Few timer and AIA fixes for KVM RISC-V
Subsystems (3)
v2 cover Few timer and AIA fixes for KVM RISC-V
The RISC-V Privileged specificaiton says the following: "WFI is also required to resume execution for locally enabled interrupts pending at any privilege level, regardless of the global interrupt enable at each privilege level."
Based on the above, if there is pending VS-timer interrupt when the host (aka HS-mode) executes WFI then such a WFI will simply become NOP and not do anything. This result in QEMU RISC-V consuming a lot of CPU time on the x86 machine where it is running. The PATCH1 solves this issue by adding appropriate cleanup in KVM RISC-V timer virtualization.
As a result PATCH1, race conditions in updating HGEI[E|P] CSRs when a VCPU is moved from one host CPU to another are being observed on QEMU so the PATCH2 tries to minimize the chances of these race conditions.
Changes since v1:
- Added more details about race condition in PATCH2 commit description.
Anup Patel (2):
RISC-V: KVM: Disable vstimecmp before exiting to user-space
RISC-V: KVM: Move HGEI[E|P] CSR access to IMSIC virtualization
arch/riscv/include/asm/kvm_aia.h | 4 ++- arch/riscv/kvm/aia.c | 51 +++++--------------------------- arch/riscv/kvm/aia_imsic.c | 45 ++++++++++++++++++++++++++++ arch/riscv/kvm/vcpu.c | 2 -- arch/riscv/kvm/vcpu_timer.c | 16 ++++++++++ 5 files changed, 71 insertions(+), 47 deletions(-)
Signature · 1 line
2.43.0
PATCH v2 1/2 RISC-V: KVM: Disable vstimecmp before exiting to user-space
If VS-timer expires when no VCPU running on a host CPU then WFI executed by such host CPU will be effective NOP resulting in no power savings. This is as-per RISC-V Privileged specificaiton which says: "WFI is also required to resume execution for locally enabled interrupts pending at any privilege level, regardless of the global interrupt enable at each privilege level."
To address the above issue, vstimecmp CSR must be set to -1UL over here when VCPU is scheduled-out or exits to user space.
--- arch/riscv/kvm/vcpu_timer.c | 16 ++++++++++++++++ 1 file changed, 16 insertions(+)
arch/riscv/kvm/vcpu_timer.c +16 -0
@@ -345,8 +345,24 @@ void kvm_riscv_vcpu_timer_save(struct kvm_vcpu *vcpu)/* * The vstimecmp CSRs are saved by kvm_riscv_vcpu_timer_sync() * upon every VM exit so no need to save here. +* +* If VS-timer expires when no VCPU running on a host CPU then +* WFI executed by such host CPU will be effective NOP resulting +* in no power savings. This is because as-per RISC-V Privileged +* specificaiton: "WFI is also required to resume execution for +* locally enabled interrupts pending at any privilege level, +* regardless of the global interrupt enable at each privilege +* level." +* +* To address the above issue, vstimecmp CSR must be set to -1UL +* over here when VCPU is scheduled-out or exits to user space. */ +csr_write (CSR_VSTIMECMP ,-1UL ); +#if defined(CONFIG_32BIT ) +csr_write (CSR_VSTIMECMPH ,-1UL ); +#endif +/* timer should be enabled for the remaining operations */ if (unlikely (!t->init_done))return ;
Signature · 1 line
2.43.0
Re: [PATCH v2 1/2] RISC-V: KVM: Disable vstimecmp before exiting to user-space
Quoted 41 lines · Anup Patel
If VS-timer expires when no VCPU running on a host CPU then WFI executed by such host CPU will be effective NOP resulting in no power savings. This is as-per RISC-V Privileged specificaiton which says: "WFI is also required to resume execution for locally enabled interrupts pending at any privilege level, regardless of the global interrupt enable at each privilege level."
To address the above issue, vstimecmp CSR must be set to -1UL over here when VCPU is scheduled-out or exits to user space.
