Skip to contentSkip to conversation

KVM: s390: fix diag258 virtual-physical confusion

PATCH In review Nico Boehr 3 participants 7 messages Sep 17, 2024 – Sep 18, 2024

Subsystems (3)
lore
Share & tools
lore

Nico Boehr

author
#0 lore raw

cover KVM: s390: fix diag258 virtual-physical confusion

The parameters for the diag 0x258 are real addresses, not virtual, but KVM was using them as virtual addresses. This only happened to work, since the Linux kernel as a guest used to have a 1:1 mapping for physical vs virtual addresses.

In addition, add handling for guest addresses outside memslots in access_guest_page() for architectural compliance and testability.

Michael Mueller (1):
KVM: s390: Change virtual to physical address access in diag 0x258
handler

Nico Boehr (1):
KVM: s390: gaccess: check if guest address is in memslot

 arch/s390/kvm/diag.c    |  2 +-
 arch/s390/kvm/gaccess.c |  7 +++++++
 arch/s390/kvm/gaccess.h | 14 ++++++++------
 3 files changed, 16 insertions(+), 7 deletions(-)
Signature · 1 line
2.46.0
author
#1 lore raw

PATCH 1/2 KVM: s390: gaccess: check if guest address is in memslot

Previously, access_guest_page() did not check whether the given guest address is inside of a memslot. This is not a problem, since kvm_write_guest_page/kvm_read_guest_page return -EFAULT in this case.

However, -EFAULT is also returned when copy_to/from_user fails.

When emulating a guest instruction, the address being outside a memslot usually means that an addressing exception should be injected into the guest.

Failure in copy_to/from_user however indicates that something is wrong in userspace and hence should be handled there.

To be able to distinguish these two cases, return PGM_ADDRESSING in access_guest_page() when the guest address is outside guest memory. In access_guest_real(), populate vcpu->arch.pgm.code such that kvm_s390_inject_prog_cond() can be used in the caller for injecting into the guest (if applicable).

Since this adds a new return value to access_guest_page(), we need to make sure that other callers are not confused by the new positive return value.

There are the following users of access_guest_page():
- access_guest_with_key() does the checking itself (in guest_range_to_gpas()), so this case should never happen. Even if, the handling is set up properly.
- access_guest_real() just passes the return code to its callers, which are:
- read_guest_real() - see below
- write_guest_real() - see below

There are the following users of read_guest_real():
- ar_translation() in gaccess.c which already returns PGM_*
- setup_apcb10(), setup_apcb00(), setup_apcb11() in vsie.c which always return -EFAULT on read_guest_read() nonzero return - no change
- shadow_crycb(), handle_stfle() always present this as validity, this could be handled better but doesn't change current behaviour - no change

There are the following users of write_guest_real():
- kvm_s390_store_status_unloaded() always returns -EFAULT on write_guest_real() failure.

Fixes 2293897805c2 ("KVM: s390: add architecture compliant guest access functions")
---
 arch/s390/kvm/gaccess.c |  7 +++++++
 arch/s390/kvm/gaccess.h | 14 ++++++++------
 2 files changed, 15 insertions(+), 6 deletions(-)
arch/s390/kvm/gaccess.c +7 -0
index e65f597e3044..004047578729 100644
@@ -828,6 +828,9 @@ static int access_guest_page(struct kvm *kvm, enum gacc_mode mode, gpa_t gpa,
 	const gfn_t gfn = gpa_to_gfn(gpa);
 	int rc;
 
+	if (!gfn_to_memslot(kvm, gfn))
+		return PGM_ADDRESSING;
+
 	if (mode == GACC_STORE)
 		rc = kvm_write_guest_page(kvm, gfn, data, offset, len);
 	else
@@ -985,6 +988,10 @@ int access_guest_real(struct kvm_vcpu *vcpu, unsigned long gra,
 		gra += fragment_len;
 		data += fragment_len;
 	}
+
+	if (rc > 0)
+		vcpu->arch.pgm.code = rc;
+
 	return rc;
 }
 
