Date:   Thu, 01 Dec 2022 18:44:56 +0100
From:   Janis Schoetterl-Glausch <scgl@...ux.ibm.com>
To:     Claudio Imbrenda <imbrenda@...ux.ibm.com>
Cc:     Christian Borntraeger <borntraeger@...ux.ibm.com>,
        Janosch Frank <frankja@...ux.ibm.com>,
        Heiko Carstens <hca@...ux.ibm.com>,
        Vasily Gorbik <gor@...ux.ibm.com>,
        Alexander Gordeev <agordeev@...ux.ibm.com>,
        David Hildenbrand <david@...hat.com>,
        Jonathan Corbet <corbet@....net>, kvm@...r.kernel.org,
        linux-doc@...r.kernel.org, linux-kernel@...r.kernel.org,
        linux-kselftest@...r.kernel.org, linux-s390@...r.kernel.org,
        Paolo Bonzini <pbonzini@...hat.com>,
        Shuah Khan <shuah@...nel.org>,
        Sven Schnelle <svens@...ux.ibm.com>
Subject: Re: [PATCH v3 1/9] KVM: s390: Extend MEM_OP ioctl by storage key
 checked cmpxchg

On Thu, 2022-12-01 at 17:15 +0100, Claudio Imbrenda wrote:
> On Thu, 17 Nov 2022 23:17:50 +0100
> Janis Schoetterl-Glausch <scgl@...ux.ibm.com> wrote:
> 
> > User space can use the MEM_OP ioctl to make storage key checked reads
> > and writes to the guest, however, it has no way of performing atomic,
> > key checked, accesses to the guest.
> > Extend the MEM_OP ioctl in order to allow for this, by adding a cmpxchg
> > mode. For now, support this mode for absolute accesses only.
> > 
> > This mode can be use, for example, to set the device-state-change
> > indicator and the adapter-local-summary indicator atomically.
> > 
> > Signed-off-by: Janis Schoetterl-Glausch <scgl@...ux.ibm.com>
> > ---
> >  include/uapi/linux/kvm.h |   5 ++
> >  arch/s390/kvm/gaccess.h  |   3 ++
> >  arch/s390/kvm/gaccess.c  | 101 +++++++++++++++++++++++++++++++++++++++
> >  arch/s390/kvm/kvm-s390.c |  35 +++++++++++++-
> >  4 files changed, 142 insertions(+), 2 deletions(-)
> > 
> > diff --git a/include/uapi/linux/kvm.h b/include/uapi/linux/kvm.h
> > index 0d5d4419139a..1f36be5493e6 100644
> > --- a/include/uapi/linux/kvm.h
> > +++ b/include/uapi/linux/kvm.h
> > @@ -588,6 +588,8 @@ struct kvm_s390_mem_op {
> >  		struct {
> >  			__u8 ar;	/* the access register number */
> >  			__u8 key;	/* access key, ignored if flag unset */
> > +			__u8 pad1[6];	/* ignored */
> > +			__u64 old_p;	/* ignored if flag unset */
> >  		};
> >  		__u32 sida_offset; /* offset into the sida */
> >  		__u8 reserved[32]; /* ignored */
> > @@ -604,6 +606,9 @@ struct kvm_s390_mem_op {
> >  #define KVM_S390_MEMOP_F_CHECK_ONLY		(1ULL << 0)
> >  #define KVM_S390_MEMOP_F_INJECT_EXCEPTION	(1ULL << 1)
> >  #define KVM_S390_MEMOP_F_SKEY_PROTECTION	(1ULL << 2)
> > +#define KVM_S390_MEMOP_F_CMPXCHG		(1ULL << 3)
> > +/* Non program exception return codes (pgm codes are 16 bit) */
> > +#define KVM_S390_MEMOP_R_NO_XCHG		((1 << 16) + 0)
> 
> are you planning to have further *_R_* macros in the near future?
> if not, remove the + 0

No, we can indeed just add it back if there ever are additional ones.