Reviewed-by Atish Patra <[email protected]>Tested-by Atish Patra <[email protected]>Tested-by Heinrich Schuchardt <[email protected]>Fixes 8f5cb44b1bae ("RISC-V: KVM: Support sstc extension")Fixes cea8896bd936 ("RISC-V: KVM: Fix kvm_riscv_vcpu_timer_pending() for Sstc")Reported-by Heinrich Schuchardt <[email protected]>Signed-off-by Anup Patel <[email protected]>--- arch/riscv/kvm/vcpu_timer.c | 16 ++++++++++++++++ 1 file changed, 16 insertions(+)arch/riscv/kvm/vcpu_timer.c +11
-0@@ -345,8 +345,24 @@ void kvm_riscv_vcpu_timer_save(struct kvm_vcpu *vcpu)/* * The vstimecmp CSRs are saved by kvm_riscv_vcpu_timer_sync() * upon every VM exit so no need to save here. +* +* If VS-timer expires when no VCPU running on a host CPU then +* WFI executed by such host CPU will be effective NOP resulting +* in no power savings. This is because as-per RISC-V Privileged +* specificaiton: "WFI is also required to resume execution for +* locally enabled interrupts pending at any privilege level, +* regardless of the global interrupt enable at each privilege +* level." +* +* To address the above issue, vstimecmp CSR must be set to -1UL +* over here when VCPU is scheduled-out or exits to user space.
arch/riscv/kvm/vcpu_timer.c void kvm_riscv_vcpu_timer_save(struct kvm_vcpu *vcpu) · line 368
Quoted 10 lines
arch/riscv/kvm/vcpu_timer.c +5
-0*/ +csr_write (CSR_VSTIMECMP ,-1UL ); +#if defined(CONFIG_32BIT ) +csr_write (CSR_VSTIMECMPH ,-1UL ); +#endif +/* timer should be enabled for the remaining operations */ if (unlikely (!t->init_done))return ;
Thanks,
Nutty
PATCH v2 2/2 RISC-V: KVM: Move HGEI[E|P] CSR access to IMSIC virtualization
In reply to Anup PatelCurrently, the common AIA functions kvm_riscv_vcpu_aia_has_interrupts() and kvm_riscv_aia_wakeon_hgei() lookup HGEI line using an array of VCPU pointers before accessing HGEI[E|P] CSR which is slow and prone to race conditions because there is a separate per-hart lock for the VCPU pointer array and a separate per-VCPU rwlock for IMSIC VS-file (including HGEI line) used by the VCPU. Due to these race conditions, it is observed on QEMU RISC-V host that Guest VCPUs sleep in WFI and never wakeup even with interrupt pending in the IMSIC VS-file because VCPUs were waiting for HGEI wakeup on the wrong host CPU.
The IMSIC virtualization already keeps track of the HGEI line and the associated IMSIC VS-file used by each VCPU so move the HGEI[E|P] CSR access to IMSIC virtualization so that costly HGEI line lookup can be avoided and likelihood of race-conditions when updating HGEI[E|P] CSR is also reduced.