arch/s390/kvm/gaccess.h +8 -6
index b320d12aa049..3fde45a151f2 100644
@@ -405,11 +405,12 @@ int read_guest_abs(struct kvm_vcpu *vcpu, unsigned long gpa, void *data,
  * @len: number of bytes to copy
  *
  * Copy @len bytes from @data (kernel space) to @gra (guest real address).
- * It is up to the caller to ensure that the entire guest memory range is
- * valid memory before calling this function.
  * Guest low address and key protection are not checked.
  *
- * Returns zero on success or -EFAULT on error.
+ * Returns zero on success, -EFAULT when copying from @data failed, or
+ * PGM_ADRESSING in case @gra is outside a memslot. In this case, pgm check info
+ * is also stored to allow injecting into the guest (if applicable) using
+ * kvm_s390_inject_prog_cond().
  *
  * If an error occurs data may have been copied partially to guest memory.
  */
@@ -428,11 +429,12 @@ int write_guest_real(struct kvm_vcpu *vcpu, unsigned long gra, void *data,
  * @len: number of bytes to copy
  *
  * Copy @len bytes from @gra (guest real address) to @data (kernel space).
- * It is up to the caller to ensure that the entire guest memory range is
- * valid memory before calling this function.
  * Guest key protection is not checked.
  *
- * Returns zero on success or -EFAULT on error.
+ * Returns zero on success, -EFAULT when copying to @data failed, or
+ * PGM_ADRESSING in case @gra is outside a memslot. In this case, pgm check info
+ * is also stored to allow injecting into the guest (if applicable) using
+ * kvm_s390_inject_prog_cond().
  *
  * If an error occurs data may have been copied partially to kernel space.
  */
Signature · 1 line
2.46.0
maintainer Reviewed-by on v1 1/2
#2 lore raw

Re: [PATCH v1 1/2] KVM: s390: gaccess: check if guest address is in memslot

Quoted 3 lines · Nico Boehr

Previously, access_guest_page() did not check whether the given guest address is inside of a memslot. This is not a problem, since kvm_write_guest_page/kvm_read_guest_page return -EFAULT in this case.

...

Quoted 18 lines

To be able to distinguish these two cases, return PGM_ADDRESSING in access_guest_page() when the guest address is outside guest memory. In access_guest_real(), populate vcpu->arch.pgm.code such that kvm_s390_inject_prog_cond() can be used in the caller for injecting into the guest (if applicable).

Since this adds a new return value to access_guest_page(), we need to make sure that other callers are not confused by the new positive return value.

There are the following users of access_guest_page():
- access_guest_with_key() does the checking itself (in guest_range_to_gpas()), so this case should never happen. Even if, the handling is set up properly.
- access_guest_real() just passes the return code to its callers, which are:
- read_guest_real() - see below
- write_guest_real() - see below

Quoted 2 lines

There are the following users of read_guest_real():
- ar_translation() in gaccess.c which already returns PGM_*

With this patch you actually fix a bug in ar_translation(), where two read_guest_real() invocations might have returned -EFAULT instead of the correct PGM_ADDRESSING.

Looks like the author assumed read_guest_real() would do the right thing. See commit 664b49735370 ("KVM: s390: Add access register mode").

Quoted 13 lines
Fixes 2293897805c2 ("KVM: s390: add architecture compliant guest access functions")
---
 arch/s390/kvm/gaccess.c |  7 +++++++
 arch/s390/kvm/gaccess.h | 14 ++++++++------
 2 files changed, 15 insertions(+), 6 deletions(-)
arch/s390/kvm/gaccess.c +0 -0
index e65f597e3044..004047578729 100644

arch/s390/kvm/gaccess.c static int access_guest_page(struct kvm *kvm, enum gacc_mode mode, gpa_t gpa, · line 835

Quoted 8 lines
arch/s390/kvm/gaccess.c +3 -0
@@ -828,6 +828,9 @@ static int access_guest_page(struct kvm *kvm, enum gacc_mode mode, gpa_t gpa,
 	const gfn_t gfn = gpa_to_gfn(gpa);
 	int rc;
 
+	if (!gfn_to_memslot(kvm, gfn))
+		return PGM_ADDRESSING;
+
 	if (mode == GACC_STORE)
 		rc = kvm_write_guest_page(kvm, gfn, data, offset, len);

It would be nice to not add random empty lines to stay consistent with the existing coding style.

arch/s390/kvm/gaccess.c int access_guest_real(struct kvm_vcpu *vcpu, unsigned long gra, · line 996

Quoted 7 lines
arch/s390/kvm/gaccess.c +4 -0
 	}
+
+	if (rc > 0)
+		vcpu->arch.pgm.code = rc;
+
 	return rc;
 }

Same here.

But whoever applies this can change this, or not.