> if yes, move the (1 << 16) to a macro, so it becomes something like
> (KVM_S390_MEMOP_R_BASE + 0)
> 
> (maybe you can find a better/shorter name)
> 
> >  
> >  /* for KVM_INTERRUPT */
> >  struct kvm_interrupt {
> > diff --git a/arch/s390/kvm/gaccess.h b/arch/s390/kvm/gaccess.h
> > index 9408d6cc8e2c..92a3b9fb31ec 100644
> > --- a/arch/s390/kvm/gaccess.h
> > +++ b/arch/s390/kvm/gaccess.h
> > @@ -206,6 +206,9 @@ int access_guest_with_key(struct kvm_vcpu *vcpu, unsigned long ga, u8 ar,
> >  int access_guest_real(struct kvm_vcpu *vcpu, unsigned long gra,
> >  		      void *data, unsigned long len, enum gacc_mode mode);
> >  
> > +int cmpxchg_guest_abs_with_key(struct kvm *kvm, gpa_t gpa, int len,
> > +			       __uint128_t *old, __uint128_t new, u8 access_key);
> > +
> >  /**
> >   * write_guest_with_key - copy data from kernel space to guest space
> >   * @vcpu: virtual cpu
> > diff --git a/arch/s390/kvm/gaccess.c b/arch/s390/kvm/gaccess.c
> > index 0243b6e38d36..be042865d8a1 100644
> > --- a/arch/s390/kvm/gaccess.c
> > +++ b/arch/s390/kvm/gaccess.c
> > @@ -1161,6 +1161,107 @@ int access_guest_real(struct kvm_vcpu *vcpu, unsigned long gra,
> >  	return rc;
> >  }
> >  
> > +/**
> > + * cmpxchg_guest_abs_with_key() - Perform cmpxchg on guest absolute address.
> > + * @kvm: Virtual machine instance.
> > + * @gpa: Absolute guest address of the location to be changed.
> > + * @len: Operand length of the cmpxchg, required: 1 <= len <= 16. Providing a
> > + *       non power of two will result in failure.
> > + * @old_p: Pointer to old value. If the location at @gpa contains this value, the
> > + *         exchange will succeed. After calling cmpxchg_guest_abs_with_key() *@old
> > + *         contains the value at @gpa before the attempt to exchange the value.
> > + * @new: The value to place at @gpa.
> > + * @access_key: The access key to use for the guest access.
> > + *
> > + * Atomically exchange the value at @gpa by @new, if it contains *@....
> > + * Honors storage keys.
> > + *
> > + * Return: * 0: successful exchange
> > + *         * 1: exchange unsuccessful
> > + *         * a program interruption code indicating the reason cmpxchg could
> > + *           not be attempted
> > + *         * -EINVAL: address misaligned or len not power of two
> > + *         * -EAGAIN: transient failure (len 1 or 2)
> 
> please also document -EOPNOTSUPP

I'd add "* -EOPNOTSUPP: should never occur", then, that ok with you?
> 
> > + */
> > +int cmpxchg_guest_abs_with_key(struct kvm *kvm, gpa_t gpa, int len,
> > +			       __uint128_t *old_p, __uint128_t new,
> > +			       u8 access_key)
> > +{
> > +	gfn_t gfn = gpa >> PAGE_SHIFT;
> > +	struct kvm_memory_slot *slot = gfn_to_memslot(kvm, gfn);
> 
> exchange the above two lines (reverse christmas tree)

