[RFC PATCH v2 1/3] KVM: arm64: SIGBUS VMM for SEA guest abort

[RFC PATCH v2 1/3] KVM: arm64: SIGBUS VMM for SEA guest abort

[Jiaqi Yan]

When APEI is unable claim or handles synchronous external abort (SEA)
today KVM handles SEA for guest by injecting an async SError into the
guest directly, bypassing VMM, usually results in guest kernel panic.

One major situation of guest SEA is when vCPU consumes uncorrectable
memory error on the physical memory. Although SError and guest kernel
panic effectively stops the propagation of corrupted memory, it is not
easy for VMM and guest to recover from memory error in a more graceful
manner.

This patch teach KVM to send a SIGBUS BUS_OBJERR to VMM/vCPU, just like
how core kernel signals SIGBUS BUS_OBJERR to a gernal poison consuming
userspace thread when APEI is unable to claim the SEA. In addition to
the benifit that KVM's handling for SEA becomes aligned with core
kernel's behavior
- VMM can inject SEA to guest. Compared to SError, the blast radius in
  VM is possible to be limited to only the consuming thread in guest,
  instead of the entire guest kernel (unless the poison consumption is
  from guest kernel).
- VMM usually tracks the poisoned guest pages. Together with [1], if
  guest consumes again the already poisoned guest pages, VMM can protect
  itself and the host by stopping the consumption at software level, by
  intercepting guest's access to poisoned pages, and again injecting
  SEA to guest.

KVM now handles SEA as follows:
1. Delegate to APEI/GHES driver to see if SEA can be claimed by them.
2. If APEI failed to claim the SEA, send current thread (i.e. VMM in EL0)
   a si_code=BUS_OBJERR SIGBUS signal. If the DIMM error's physical
   address is available from FAR_EL2, si_addr will be the DIMM error's
   host virtual address in VMM/vCPU's memory space.
3. Otherwise bypass VMM and inject async SError to guest.

Tested on a machine running Siryn AmpereOne processor. A dummy application
in VM allocated some memory buffer. The test used EINJ to inject an
uncorrectable recoverable memory error at a page in the allocated memory
buffer. The dummy application then consumed the memory error. Some hack
was done to make core kernel's memory_failure triggered by poison
generation to fail, so KVM had to deal with the SEA guest abort due to
poison consumption. vCPU thread in VMM received SIGBUS BUS_OBJERR with
valid host virtual address of the poisoned page. VMM then injected a SEA
into guest using KVM_SET_VCPU_EVENTS with ext_dabt_pending=1. At last
the dummy application in guest was killed by SIGBUS BUS_OBJERR, while the
guest survived and continued to run.

[1] https://lpc.events/event/18/contributions/1757/attachments/1442/3073/LPC_%20KVM%20Userfault.pdf

Changelog

RFC v2 -> RFC v1
- reword commit msg
- drop unused parameters from kvm_delegate_guest_sea
- remove KVM_CAP_ARM_SIGBUS_ON_SEA and its opt in code
- set FnV bit in vcpu's ESR_ELx if host ESR_EL2's FnV is set
- add documentation for this new SIGBUS feature

Signed-off-by: Jiaqi Yan <jiaqiyan@google.com>
---
 arch/arm64/include/asm/kvm_ras.h | 24 ++++++----
 arch/arm64/kvm/Makefile          |  2 +-
 arch/arm64/kvm/kvm_ras.c         | 81 ++++++++++++++++++++++++++++++++
 arch/arm64/kvm/mmu.c             |  8 +---
 4 files changed, 98 insertions(+), 17 deletions(-)
 create mode 100644 arch/arm64/kvm/kvm_ras.c

diff --git a/arch/arm64/include/asm/kvm_ras.h b/arch/arm64/include/asm/kvm_ras.h
index 87e10d9a635b5..5b4bec6f4f32b 100644
--- a/arch/arm64/include/asm/kvm_ras.h
+++ b/arch/arm64/include/asm/kvm_ras.h
@@ -1,5 +1,4 @@
 /* SPDX-License-Identifier: GPL-2.0 */
-/* Copyright (C) 2018 - Arm Ltd */
 
 #ifndef __ARM64_KVM_RAS_H__
 #define __ARM64_KVM_RAS_H__
@@ -11,15 +10,22 @@
 #include <asm/acpi.h>
 
