Skip to contentSkip to conversation

RISC-V: KVM: Using user-mode pte within kvm_riscv_gstage_ioremap

PATCH In review [email protected] 5 participants 8 messages Aug 7, 2025 – Aug 19, 2025

Subsystems (3)
lore
Share & tools
lore
#0 lore raw

PATCH 1/1 RISC-V: KVM: Using user-mode pte within kvm_riscv_gstage_ioremap

From: Fangyu Yu <[email protected]>

Currently we use kvm_riscv_gstage_ioremap to map IMSIC gpa to the spa of guest interrupt file within IMSIC.

The PAGE_KERNEL_IO property does not include user mode settings, so when accessing the IMSIC address in the virtual machine, a guest page fault will occur, this is not expected.

According to the RISC-V Privileged Architecture Spec, for G-stage address translation, all memory accesses are considered to be user-level accesses as though executed in Umode.

---
 arch/riscv/kvm/mmu.c | 6 +++++-
 1 file changed, 5 insertions(+), 1 deletion(-)
arch/riscv/kvm/mmu.c +5 -1
index 1087ea74567b..800064e96ef6 100644
@@ -351,6 +351,7 @@ int kvm_riscv_gstage_ioremap(struct kvm *kvm, gpa_t gpa,
 	int ret = 0;
 	unsigned long pfn;
 	phys_addr_t addr, end;
+	pgprot_t prot;
 	struct kvm_mmu_memory_cache pcache = {
 		.gfp_custom = (in_atomic) ? GFP_ATOMIC | __GFP_ACCOUNT : 0,
 		.gfp_zero = __GFP_ZERO,
@@ -359,8 +360,11 @@ int kvm_riscv_gstage_ioremap(struct kvm *kvm, gpa_t gpa,
 	end = (gpa + size + PAGE_SIZE - 1) & PAGE_MASK;
 	pfn = __phys_to_pfn(hpa);
 
+	prot = pgprot_noncached(PAGE_WRITE);
+
 	for (addr = gpa; addr < end; addr += PAGE_SIZE) {
-		pte = pfn_pte(pfn, PAGE_KERNEL_IO);
+		pte = pfn_pte(pfn, prot);
+		pte = pte_mkdirty(pte);
 
 		if (!writable)
 			pte = pte_wrprotect(pte);
Signature · 1 line
2.39.3 (Apple Git-146)
Reviewed-by on v1 1/1
#1 lore raw
Quoted 9 lines · On 8/7/2025 3:07 PM, [email protected] wrote:

From: Fangyu Yu <[email protected]>

Currently we use kvm_riscv_gstage_ioremap to map IMSIC gpa to the spa of guest interrupt file within IMSIC.

The PAGE_KERNEL_IO property does not include user mode settings, so when accessing the IMSIC address in the virtual machine, a guest page fault will occur, this is not expected.

Quoted 8 lines

According to the RISC-V Privileged Architecture Spec, for G-stage address translation, all memory accesses are considered to be user-level accesses as though executed in Umode.

---
  arch/riscv/kvm/mmu.c | 6 +++++-
  1 file changed, 5 insertions(+), 1 deletion(-)

Thanks,
Nutty

Quoted 25 lines
arch/riscv/kvm/mmu.c +5 -1
index 1087ea74567b..800064e96ef6 100644
@@ -351,6 +351,7 @@ int kvm_riscv_gstage_ioremap(struct kvm *kvm, gpa_t gpa,
  	int ret = 0;
  	unsigned long pfn;
  	phys_addr_t addr, end;
+	pgprot_t prot;
  	struct kvm_mmu_memory_cache pcache = {
  		.gfp_custom = (in_atomic) ? GFP_ATOMIC | __GFP_ACCOUNT : 0,
  		.gfp_zero = __GFP_ZERO,
@@ -359,8 +360,11 @@ int kvm_riscv_gstage_ioremap(struct kvm *kvm, gpa_t gpa,
  	end = (gpa + size + PAGE_SIZE - 1) & PAGE_MASK;
  	pfn = __phys_to_pfn(hpa);
 
+	prot = pgprot_noncached(PAGE_WRITE);
+
  	for (addr = gpa; addr < end; addr += PAGE_SIZE) {
-		pte = pfn_pte(pfn, PAGE_KERNEL_IO);
+		pte = pfn_pte(pfn, prot);
+		pte = pte_mkdirty(pte);
 
  		if (!writable)
  			pte = pte_wrprotect(pte);
Reviewed-by on v1 1/1
#2 lore raw
In reply to [email protected]
2025-08-07T15:07:29+08:00, <[email protected]>:
Quoted 3 lines

From: Fangyu Yu <[email protected]>

Currently we use kvm_riscv_gstage_ioremap to map IMSIC gpa to the spa of

                                                                    ^^^
                                                                    hpa?
Quoted 5 lines

guest interrupt file within IMSIC.

The PAGE_KERNEL_IO property does not include user mode settings, so when accessing the IMSIC address in the virtual machine, a guest page fault will occur, this is not expected.