--- arch/riscv/include/asm/kvm_aia.h | 4 ++- arch/riscv/kvm/aia.c | 51 +++++--------------------------- arch/riscv/kvm/aia_imsic.c | 45 ++++++++++++++++++++++++++++ arch/riscv/kvm/vcpu.c | 2 -- 4 files changed, 55 insertions(+), 47 deletions(-)
arch/riscv/include/asm/kvm_aia.h +3 -1
@@ -87,6 +87,9 @@ DECLARE_STATIC_KEY_FALSE(kvm_riscv_aia_available);extern struct kvm_device_ops kvm_riscv_aia_device_ops; +bool kvm_riscv_vcpu_aia_imsic_has_interrupt (struct kvm_vcpu *vcpu); +void kvm_riscv_vcpu_aia_imsic_load (struct kvm_vcpu *vcpu,int cpu); +void kvm_riscv_vcpu_aia_imsic_put (struct kvm_vcpu *vcpu);void kvm_riscv_vcpu_aia_imsic_release (struct kvm_vcpu *vcpu);int kvm_riscv_vcpu_aia_imsic_update (struct kvm_vcpu *vcpu);
@@ -161,7 +164,6 @@ void kvm_riscv_aia_destroy_vm(struct kvm *kvm);int kvm_riscv_aia_alloc_hgei (int cpu,struct kvm_vcpu *owner,void __iomem **hgei_va,phys_addr_t *hgei_pa);void kvm_riscv_aia_free_hgei (int cpu,int hgei); -void kvm_riscv_aia_wakeon_hgei (struct kvm_vcpu *owner,bool enable);void kvm_riscv_aia_enable (void );void kvm_riscv_aia_disable (void );
arch/riscv/kvm/aia.c +7 -44
87 lines not on this page · raw message
arch/riscv/kvm/aia_imsic.c +45 -0
57 lines not on this page · raw message
arch/riscv/kvm/vcpu.c +0 -2
@@ -211,12 +211,10 @@ int kvm_cpu_has_pending_timer(struct kvm_vcpu *vcpu)void kvm_arch_vcpu_blocking (struct kvm_vcpu *vcpu) { -kvm_riscv_aia_wakeon_hgei (vcpu, true); }void kvm_arch_vcpu_unblocking (struct kvm_vcpu *vcpu) { -kvm_riscv_aia_wakeon_hgei (vcpu, false); }int kvm_arch_vcpu_runnable (struct kvm_vcpu *vcpu)
Signature · 1 line
2.43.0
Re: [PATCH v2 2/2] RISC-V: KVM: Move HGEI[E|P] CSR access to IMSIC virtualization
Quoted 216 lines · Anup Patel
Currently, the common AIA functions
kvm_riscv_vcpu_aia_has_interrupts()andkvm_riscv_aia_wakeon_hgei()lookup HGEI line using an array of VCPU pointers before accessing HGEI[E|P] CSR which is slow and prone to race conditions because there is a separate per-hart lock for the VCPU pointer array and a separate per-VCPU rwlock for IMSIC VS-file (including HGEI line) used by the VCPU. Due to these race conditions, it is observed on QEMU RISC-V host that Guest VCPUs sleep in WFI and never wakeup even with interrupt pending in the IMSIC VS-file because VCPUs were waiting for HGEI wakeup on the wrong host CPU.The IMSIC virtualization already keeps track of the HGEI line and the associated IMSIC VS-file used by each VCPU so move the HGEI[E|P] CSR access to IMSIC virtualization so that costly HGEI line lookup can be avoided and likelihood of race-conditions when updating HGEI[E|P] CSR is also reduced.
Reviewed-by Atish Patra <[email protected]>Tested-by Atish Patra <[email protected]>Tested-by Heinrich Schuchardt <[email protected]>Fixes 3385339296d1 ("RISC-V: KVM: Use IMSIC guest files when available")Signed-off-by Anup Patel <[email protected]>--- arch/riscv/include/asm/kvm_aia.h | 4 ++- arch/riscv/kvm/aia.c | 51 +++++--------------------------- arch/riscv/kvm/aia_imsic.c | 45 ++++++++++++++++++++++++++++ arch/riscv/kvm/vcpu.c | 2 -- 4 files changed, 55 insertions(+), 47 deletions(-)arch/riscv/include/asm/kvm_aia.h +3