 /*
- * Was this synchronous external abort a RAS notification?
- * Returns '0' for errors handled by some RAS subsystem, or -ENOENT.
+ * For synchrnous external abort taken to KVM at EL2, not on translation
+ * table walk or hardware update of translation table, is FAR_EL2 valid?
  */
-static inline int kvm_handle_guest_sea(phys_addr_t addr, u64 esr)
-{
-	/* apei_claim_sea(NULL) expects to mask interrupts itself */
-	lockdep_assert_irqs_enabled();
+bool kvm_vcpu_sea_far_valid(const struct kvm_vcpu *vcpu);
 
-	return apei_claim_sea(NULL);
-}
+/*
+ * Handle synchronous external abort (SEA) in the following order:
+ * 1. Delegate to APEI/GHES to see if they can claim SEA. If so, all done.
+ * 2. If the SEA is NOT about S2 translation table, send SIGBUS to current
+ *    with BUS_OBJERR and si_addr set to faulting/poisoned host virtual
+ *    address. When accurate HVA is unavailable, si_addr will be 0.
+ * 3. Otherwise, directly inject an async SError to guest.
+ *
+ * Note this applies to both instruction and data abort (ESR_ELx_EC_IABT_*
+ * and ESR_ELx_EC_DABT_*).
+ */
+void kvm_handle_guest_sea(struct kvm_vcpu *vcpu);
 
 #endif /* __ARM64_KVM_RAS_H__ */
diff --git a/arch/arm64/kvm/Makefile b/arch/arm64/kvm/Makefile
index 3cf7adb2b5038..c4a3a6d4870e6 100644
--- a/arch/arm64/kvm/Makefile
+++ b/arch/arm64/kvm/Makefile
@@ -23,7 +23,7 @@ kvm-y += arm.o mmu.o mmio.o psci.o hypercalls.o pvtime.o \
 	 vgic/vgic-v3.o vgic/vgic-v4.o \
 	 vgic/vgic-mmio.o vgic/vgic-mmio-v2.o \
 	 vgic/vgic-mmio-v3.o vgic/vgic-kvm-device.o \
-	 vgic/vgic-its.o vgic/vgic-debug.o
+	 vgic/vgic-its.o vgic/vgic-debug.o kvm_ras.o
 
