arch/riscv/include/asm/kvm_vcpu_vector.h | 24 +++++++ arch/riscv/include/asm/processor.h | 8 +++ arch/riscv/include/asm/simd.h | 16 +---- arch/riscv/include/asm/vector.h | 5 ++ arch/riscv/kernel/kernel_mode_vector.c | 88 +++++++++++++++++------- arch/riscv/kvm/main.c | 4 ++ arch/riscv/kvm/vcpu.c | 12 ++++ arch/riscv/kvm/vcpu_vector.c | 22 +++++- 8 files changed, 139 insertions(+), 40 deletions(-)
This series fixes a vtype corruption encountered when running perf +
vector workload on KVM.
The root cause of the bug is that the kernel-mode vector (KMV)
misattributes the guest's vcpu context as the user's context. To solve
this, we need to correctly save the vcpu context when the kernel-mode
vector is serving a guest.
However, calling directly into KVM from RISC-V generic architecture code
creates a reverse dependency, which is problematic when KVM is built as
a module. To address this, we introduce an RCU-protected callback for
context flushing, which KVM registers during module init.
Patch 1 is a preparatory cleanup that refactors
riscv_v_start_kernel_context().
Patch 2 prepares get/put_cpu_vector_context() for gaurding the use of
vector in kvm_arch_vcpu_load/put()
Patch 3 implements the callback mechanism and fixes the context handling.
Reason for this respin:
put_cpu_vector_context may casue a volutary ctxswch. If we prematurely
set RISCV_PREEMPT_V, the context restore path will rasise NEED_RESTORE
for this kernel thread. As the result, any irq taken during the
subsequence preempt_v would trigger a restore from the empty/stale
context memory on the irq return path in riscv_v_context_nesting_end()
The fix is summarized as below:
--- a/arch/riscv/kernel/kernel_mode_vector.c
+++ b/arch/riscv/kernel/kernel_mode_vector.c
@@ -181,8 +181,8 @@ static int riscv_v_start_kernel_context(void)
/* Transfer the ownership of V from user to kernel, then save */
get_cpu_vector_context();
__riscv_flush_vector_context();
- riscv_v_start(RISCV_PREEMPT_V);
put_cpu_vector_context();
+ riscv_v_start(RISCV_PREEMPT_V);
return 0;
}
We will send a optimization patch that removes the costly riscv_v_is_on()
for voluntary switch detection. A kselftest that stresses the vectorized
user copy will also be included into that series.
Patch summary:
- unchanged patch: 1, 2
- new patch: none
- modified patch: 3
Changelog v5:
- add the r-b from Yong-Xuan
- fix a guest boot fail that hits ~1/100 in host kernel-mode vector
- Link to v4: https://lore.kernel.org/all/20260725001749.2579274-1-tchiu@tenstorrent.com/
Changelog v4:
- Include a header to solve a mid-series build fail (patchwork ci)
- Drop preempt_v_started test in may_use_simd (Sashiko)
- Link to v3: https://lore.kernel.org/all/20260724165001.2317788-1-tchiu@tenstorrent.com/
Changelog v3:
- Limit the export scope for {get,put}_cpu_vector_context()
- clears RISCV_V_VCPU_NEED_RESTORE flag in host restore to prevent
leaking
- consolidates guest vector restore at returning to guest to prevent
unnecessary save/restore between the preemptible window from
vcpu_load to vcpu_enter_exit
- Document flags added to riscv_v_flags
- Link to v2: https://lore.kernel.org/all/20260715051629.1169645-1-tchiu@tenstorrent.com/
Changelog v2:
- Address issues pointed out by sashiko (2, 3)
- Link to v1: https://lore.kernel.org/all/20260711015835.767259-1-tchiu@tenstorrent.com/
Andy Chiu (3):
riscv: vector: refactor riscv_v_start_kernel_context
riscv: vector: allow non-preemptible kernel-mode vector with IRQs off
RISC-V: KVM: fix vcpu vector context handling for kernel-mode vector
arch/riscv/include/asm/kvm_vcpu_vector.h | 24 +++++++
arch/riscv/include/asm/processor.h | 8 +++
arch/riscv/include/asm/simd.h | 16 +----
arch/riscv/include/asm/vector.h | 5 ++
arch/riscv/kernel/kernel_mode_vector.c | 88 +++++++++++++++++-------
arch/riscv/kvm/main.c | 4 ++
arch/riscv/kvm/vcpu.c | 12 ++++
arch/riscv/kvm/vcpu_vector.c | 22 +++++-
8 files changed, 139 insertions(+), 40 deletions(-)
--
2.43.0On Tue, Aug 4, 2026 at 3:23 AM Andy Chiu <tchiu@tenstorrent.com> wrote:
>
> This series fixes a vtype corruption encountered when running perf +
> vector workload on KVM.