-1@@ -87,6 +87,9 @@ DECLARE_STATIC_KEY_FALSE(kvm_riscv_aia_available);extern struct kvm_device_ops kvm_riscv_aia_device_ops; +bool kvm_riscv_vcpu_aia_imsic_has_interrupt (struct kvm_vcpu *vcpu); +void kvm_riscv_vcpu_aia_imsic_load (struct kvm_vcpu *vcpu,int cpu); +void kvm_riscv_vcpu_aia_imsic_put (struct kvm_vcpu *vcpu);void kvm_riscv_vcpu_aia_imsic_release (struct kvm_vcpu *vcpu);int kvm_riscv_vcpu_aia_imsic_update (struct kvm_vcpu *vcpu);@@ -161,7 +164,6 @@ void kvm_riscv_aia_destroy_vm(struct kvm *kvm);int kvm_riscv_aia_alloc_hgei (int cpu,struct kvm_vcpu *owner,void __iomem **hgei_va,phys_addr_t *hgei_pa);void kvm_riscv_aia_free_hgei (int cpu,int hgei); -void kvm_riscv_aia_wakeon_hgei (struct kvm_vcpu *owner,bool enable);void kvm_riscv_aia_enable (void );void kvm_riscv_aia_disable (void );arch/riscv/kvm/aia.c +7
-44@@ -30,28 +30,6 @@ unsigned int kvm_riscv_aia_nr_hgei;unsigned int kvm_riscv_aia_max_ids;DEFINE_STATIC_KEY_FALSE (kvm_riscv_aia_available); -static int aia_find_hgei (struct kvm_vcpu *owner) -{ -int i, hgei; -unsigned long flags; -struct aia_hgei_control *hgctrl =get_cpu_ptr (&aia_hgei); - -raw_spin_lock_irqsave (&hgctrl->lock, flags); - - hgei =-1 ; -for (i =1 ; i <= kvm_riscv_aia_nr_hgei; i++) { -if (hgctrl->owners[i] == owner) { - hgei = i; -break ; - } - } - -raw_spin_unlock_irqrestore (&hgctrl->lock, flags); - -put_cpu_ptr (&aia_hgei); -return hgei; -} -static inline unsigned long aia_hvictl_value (bool ext_irq_pending) {unsigned long hvictl;@@ -95,7 +73,6 @@ void kvm_riscv_vcpu_aia_sync_interrupts(struct kvm_vcpu *vcpu)bool kvm_riscv_vcpu_aia_has_interrupts (struct kvm_vcpu *vcpu,u64 mask) { -int hgei;unsigned long seip;if (!kvm_riscv_aia_available ())@@ -114,11 +91,7 @@ bool kvm_riscv_vcpu_aia_has_interrupts(struct kvm_vcpu *vcpu, u64 mask)if (!kvm_riscv_aia_initialized (vcpu->kvm) || !seip)return false; - hgei =aia_find_hgei (vcpu); -if (hgei >0 ) -return !!(ncsr_read (CSR_HGEIP ) &BIT (hgei)); - -return false; +return kvm_riscv_vcpu_aia_imsic_has_interrupt (vcpu); }void kvm_riscv_vcpu_aia_update_hvip (struct kvm_vcpu *vcpu)@@ -164,6 +137,9 @@ void kvm_riscv_vcpu_aia_load(struct kvm_vcpu *vcpu, int cpu)csr_write (CSR_HVIPRIO2H , csr->hviprio2h);#endif } + +if (kvm_riscv_aia_initialized (vcpu->kvm)) +kvm_riscv_vcpu_aia_imsic_load (vcpu, cpu); }void kvm_riscv_vcpu_aia_put (struct kvm_vcpu *vcpu)@@ -174,6 +150,9 @@ void kvm_riscv_vcpu_aia_put(struct kvm_vcpu *vcpu)if (!kvm_riscv_aia_available ())return ; +if (kvm_riscv_aia_initialized (vcpu->kvm)) +kvm_riscv_vcpu_aia_imsic_put (vcpu); +if (kvm_riscv_nacl_available ()) { nsh =nacl_shmem (); csr->vsiselect =nacl_csr_read (nsh,CSR_VSISELECT );@@ -472,22 +451,6 @@ void kvm_riscv_aia_free_hgei(int cpu, int hgei)raw_spin_unlock_irqrestore (&hgctrl->lock, flags); } -void kvm_riscv_aia_wakeon_hgei (struct kvm_vcpu *owner,bool enable) -{ -int hgei; - -if (!kvm_riscv_aia_available ()) -return ; - - hgei =aia_find_hgei (owner); -if (hgei >0 ) { -if (enable) -csr_set (CSR_HGEIE ,BIT (hgei)); -else -csr_clear (CSR_HGEIE ,BIT (hgei)); - } -} -static irqreturn_t hgei_interrupt (int irq,void *dev_id) {int i;arch/riscv/kvm/aia_imsic.c +45