 kvm-$(CONFIG_HW_PERF_EVENTS)  += pmu-emul.o pmu.o
 kvm-$(CONFIG_ARM64_PTR_AUTH)  += pauth.o
diff --git a/arch/arm64/kvm/kvm_ras.c b/arch/arm64/kvm/kvm_ras.c
new file mode 100644
index 0000000000000..88d5c57f14bc7
--- /dev/null
+++ b/arch/arm64/kvm/kvm_ras.c
@@ -0,0 +1,81 @@
+// SPDX-License-Identifier: GPL-2.0-only
+
+#include <linux/bitops.h>
+#include <linux/kvm_host.h>
+
+#include <asm/kvm_emulate.h>
+#include <asm/kvm_ras.h>
+#include <asm/system_misc.h>
+
+bool kvm_vcpu_sea_far_valid(const struct kvm_vcpu *vcpu)
+{
+	/*
+	 * FnV is valid only for Data/Instruction aborts and if DFSC/IFSC
+	 * is ESR_ELx_FSC_EXTABT(0b010000).
+	 */
+	if (kvm_vcpu_trap_get_fault(vcpu) == ESR_ELx_FSC_EXTABT)
+		return !(vcpu->arch.fault.esr_el2 & ESR_ELx_FnV);
+
+	/* Other exception classes or aborts don't care about FnV field. */
+	return true;
+}
+
+/*
+ * Was this synchronous external abort a RAS notification?
+ * Returns '0' for errors handled by some RAS subsystem, or -ENOENT.
+ */
+static int kvm_delegate_guest_sea(void)
+{
+	/* apei_claim_sea(NULL) expects to mask interrupts itself */
+	lockdep_assert_irqs_enabled();
+	return apei_claim_sea(NULL);
+}
+
+void kvm_handle_guest_sea(struct kvm_vcpu *vcpu)
+{
+	bool sigbus_on_sea;
+	int idx;
+	u64 vcpu_esr = kvm_vcpu_get_esr(vcpu);
+	u8 fsc = kvm_vcpu_trap_get_fault(vcpu);
+	phys_addr_t fault_ipa = kvm_vcpu_get_fault_ipa(vcpu);
+	gfn_t gfn = fault_ipa >> PAGE_SHIFT;
+	/* When FnV is set, send 0 as si_addr like what do_sea() does. */
+	unsigned long hva = 0UL;
+
+	/*
+	 * For RAS the host kernel may handle this abort.
+	 * There is no need to SIGBUS VMM, or pass the error into the guest.
+	 */
+	if (kvm_delegate_guest_sea() == 0)
+		return;
+
+	/*
+	 * In addition to userspace opt-in, SIGBUS only makes sense if the
+	 * abort is NOT about stage 2 translation table walk and NOT about
+	 * hardware update of stage 2 translation table.
+	 */
+	sigbus_on_sea = (fsc == ESR_ELx_FSC_EXTABT ||
+			 fsc == ESR_ELx_FSC_SECC ||
+			 fsc == ESR_ELx_FSC_SEA_TTW(1) ||
+			 fsc == ESR_ELx_FSC_SECC_TTW(1));
+
+	/* Pass the error directly into the guest. */
+	if (!sigbus_on_sea) {
+		kvm_inject_vabt(vcpu);
+		return;
+	}
+
+	if (kvm_vcpu_sea_far_valid(vcpu)) {
+		idx = srcu_read_lock(&vcpu->kvm->srcu);
+		hva = gfn_to_hva(vcpu->kvm, gfn);
+		srcu_read_unlock(&vcpu->kvm->srcu, idx);
+	}
+
+	/*
+	 * Send a SIGBUS BUS_OBJERR to vCPU thread (the userspace thread that
+	 * runs KVM_RUN) or VMM, which aligns with what host kernel do_sea()
+	 * does if apei_claim_sea() fails.
+	 */
+	arm64_notify_die("synchronous external abort",
+			 current_pt_regs(), SIGBUS, BUS_OBJERR, hva, vcpu_esr);
+}
diff --git a/arch/arm64/kvm/mmu.c b/arch/arm64/kvm/mmu.c
index a71fe6f6bd90f..f5335953827ec 100644
--- a/arch/arm64/kvm/mmu.c
+++ b/arch/arm64/kvm/mmu.c
@@ -1766,13 +1766,7 @@ int kvm_handle_guest_abort(struct kvm_vcpu *vcpu)
 
 	/* Synchronous External Abort? */
 	if (kvm_vcpu_abt_issea(vcpu)) {
-		/*
-		 * For RAS the host kernel may handle this abort.
-		 * There is no need to pass the error into the guest.
-		 */
-		if (kvm_handle_guest_sea(fault_ipa, kvm_vcpu_get_esr(vcpu)))
-			kvm_inject_vabt(vcpu);
-
+		kvm_handle_guest_sea(vcpu);
 		return 1;
 	}
 
-- 
2.47.0.338.g60cca15819-goog

[RFC PATCH v2 2/3] KVM: arm64: set FnV in vcpu's ESR_ELx when host FAR_EL2 is invalid

[Jiaqi Yan]

Certain microarchitectures (e.g. Neoverse V2) do not keep track of
the faulting address for a memory load that consumes poisoned data
and results in a synchronous external abort (SEA). This means the
poisoned guest physical address is unavailable when KVM handles such
SEA in EL2, and FAR_EL2 just holds a garbage value. KVM sends SIGBUS
to interrupt VMM/vCPU but the si_addr will be zero.

In case VMM later asks KVM to synchronously inject a SEA into the
guest, KVM should set FnV bit
- in vcpu's ESR_EL1 to let guest kernel know that FAR_EL1
- in vcpu's ESR_EL2 to let nested virtualization know that FAR_EL2
is also invalid and holds garbage value.