>
> The root cause of the bug is that the kernel-mode vector (KMV)
> misattributes the guest's vcpu context as the user's context. To solve
> this, we need to correctly save the vcpu context when the kernel-mode
> vector is serving a guest.
>
> However, calling directly into KVM from RISC-V generic architecture code
> creates a reverse dependency, which is problematic when KVM is built as
> a module. To address this, we introduce an RCU-protected callback for
> context flushing, which KVM registers during module init.
>
> Patch 1 is a preparatory cleanup that refactors
> riscv_v_start_kernel_context().
> Patch 2 prepares get/put_cpu_vector_context() for gaurding the use of
> vector in kvm_arch_vcpu_load/put()
> Patch 3 implements the callback mechanism and fixes the context handling.
>
> Reason for this respin:
>
> put_cpu_vector_context may casue a volutary ctxswch. If we prematurely
> set RISCV_PREEMPT_V, the context restore path will rasise NEED_RESTORE
> for this kernel thread. As the result, any irq taken during the
> subsequence preempt_v would trigger a restore from the empty/stale
> context memory on the irq return path in riscv_v_context_nesting_end()
>
> The fix is summarized as below:
> --- a/arch/riscv/kernel/kernel_mode_vector.c
> +++ b/arch/riscv/kernel/kernel_mode_vector.c
> @@ -181,8 +181,8 @@ static int riscv_v_start_kernel_context(void)
> /* Transfer the ownership of V from user to kernel, then save */
> get_cpu_vector_context();
> __riscv_flush_vector_context();
> - riscv_v_start(RISCV_PREEMPT_V);
> put_cpu_vector_context();
> + riscv_v_start(RISCV_PREEMPT_V);
> return 0;
> }
>
> We will send a optimization patch that removes the costly riscv_v_is_on()
> for voluntary switch detection. A kselftest that stresses the vectorized
> user copy will also be included into that series.
>
> Patch summary:
> - unchanged patch: 1, 2
> - new patch: none
> - modified patch: 3
>
> Changelog v5:
> - add the r-b from Yong-Xuan
> - fix a guest boot fail that hits ~1/100 in host kernel-mode vector
> - Link to v4: https://lore.kernel.org/all/20260725001749.2579274-1-tchiu@tenstorrent.com/
>
> Changelog v4:
> - Include a header to solve a mid-series build fail (patchwork ci)
> - Drop preempt_v_started test in may_use_simd (Sashiko)
> - Link to v3: https://lore.kernel.org/all/20260724165001.2317788-1-tchiu@tenstorrent.com/
>
> Changelog v3:
> - Limit the export scope for {get,put}_cpu_vector_context()
> - clears RISCV_V_VCPU_NEED_RESTORE flag in host restore to prevent
> leaking
> - consolidates guest vector restore at returning to guest to prevent
> unnecessary save/restore between the preemptible window from
> vcpu_load to vcpu_enter_exit
> - Document flags added to riscv_v_flags
> - Link to v2: https://lore.kernel.org/all/20260715051629.1169645-1-tchiu@tenstorrent.com/
>
> Changelog v2:
> - Address issues pointed out by sashiko (2, 3)
> - Link to v1: https://lore.kernel.org/all/20260711015835.767259-1-tchiu@tenstorrent.com/
>
>
> Andy Chiu (3):
> riscv: vector: refactor riscv_v_start_kernel_context
> riscv: vector: allow non-preemptible kernel-mode vector with IRQs off
> RISC-V: KVM: fix vcpu vector context handling for kernel-mode vector
>
> arch/riscv/include/asm/kvm_vcpu_vector.h | 24 +++++++
> arch/riscv/include/asm/processor.h | 8 +++
> arch/riscv/include/asm/simd.h | 16 +----
> arch/riscv/include/asm/vector.h | 5 ++
> arch/riscv/kernel/kernel_mode_vector.c | 88 +++++++++++++++++-------
> arch/riscv/kvm/main.c | 4 ++
> arch/riscv/kvm/vcpu.c | 12 ++++
> arch/riscv/kvm/vcpu_vector.c | 22 +++++-
> 8 files changed, 139 insertions(+), 40 deletions(-)
>
> --
> 2.43.0
LGTM. I did not see any obvious issue on QEMU virt machine.
Reviewed-by: Anup Patel <anup@brainfault.org>
Queued this patch for Linux-7.3
Thanks,
Anup
© 2016 - 2026 Red Hat, Inc.