-0@@ -677,6 +677,48 @@ static void imsic_swfile_update(struct kvm_vcpu *vcpu,imsic_swfile_extirq_update (vcpu); } +bool kvm_riscv_vcpu_aia_imsic_has_interrupt (struct kvm_vcpu *vcpu) +{ +struct imsic *imsic = vcpu->arch.aia_context.imsic_state; +unsigned long flags; +bool ret = false; + +/* +* The IMSIC SW-file directly injects interrupt via hvip so +* only check for interrupt when IMSIC VS-file is being used. +*/ + +read_lock_irqsave (&imsic->vsfile_lock, flags); +if (imsic->vsfile_cpu >-1 ) + ret = !!(csr_read (CSR_HGEIP ) &BIT (imsic->vsfile_hgei)); +read_unlock_irqrestore (&imsic->vsfile_lock, flags); + +return ret; +} + +void kvm_riscv_vcpu_aia_imsic_load (struct kvm_vcpu *vcpu,int cpu) +{ +/* +* No need to explicitly clear HGEIE CSR bits because the +* hgei interrupt handler (aka hgei_interrupt()) will always +* clear it for us. +*/ +} + +void kvm_riscv_vcpu_aia_imsic_put (struct kvm_vcpu *vcpu) +{ +struct imsic *imsic = vcpu->arch.aia_context.imsic_state; +unsigned long flags; + +if (!kvm_vcpu_is_blocking (vcpu)) +return ; + +read_lock_irqsave (&imsic->vsfile_lock, flags); +if (imsic->vsfile_cpu >-1 ) +csr_set (CSR_HGEIE ,BIT (imsic->vsfile_hgei)); +read_unlock_irqrestore (&imsic->vsfile_lock, flags); +} +void kvm_riscv_vcpu_aia_imsic_release (struct kvm_vcpu *vcpu) {unsigned long flags;@@ -781,6 +823,9 @@ int kvm_riscv_vcpu_aia_imsic_update(struct kvm_vcpu *vcpu)* producers to the new IMSIC VS-file. */ +/* Ensure HGEIE CSR bit is zero before using the new IMSIC VS-file */ +csr_clear (CSR_HGEIE ,BIT (new_vsfile_hgei)); +/* Zero-out new IMSIC VS-file */ imsic_vsfile_local_clear (new_vsfile_hgei, imsic->nr_hw_eix);arch/riscv/kvm/vcpu.c +0
-0@@ -211,12 +211,10 @@ int kvm_cpu_has_pending_timer(struct kvm_vcpu *vcpu)
arch/riscv/kvm/vcpu.c int kvm_cpu_has_pending_timer(struct kvm_vcpu *vcpu) · line 219
Quoted 10 lines
arch/riscv/kvm/vcpu.c +0
-2void kvm_arch_vcpu_blocking (struct kvm_vcpu *vcpu) { -kvm_riscv_aia_wakeon_hgei (vcpu, true); }void kvm_arch_vcpu_unblocking (struct kvm_vcpu *vcpu) { -kvm_riscv_aia_wakeon_hgei (vcpu, false); }
Nitpick:
Should the above two empty functions be removed ?
Thanks,
Nutty
Quoted 1 line
arch/riscv/kvm/vcpu.c +0
-0int kvm_arch_vcpu_runnable (struct kvm_vcpu *vcpu)
Signature · 3 lines
kvm-riscv mailing list [email protected] http://lists.infradead.org/mailman/listinfo/kvm-riscv
Re: [PATCH v2 2/2] RISC-V: KVM: Move HGEI[E|P] CSR access to IMSIC virtualization
Quoted 223 lines · Nutty Liu
Quoted 221 lines · Anup Patel
Currently, the common AIA functions
kvm_riscv_vcpu_aia_has_interrupts()andkvm_riscv_aia_wakeon_hgei()lookup HGEI line using an array of VCPU pointers before accessing HGEI[E|P] CSR which is slow and prone to race conditions because there is a separate per-hart lock for the VCPU pointer array and a separate per-VCPU rwlock for IMSIC VS-file (including HGEI line) used by the VCPU. Due to these race conditions, it is observed on QEMU RISC-V host that Guest VCPUs sleep in WFI and never wakeup even with interrupt pending in the IMSIC VS-file because VCPUs were waiting for HGEI wakeup on the wrong host CPU.The IMSIC virtualization already keeps track of the HGEI line and the associated IMSIC VS-file used by each VCPU so move the HGEI[E|P] CSR access to IMSIC virtualization so that costly HGEI line lookup can be avoided and likelihood of race-conditions when updating HGEI[E|P] CSR is also reduced.