Signed-off-by: Jiaqi Yan <jiaqiyan@google.com>
---
 arch/arm64/kvm/inject_fault.c | 4 ++++
 1 file changed, 4 insertions(+)

diff --git a/arch/arm64/kvm/inject_fault.c b/arch/arm64/kvm/inject_fault.c
index a640e839848e6..2b01b331a4879 100644
--- a/arch/arm64/kvm/inject_fault.c
+++ b/arch/arm64/kvm/inject_fault.c
@@ -13,6 +13,7 @@
 #include <linux/kvm_host.h>
 #include <asm/kvm_emulate.h>
 #include <asm/kvm_nested.h>
+#include <asm/kvm_ras.h>
 #include <asm/esr.h>
 
 static void pend_sync_exception(struct kvm_vcpu *vcpu)
@@ -81,6 +82,9 @@ static void inject_abt64(struct kvm_vcpu *vcpu, bool is_iabt, unsigned long addr
 	if (!is_iabt)
 		esr |= ESR_ELx_EC_DABT_LOW << ESR_ELx_EC_SHIFT;
 
+	if (!kvm_vcpu_sea_far_valid(vcpu))
+		esr |= ESR_ELx_FnV;
+
 	esr |= ESR_ELx_FSC_EXTABT;
 
 	if (match_target_el(vcpu, unpack_vcpu_flag(EXCEPT_AA64_EL1_SYNC))) {
-- 
2.47.0.338.g60cca15819-goog

[RFC PATCH v2 3/3] Documentation: kvm: new UAPI when arm64 guest consumes UER

[Jiaqi Yan]

Add the documentation for new UAPI when guest consumes uncorrectable
but recoverable memory error (UER). This new UAPI enables userspace
to inject SEA into the guest.

Tested:
make htmldocs and proofreading

Signed-off-by: Jiaqi Yan <jiaqiyan@google.com>
---
 Documentation/virt/kvm/api.rst | 46 ++++++++++++++++++++++++++++------
 1 file changed, 39 insertions(+), 7 deletions(-)

diff --git a/Documentation/virt/kvm/api.rst b/Documentation/virt/kvm/api.rst
index 454c2aaa155e5..67645bdb66fe5 100644
--- a/Documentation/virt/kvm/api.rst
+++ b/Documentation/virt/kvm/api.rst
@@ -1288,16 +1288,48 @@ ARM64:
 
 User space may need to inject several types of events to the guest.
 
+Inject SError
+~~~~~~~~~~~~~
+
 Set the pending SError exception state for this VCPU. It is not possible to
 'cancel' an Serror that has been made pending.
 
-If the guest performed an access to I/O memory which could not be handled by
-userspace, for example because of missing instruction syndrome decode
-information or because there is no device mapped at the accessed IPA, then
-userspace can ask the kernel to inject an external abort using the address
-from the exiting fault on the VCPU. It is a programming error to set
-ext_dabt_pending after an exit which was not either KVM_EXIT_MMIO or
-KVM_EXIT_ARM_NISV. This feature is only available if the system supports
+Inject SEA
+~~~~~~~~~~
+
+- If the guest performed an access to I/O memory which could not be handled by
+  userspace, for example because of missing instruction syndrome decode
+  information or because there is no device mapped at the accessed IPA, then
+  userspace can ask the kernel to inject an external abort using the address
+  from the exiting fault on the VCPU.
+
+- If the guest consumed an uncorrectable memory error, and RAS extension in
+  Trusted Firmware choosed to notify PE with SEA, KVM and core kernel may have
+  to handle the memory poison consumption when host APEI was unable to claim
+  the SEA. For the following type of faults, KVM sends SIGBUS to current thread
+  (i.e. VMM in EL0) with si_code=BUS_OBJERR:
+
+  - Synchronous External abort, not on translation table walk or hardware
+    update of translation table.
+
+  - Synchronous External abort on translation table walk or hardware update of
+    translation table, level 1.
+
+  - Synchronous parity or ECC error on memory access, not on translation table
+    walk.
+
+  - Synchronous parity or ECC error on memory access on translation table walk
+    or hardware update of translation table, level 1.
+
+  If the memory error's physical address is available, si_addr will be the
+  error's host virtual address in VM's memory space; otherwise si_addr is zero.
+  When userspace VMM is interrupted by such SIGBUS, it can ask KVM to replay
+  an external abort into guest.
+
+It is a programming error to set ext_dabt_pending after an exit which was not
+KVM_EXIT_MMIO, not KVM_EXIT_ARM_NISV, and not interrupted by BUS_OBJERR SIGBUS.
+
+This feature is only available if the system supports
 KVM_CAP_ARM_INJECT_EXT_DABT. This is a helper which provides commonality in
 how userspace reports accesses for the above cases to guests, across different
 userspace implementations. Nevertheless, userspace can still emulate all Arm
-- 
2.47.0.338.g60cca15819-goog