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@linux.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@linux.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 *@old.
• • 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;