Reviewed-by Atish Patra <[email protected]>Tested-by Atish Patra <[email protected]>Tested-by Heinrich Schuchardt <[email protected]>Fixes 3385339296d1 ("RISC-V: KVM: Use IMSIC guest files when available")Signed-off-by Anup Patel <[email protected]>--- arch/riscv/include/asm/kvm_aia.h | 4 ++- arch/riscv/kvm/aia.c | 51 +++++--------------------------- arch/riscv/kvm/aia_imsic.c | 45 ++++++++++++++++++++++++++++ arch/riscv/kvm/vcpu.c | 2 -- 4 files changed, 55 insertions(+), 47 deletions(-)arch/riscv/include/asm/kvm_aia.h +3
-1@@ -87,6 +87,9 @@ DECLARE_STATIC_KEY_FALSE(kvm_riscv_aia_available);extern struct kvm_device_ops kvm_riscv_aia_device_ops; +bool kvm_riscv_vcpu_aia_imsic_has_interrupt (struct kvm_vcpu *vcpu); +void kvm_riscv_vcpu_aia_imsic_load (struct kvm_vcpu *vcpu,int cpu); +void kvm_riscv_vcpu_aia_imsic_put (struct kvm_vcpu *vcpu);void kvm_riscv_vcpu_aia_imsic_release (struct kvm_vcpu *vcpu);int kvm_riscv_vcpu_aia_imsic_update (struct kvm_vcpu *vcpu);@@ -161,7 +164,6 @@ void kvm_riscv_aia_destroy_vm(struct kvm *kvm);int kvm_riscv_aia_alloc_hgei (int cpu,struct kvm_vcpu *owner,void __iomem **hgei_va,phys_addr_t *hgei_pa);void kvm_riscv_aia_free_hgei (int cpu,int hgei); -void kvm_riscv_aia_wakeon_hgei (struct kvm_vcpu *owner,bool enable);void kvm_riscv_aia_enable (void );void kvm_riscv_aia_disable (void );arch/riscv/kvm/aia.c +7
-44@@ -30,28 +30,6 @@ unsigned int kvm_riscv_aia_nr_hgei;unsigned int kvm_riscv_aia_max_ids;DEFINE_STATIC_KEY_FALSE (kvm_riscv_aia_available); -static int aia_find_hgei (struct kvm_vcpu *owner) -{ -int i, hgei; -unsigned long flags; -struct aia_hgei_control *hgctrl =get_cpu_ptr (&aia_hgei); - -raw_spin_lock_irqsave (&hgctrl->lock, flags); - - hgei =-1 ; -for (i =1 ; i <= kvm_riscv_aia_nr_hgei; i++) { -if (hgctrl->owners[i] == owner) { - hgei = i; -break ; - } - } - -raw_spin_unlock_irqrestore (&hgctrl->lock, flags); - -put_cpu_ptr (&aia_hgei); -return hgei; -} -static inline unsigned long aia_hvictl_value (bool ext_irq_pending) {unsigned long hvictl;@@ -95,7 +73,6 @@ void kvm_riscv_vcpu_aia_sync_interrupts(struct kvm_vcpu *vcpu)bool kvm_riscv_vcpu_aia_has_interrupts (struct kvm_vcpu *vcpu,u64 mask) { -int hgei;unsigned long seip;if (!kvm_riscv_aia_available ())@@ -114,11 +91,7 @@ bool kvm_riscv_vcpu_aia_has_interrupts(struct kvm_vcpu *vcpu, u64 mask)if (!kvm_riscv_aia_initialized (vcpu->kvm) || !seip)return false; - hgei =aia_find_hgei (vcpu); -if (hgei >0 ) -return !!(ncsr_read (CSR_HGEIP ) &BIT (hgei)); - -return false; +return kvm_riscv_vcpu_aia_imsic_has_interrupt (vcpu); }void kvm_riscv_vcpu_aia_update_hvip (struct kvm_vcpu *vcpu)@@ -164,6 +137,9 @@ void kvm_riscv_vcpu_aia_load(struct kvm_vcpu *vcpu, int cpu)csr_write (CSR_HVIPRIO2H , csr->hviprio2h);#endif } + +if (kvm_riscv_aia_initialized (vcpu->kvm)) +kvm_riscv_vcpu_aia_imsic_load (vcpu, cpu); }void kvm_riscv_vcpu_aia_put (struct kvm_vcpu *vcpu)@@ -174,6 +150,9 @@ void