do any non-zero initialization in a separate
> block following the declarations.

OK
> 
>> +
>> +	/*
>> +	 * Among supported registers, only GITS_CTLR (0x0) and GITS_IIDR (0x4)
>> +	 * are 32 bits. Others are 64 bits.
>> +	 */
>> +	if ((offset < GITS_TYPER && offset & 0x3) ||
>> +	    (offset >= GITS_TYPER && offset & 0x7))
>> +		return -EINVAL;
>> +
>> +	mutex_lock(&dev->kvm->lock);
>> +
>> +	if (IS_VGIC_ADDR_UNDEF(its->vgic_its_base)) {
>> +		ret = -ENXIO;
>> +		goto out;
>> +	}
>> +
>> +	region = vgic_find_mmio_region(iodev.regions,
>> +				       iodev.nr_regions,
>> +				       offset);
>> +	if (!region) {
>> +		ret = -ENXIO;
>> +		goto out;
>> +	}
>> +
>> +	if (!lock_all_vcpus(dev->kvm)) {
>> +		ret = -EBUSY;
>> +		goto out;
>> +	}
>> +
>> +	addr = its->vgic_its_base + offset;
>> +
>> +	/*
>> +	 * Only full length register accesses are supported although
>> +	 * the architecture spec theoretically allows upper/lower 32
> 
> does the spec allow 32-bit accesses, or only theoretically ?  :)
Yes the spec allows 32-bit access.
> 
>> +	 * bits to be accessed independently
>> +	 */
> 
> In any case, the comment is a bit confusing, because it seems to imply
> that we only support 64-bit accesses, but we do set the length below to
> 4 or 8.
> 
> Did you mean:
> 
> 	/* 
> 	 * Althought the spec supports upper/lower 32-bit accesses to
> 	 * 64-bit ITS registers, the userspace ABI requires 64-bit
> 	 * accesses to all 64-bit wide registers.  We therefore only 32-bit
> 	 * accesses to the GITS_CTLR, GITS_IIDR registers.
> 	 */
Yes this is what I meant. I will rephrase as you suggest.
> 
> Also, I don't understand how this works with the ID registers?  For
> example, if userspace wants to read GITS_PIDR1 does it have to read
> GITS_PIDR0 as a 64-bit register and split it afterwards?  (that doesn't
> work with this implementation)
> 
Hum the IDREGS were not described in the archi spec 8.19 (ITS register
descriptions) and I "forgot" them. they are 32 bit and 32 bit aligned
access should be allowed. Thanks for spotting this.
>> +	len = region->access_flags & VGIC_ACCESS_64bit ? 8 : 4;
>> +
>> +	if (is_write) {
>> +		data = vgic_data_mmio_bus_to_host(reg, len);
> 
> I don't think we need this anymore; we no longer share the guest
> trapping MMIO path with uaccesses.
agreed
> 
>> +		if (region->uaccess_its_write)
>> +			ret = region->uaccess_its_write(dev->kvm, its, addr,
>> +							len, data);
>> +		else
>> +			region->its_write(dev->kvm, its, addr, len, data);
>> +	} else {
>> +		data = region->its_read(dev->kvm, its, addr, len);
>> +		vgic_data_host_to_mmio_bus(reg, len, data);
> 
> same here.
OK

Thanks

Eric
> 
>> +	}
>> +	unlock_all_vcpus(dev->kvm);
>> +out:
>> +	mutex_unlock(&dev->kvm->lock);
>> +	return ret;
>>  }
>>  
>>  static int vgic_its_has_attr(struct kvm_device *dev,
>> diff --git a/virt/kvm/arm/vgic/vgic-mmio.h b/virt/kvm/arm/vgic/vgic-mmio.h
>> index 6eec91b..ea4171a 100644
>> --- a/virt/kvm/arm/vgic/vgic-mmio.h
>> +++ b/virt/kvm/arm/vgic/vgic-mmio.h
>> @@ -36,8 +36,13 @@ struct vgic_register_region {
>>  	};
>>  	unsigned long (*uaccess_read)(struct kvm_vcpu *vcpu, gpa_t addr,
>>  				      unsigned int len);
>> -	void (*uaccess_write)(struct kvm_vcpu *vcpu, gpa_t addr,
>> -			      unsigned int len, unsigned long val);
>> +	union {
>> +		void (*uaccess_write)(struct kvm_vcpu *vcpu, gpa_t addr,
>> +				      unsigned int len, unsigned long val);
>> +		int (*uaccess_its_write)(struct kvm *kvm, struct vgic_its *its,
>> +					 gpa_t addr, unsigned int len,
>> +					 unsigned long val);
>> +	};
>>  };
>>  
>>  extern struct kvm_io_device_ops kvm_io_gic_ops;
>> -- 
>> 2.5.5
>>
> 
> Thanks,
> -Christoffer
>