Is this a hard requirement? Since there is a dependency.
If I do the initialization further down, the order wouldn't actually change.
> 
> > +	bool writable;
> > +	hva_t hva;
> > +	int ret;
> > +
> > +	if (!IS_ALIGNED(gpa, len))
> > +		return -EINVAL;
> > +
> > +	hva = gfn_to_hva_memslot_prot(slot, gfn, &writable);
> > +	if (kvm_is_error_hva(hva))
> > +		return PGM_ADDRESSING;
> > +	/*
> > +	 * Check if it's a read-only memslot, even though that cannot occur
> > +	 * since those are unsupported.
> > +	 * Don't try to actually handle that case.
> > +	 */
> > +	if (!writable)
> > +		return -EOPNOTSUPP;
> 
> either you document this, or you return something else (like -EINVAL)
> 
> > +
> > +	hva += offset_in_page(gpa);
> > +	switch (len) {
> > +	case 1: {
> > +		u8 old;
> > +
> > +		ret = cmpxchg_user_key((u8 *)hva, &old, *old_p, new, access_key);
> > +		ret = ret < 0 ? ret : old != *old_p;
> > +		*old_p = old;
> > +		break;
> > +	}
> > +	case 2: {
> > +		u16 old;
> > +
> > +		ret = cmpxchg_user_key((u16 *)hva, &old, *old_p, new, access_key);
> > +		ret = ret < 0 ? ret : old != *old_p;
> > +		*old_p = old;
> > +		break;
> > +	}
> > +	case 4: {
> > +		u32 old;
> > +
> > +		ret = cmpxchg_user_key((u32 *)hva, &old, *old_p, new, access_key);
> > +		ret = ret < 0 ? ret : old != *old_p;
> > +		*old_p = old;
> > +		break;
> > +	}
> > +	case 8: {
> > +		u64 old;
> > +
> > +		ret = cmpxchg_user_key((u64 *)hva, &old, *old_p, new, access_key);
> > +		ret = ret < 0 ? ret : old != *old_p;
> > +		*old_p = old;
> > +		break;
> > +	}
> > +	case 16: {
> > +		__uint128_t old;
> > +
> > +		ret = cmpxchg_user_key((__uint128_t *)hva, &old, *old_p, new, access_key);
> > +		ret = ret < 0 ? ret : old != *old_p;
> > +		*old_p = old;
> > +		break;
> 
> I really dislike repeating the same code 5 times, but I guess there was
> no other way?

I could use the function called by cmpxchg_user_key directly, but Heiko won't agree to that.
A macro would work too, of course, not sure if I prefer that tho.
> 
> > +	}
> > +	default:
> > +		return -EINVAL;
> > +	}
> > +	mark_page_dirty_in_slot(kvm, slot, gfn);
> > +	/*
> > +	 * Assume that the fault is caused by protection, either key protection
> > +	 * or user page write protection.
> > +	 */
> > +	if (ret == -EFAULT)
> > +		ret = PGM_PROTECTION;
> > +	return ret;
> > +}
> > +
> >  /**
> >   * guest_translate_address_with_key - translate guest logical into guest absolute address
> >   * @vcpu: virtual cpu
> > diff --git a/arch/s390/kvm/kvm-s390.c b/arch/s390/kvm/kvm-s390.c
> > index 45d4b8182b07..2410b4044283 100644
> > --- a/arch/s390/kvm/kvm-s390.c
> > +++ b/arch/s390/kvm/kvm-s390.c
> > @@ -576,7 +576,6 @@ int kvm_vm_ioctl_check_extension(struct kvm *kvm, long ext)
> >  	case KVM_CAP_S390_VCPU_RESETS:
> >  	case KVM_CAP_SET_GUEST_DEBUG:
> >  	case KVM_CAP_S390_DIAG318:
> > -	case KVM_CAP_S390_MEM_OP_EXTENSION:
> >  		r = 1;
> >  		break;
> >  	case KVM_CAP_SET_GUEST_DEBUG2:
> > @@ -590,6 +589,14 @@ int kvm_vm_ioctl_check_extension(struct kvm *kvm, long ext)
> >  	case KVM_CAP_S390_MEM_OP:
> >  		r = MEM_OP_MAX_SIZE;
> >  		break;
> > +	case KVM_CAP_S390_MEM_OP_EXTENSION:
> > +		/*
> > +		 * Flag bits indicating which extensions are supported.
> > +		 * The first extension doesn't use a flag, but pretend it does,
> > +		 * this way that can be changed in the future.
> > +		 */
> > +		r = 0x3;
> > +		break;
> >  	case KVM_CAP_NR_VCPUS:
> >  	case KVM_CAP_MAX_VCPUS:
> >  	case KVM_CAP_MAX_VCPU_ID:
> > @@ -2714,12 +2721,19 @@ static bool access_key_invalid(u8 access_key)
> >  static int kvm_s390_vm_mem_op(struct kvm *kvm, struct kvm_s390_mem_op *mop)
> >  {
> >  	void __user *uaddr = (void __user *)mop->buf;
> > +	void __user *old_p = (void __user *)mop->old_p;
> > +	union {
> > +		__uint128_t quad;
> > +		char raw[sizeof(__uint128_t)];
> > +	} old = { .quad = 0}, new = { .quad = 0 };
> > +	unsigned int off_in_quad = sizeof(__uint128_t) - mop->size;
> >  	u64 supported_flags;
> >  	void *tmpbuf = NULL;
> >  	int r, srcu_idx;
> >  
> >  	supported_flags = KVM_S390_MEMOP_F_SKEY_PROTECTION
> > -			  | KVM_S390_MEMOP_F_CHECK_ONLY;
> > +			  | KVM_S390_MEMOP_F_CHECK_ONLY
> > +			  | KVM_S390_MEMOP_F_CMPXCHG;
> >  	if (mop->flags & ~supported_flags || !mop->size)
> >  		return -EINVAL;
> >  	if (mop->size > MEM_OP_MAX_SIZE)
> > @@ -2741,6 +2755,15 @@ static int kvm_s390_vm_mem_op(struct kvm *kvm, struct kvm_s390_mem_op *mop)
> >  	} else {
> >  		mop->key = 0;
> >  	}
> > +	if (mop->flags & KVM_S390_MEMOP_F_CMPXCHG) {
> 
> add a quick comment here to explain that this check validates
> off_in_quad, and that it does not do a full validation of mop->size,
> which will happen in cmpxchg_guest_abs_with_key.