kvm_riscv_vcpu_aia_put(struct kvm_vcpu *vcpu)if (!kvm_riscv_aia_available ())return ; +if (kvm_riscv_aia_initialized (vcpu->kvm)) +kvm_riscv_vcpu_aia_imsic_put (vcpu); +if (kvm_riscv_nacl_available ()) { nsh =nacl_shmem (); csr->vsiselect =nacl_csr_read (nsh,CSR_VSISELECT );@@ -472,22 +451,6 @@ void kvm_riscv_aia_free_hgei(int cpu, int hgei)raw_spin_unlock_irqrestore (&hgctrl->lock, flags); } -void kvm_riscv_aia_wakeon_hgei (struct kvm_vcpu *owner,bool enable) -{ -int hgei; - -if (!kvm_riscv_aia_available ()) -return ; - - hgei =aia_find_hgei (owner); -if (hgei >0 ) { -if (enable) -csr_set (CSR_HGEIE ,BIT (hgei)); -else -csr_clear (CSR_HGEIE ,BIT (hgei)); - } -} -static irqreturn_t hgei_interrupt (int irq,void *dev_id) {int i;arch/riscv/kvm/aia_imsic.c +45
-0@@ -677,6 +677,48 @@ static void imsic_swfile_update(struct kvm_vcpu *vcpu,imsic_swfile_extirq_update (vcpu); } +bool kvm_riscv_vcpu_aia_imsic_has_interrupt (struct kvm_vcpu *vcpu) +{ +struct imsic *imsic = vcpu->arch.aia_context.imsic_state; +unsigned long flags; +bool ret = false; + +/* +* The IMSIC SW-file directly injects interrupt via hvip so +* only check for interrupt when IMSIC VS-file is being used. +*/ + +read_lock_irqsave (&imsic->vsfile_lock, flags); +if (imsic->vsfile_cpu >-1 ) + ret = !!(csr_read (CSR_HGEIP ) &BIT (imsic->vsfile_hgei)); +read_unlock_irqrestore (&imsic->vsfile_lock, flags); + +return ret; +} + +void kvm_riscv_vcpu_aia_imsic_load (struct kvm_vcpu *vcpu,int cpu) +{ +/* +* No need to explicitly clear HGEIE CSR bits because the +* hgei interrupt handler (aka hgei_interrupt()) will always +* clear it for us. +*/ +} + +void kvm_riscv_vcpu_aia_imsic_put (struct kvm_vcpu *vcpu) +{ +struct imsic *imsic = vcpu->arch.aia_context.imsic_state; +unsigned long flags; + +if (!kvm_vcpu_is_blocking (vcpu)) +return ; + +read_lock_irqsave (&imsic->vsfile_lock, flags); +if (imsic->vsfile_cpu >-1 ) +csr_set (CSR_HGEIE ,BIT (imsic->vsfile_hgei)); +read_unlock_irqrestore (&imsic->vsfile_lock, flags); +} +void kvm_riscv_vcpu_aia_imsic_release (struct kvm_vcpu *vcpu) {unsigned long flags;@@ -781,6 +823,9 @@ int kvm_riscv_vcpu_aia_imsic_update(struct kvm_vcpu *vcpu)* producers to the new IMSIC VS-file. */ +/* Ensure HGEIE CSR bit is zero before using the new IMSIC VS-file */ +csr_clear (CSR_HGEIE ,BIT (new_vsfile_hgei)); +/* Zero-out new IMSIC VS-file */ imsic_vsfile_local_clear (new_vsfile_hgei, imsic->nr_hw_eix);arch/riscv/kvm/vcpu.c +0
-1@@ -211,12 +211,10 @@ int kvm_cpu_has_pending_timer(struct kvm_vcpu *vcpu)void kvm_arch_vcpu_blocking (struct kvm_vcpu *vcpu) { -kvm_riscv_aia_wakeon_hgei (vcpu, true); }
Quoted 9 lines
arch/riscv/kvm/vcpu.c int kvm_cpu_has_pending_timer(struct kvm_vcpu *vcpu) · line 219
Quoted 5 lines
arch/riscv/kvm/vcpu.c +0
-1void kvm_arch_vcpu_unblocking (struct kvm_vcpu *vcpu) { -kvm_riscv_aia_wakeon_hgei (vcpu, false); }Nitpick:
Should the above two empty functions be removed ?