In any case:

#3 lore raw

Re: [PATCH v1 1/2] KVM: s390: gaccess: check if guest address is in memslot

In reply to Nico Boehr

Am 17.09.24 um 17:18 schrieb Nico Boehr:

arch/s390/kvm/gaccess.c int access_guest_real(struct kvm_vcpu *vcpu, unsigned long gra, · line 993

Quoted 7 lines
arch/s390/kvm/gaccess.c +3 -0
@@ -985,6 +988,10 @@ int access_guest_real(struct kvm_vcpu *vcpu, unsigned long gra,
  		gra += fragment_len;
  		data += fragment_len;
  	}
+
+	if (rc > 0)
+		vcpu->arch.pgm.code = rc;

This will work but using trans_exc might be more future proof I guess? Otherwise this looks good with the nits fixed.

author
#4 lore raw

PATCH 2/2 KVM: s390: Change virtual to physical address access in diag 0x258 handler

In reply to Nico Boehr

From: Michael Mueller <[email protected]>

The parameters for the diag 0x258 are real addresses, not virtual, but KVM was using them as virtual addresses. This only happened to work, since the Linux kernel as a guest used to have a 1:1 mapping for physical vs virtual addresses.

Fix KVM so that it correctly uses the addresses as real addresses.

Fixes 8ae04b8f500b ("KVM: s390: Guest's memory access functions get access registers")
[ nrb: drop tested-by tags ]
---
 arch/s390/kvm/diag.c | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)
arch/s390/kvm/diag.c +1 -1
index 2a32438e09ce..74f73141f9b9 100644
@@ -77,7 +77,7 @@ static int __diag_page_ref_service(struct kvm_vcpu *vcpu)
 	vcpu->stat.instruction_diagnose_258++;
 	if (vcpu->run->s.regs.gprs[rx] & 7)
 		return kvm_s390_inject_program_int(vcpu, PGM_SPECIFICATION);
-	rc = read_guest(vcpu, vcpu->run->s.regs.gprs[rx], rx, &parm, sizeof(parm));
+	rc = read_guest_real(vcpu, vcpu->run->s.regs.gprs[rx], &parm, sizeof(parm));
 	if (rc)
 		return kvm_s390_inject_prog_cond(vcpu, rc);
 	if (parm.parm_version != 2 || parm.parm_len < 5 || parm.code != 0x258)
Signature · 1 line
2.46.0
maintainer Reviewed-by on v1 2/2
#5 lore raw

Re: [PATCH v1 2/2] KVM: s390: Change virtual to physical address access in diag 0x258 handler

Am 17.09.24 um 17:18 schrieb Nico Boehr:

Quoted 7 lines

From: Michael Mueller <[email protected]>

The parameters for the diag 0x258 are real addresses, not virtual, but KVM was using them as virtual addresses. This only happened to work, since the Linux kernel as a guest used to have a 1:1 mapping for physical vs virtual addresses.

Quoted 8 lines

Fix KVM so that it correctly uses the addresses as real addresses.

Fixes 8ae04b8f500b ("KVM: s390: Guest's memory access functions get access registers")
[ nrb: drop tested-by tags ]
Quoted 17 lines
---
  arch/s390/kvm/diag.c | 2 +-
  1 file changed, 1 insertion(+), 1 deletion(-)
arch/s390/kvm/diag.c +1 -1
index 2a32438e09ce..74f73141f9b9 100644
@@ -77,7 +77,7 @@ static int __diag_page_ref_service(struct kvm_vcpu *vcpu)
  	vcpu->stat.instruction_diagnose_258++;
  	if (vcpu->run->s.regs.gprs[rx] & 7)
  		return kvm_s390_inject_program_int(vcpu, PGM_SPECIFICATION);
-	rc = read_guest(vcpu, vcpu->run->s.regs.gprs[rx], rx, &parm, sizeof(parm));
+	rc = read_guest_real(vcpu, vcpu->run->s.regs.gprs[rx], &parm, sizeof(parm));
  	if (rc)
  		return kvm_s390_inject_prog_cond(vcpu, rc);
  	if (parm.parm_version != 2 || parm.parm_len < 5 || parm.code != 0x258)
maintainer Reviewed-by on v1 2/2
#6 lore raw

Re: [PATCH v1 2/2] KVM: s390: Change virtual to physical address access in diag 0x258 handler

In reply to Nico Boehr
Quoted 7 lines · Nico Boehr

From: Michael Mueller <[email protected]>

The parameters for the diag 0x258 are real addresses, not virtual, but KVM was using them as virtual addresses. This only happened to work, since the Linux kernel as a guest used to have a 1:1 mapping for physical vs virtual addresses.

Quoted 7 lines

Fix KVM so that it correctly uses the addresses as real addresses.

Fixes 8ae04b8f500b ("KVM: s390: Guest's memory access functions get access registers")
[ nrb: drop tested-by tags ]

^^^^^^^^^^^^^^^^^^^^^^^^^^^^

This shouldn't be part of the commit message.

Quoted 4 lines
---
 arch/s390/kvm/diag.c | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)