Will do.
> 
> > +		if (mop->size > sizeof(new))
> > +			return -EINVAL;
> > +		/* off_in_quad has been validated */
> > +		if (copy_from_user(&new.raw[off_in_quad], uaddr, mop->size))
> > +			return -EFAULT;
> > +		if (copy_from_user(&old.raw[off_in_quad], old_p, mop->size))
> > +			return -EFAULT;
> > +	}
> >  	if (!(mop->flags & KVM_S390_MEMOP_F_CHECK_ONLY)) {
> >  		tmpbuf = vmalloc(mop->size);
> >  		if (!tmpbuf)
> > @@ -2771,6 +2794,14 @@ static int kvm_s390_vm_mem_op(struct kvm *kvm, struct kvm_s390_mem_op *mop)
> >  	case KVM_S390_MEMOP_ABSOLUTE_WRITE: {
> >  		if (mop->flags & KVM_S390_MEMOP_F_CHECK_ONLY) {
> >  			r = check_gpa_range(kvm, mop->gaddr, mop->size, GACC_STORE, mop->key);
> > +		} else if (mop->flags & KVM_S390_MEMOP_F_CMPXCHG) {
> > +			r = cmpxchg_guest_abs_with_key(kvm, mop->gaddr, mop->size,
> > +						       &old.quad, new.quad, mop->key);
> > +			if (r == 1) {
> > +				r = KVM_S390_MEMOP_R_NO_XCHG;
> 
> I wonder if you could not simplify things by returning directly
> KVM_S390_MEMOP_R_NO_XCHG instead of 1

To me it feels like KVM_S390_MEMOP_R_NO_XCHG is api surface and should be referenced here.
cmpxchg_guest_abs_with_key isn't mem op specific
(of course that's the only thing it is currently used for).
> 
> > +				if (copy_to_user(old_p, &old.raw[off_in_quad], mop->size))
> > +					r = -EFAULT;
> > +			}
> >  		} else {
> >  			if (copy_from_user(tmpbuf, uaddr, mop->size)) {
> >  				r = -EFAULT;
> 

Powered by blists - more mailing lists