We cannot remove these functions but we can certainly make them static inline. I will update in the next revision.
Thanks,
Anup
Anup Patel
maintainerQuoted 23 lines · Anup Patel
The RISC-V Privileged specificaiton says the following: "WFI is also required to resume execution for locally enabled interrupts pending at any privilege level, regardless of the global interrupt enable at each privilege level."
Based on the above, if there is pending VS-timer interrupt when the host (aka HS-mode) executes WFI then such a WFI will simply become NOP and not do anything. This result in QEMU RISC-V consuming a lot of CPU time on the x86 machine where it is running. The PATCH1 solves this issue by adding appropriate cleanup in KVM RISC-V timer virtualization.
As a result PATCH1, race conditions in updating HGEI[E|P] CSRs when a VCPU is moved from one host CPU to another are being observed on QEMU so the PATCH2 tries to minimize the chances of these race conditions.
Changes since v1:
- Added more details about race condition in PATCH2 commit description.Anup Patel (2):
RISC-V: KVM: Disable vstimecmp before exiting to user-space
RISC-V: KVM: Move HGEI[E|P] CSR access to IMSIC virtualization
Quoted 7 lines
arch/riscv/include/asm/kvm_aia.h | 4 ++- arch/riscv/kvm/aia.c | 51 +++++--------------------------- arch/riscv/kvm/aia_imsic.c | 45 ++++++++++++++++++++++++++++ arch/riscv/kvm/vcpu.c | 2 -- arch/riscv/kvm/vcpu_timer.c | 16 ++++++++++ 5 files changed, 71 insertions(+), 47 deletions(-)
Queued this series as fixes for Linux-6.16
I have taken care of the comment on PATCH2 at the time of queuing.
Thanks,
Anup
This series was applied to riscv/linux.git (for-next) by Anup Patel <[email protected]>:
Hello:
This series was applied to riscv/linux.git (for-next) by Anup Patel <[email protected]>:
Quoted 5 lines · On Mon, 7 Jul 2025 09:23:42 +0530 you wrote:
The RISC-V Privileged specificaiton says the following: "WFI is also required to resume execution for locally enabled interrupts pending at any privilege level, regardless of the global interrupt enable at each privilege level."
Quoted 7 lines
Based on the above, if there is pending VS-timer interrupt when the host (aka HS-mode) executes WFI then such a WFI will simply become NOP and not do anything. This result in QEMU RISC-V consuming a lot of CPU time on the x86 machine where it is running. The PATCH1 solves this issue by adding appropriate cleanup in KVM RISC-V timer virtualization.
[...]
Here is the summary with links:
- [v2,1/2] RISC-V: KVM: Disable vstimecmp before exiting to user-space
https://git.kernel.org/riscv/c/57f576e860d3
- [v2,2/2] RISC-V: KVM: Move HGEI[E|P] CSR access to IMSIC virtualization (no matching commit)
You are awesome, thank you!
Signature · 2 lines
Deet-doot-dot, I am a bot. https://korg.docs.kernel.org/patchwork/pwbot.html