PAGE_KERNEL_IO also set the reserved G bit, so you're fixing two issues with a single change. :)

Quoted 3 lines

According to the RISC-V Privileged Architecture Spec, for G-stage address translation, all memory accesses are considered to be user-level accesses as though executed in Umode.

What implementation are you using? I would have assume that the original code was tested on QEMU, so we might have a bug there.

arch/riscv/kvm/mmu.c int kvm_riscv_gstage_ioremap(struct kvm *kvm, gpa_t gpa, · line 367

Quoted 13 lines

---

arch/riscv/kvm/mmu.c +4 -1
@@ -359,8 +360,11 @@ int kvm_riscv_gstage_ioremap(struct kvm *kvm, gpa_t gpa,
 	end = (gpa + size + PAGE_SIZE - 1) & PAGE_MASK;
 	pfn = __phys_to_pfn(hpa);
 
+	prot = pgprot_noncached(PAGE_WRITE);
+
 	for (addr = gpa; addr < end; addr += PAGE_SIZE) {
-		pte = pfn_pte(pfn, PAGE_KERNEL_IO);
+		pte = pfn_pte(pfn, prot);
+		pte = pte_mkdirty(pte);

Is it necessary to dirty the pte?

It was dirtied before, so it definitely doesn't hurt,

Thanks.

#3 lore raw
Quoted 6 lines
Quoted 3 lines

From: Fangyu Yu <[email protected]>

Currently we use kvm_riscv_gstage_ioremap to map IMSIC gpa to the spa of

                                                                   ^^^
                                                                   hpa?

Yes, I think they mean the same thing, RISC-V IOMMU Spec defines spa (Supervisor Physical Address).

Quoted 9 lines
Quoted 5 lines

guest interrupt file within IMSIC.

The PAGE_KERNEL_IO property does not include user mode settings, so when accessing the IMSIC address in the virtual machine, a guest page fault will occur, this is not expected.

PAGE_KERNEL_IO also set the reserved G bit, so you're fixing two issues with a single change. :)

Right, The G bit in all G-stage PTEs is reserved for future standard use.

Quoted 7 lines
Quoted 3 lines

According to the RISC-V Privileged Architecture Spec, for G-stage address translation, all memory accesses are considered to be user-level accesses as though executed in Umode.

What implementation are you using? I would have assume that the original code was tested on QEMU, so we might have a bug there.

This issue can be reproduced using QEMU.
Since kvm has registered the MMIO Bus for IMSIC gpa, when a guest page fault occurs, it will call the imsic_mmio_write function,the guest irq will be written to the guest interrupt file by kvm.

Quoted 16 lines

arch/riscv/kvm/mmu.c int kvm_riscv_gstage_ioremap(struct kvm *kvm, gpa_t gpa, · line 367

Quoted 13 lines

---

arch/riscv/kvm/mmu.c +4 -1
@@ -359,8 +360,11 @@ int kvm_riscv_gstage_ioremap(struct kvm *kvm, gpa_t gpa,
 	end = (gpa + size + PAGE_SIZE - 1) & PAGE_MASK;
 	pfn = __phys_to_pfn(hpa);
 
+	prot = pgprot_noncached(PAGE_WRITE);
+
 	for (addr = gpa; addr < end; addr += PAGE_SIZE) {
-		pte = pfn_pte(pfn, PAGE_KERNEL_IO);
+		pte = pfn_pte(pfn, prot);
+		pte = pte_mkdirty(pte);

Is it necessary to dirty the pte?

Quoted 2 lines

It was dirtied before, so it definitely doesn't hurt,

Make pte dirty is necessary(for hardware without Svadu), and here is the first time to make this pte dirty.

Quoted 3 lines

Thanks.

Thanks,
fangyu

Signature · 3 lines
kvm-riscv mailing list
[email protected]
http://lists.infradead.org/mailman/listinfo/kvm-riscv
#4 lore raw
2025-08-09T11:20:20+08:00, <[email protected]>:
Quoted 5 lines
Quoted 5 lines
Quoted 4 lines

From: Fangyu Yu <[email protected]>
According to the RISC-V Privileged Architecture Spec, for G-stage address translation, all memory accesses are considered to be user-level accesses as though executed in Umode.

Quoted 8 lines
Quoted 3 lines

What implementation are you using? I would have assume that the original code was tested on QEMU, so we might have a bug there.

This issue can be reproduced using QEMU.
Since kvm has registered the MMIO Bus for IMSIC gpa, when a guest page fault occurs, it will call the imsic_mmio_write function,the guest irq will be written to the guest interrupt file by kvm.

Oh, so the interrupts were "just" slower. Great job catching that!

Quoted 11 lines
Quoted 8 lines

arch/riscv/kvm/mmu.c int kvm_riscv_gstage_ioremap(struct kvm *kvm, gpa_t gpa, · line 360

Quoted 4 lines

---

arch/riscv/kvm/mmu.c +1 -0
@@ -359,8 +360,11 @@ int kvm_riscv_gstage_ioremap(struct kvm *kvm, gpa_t gpa,
+		pte = pte_mkdirty(pte);

Is it necessary to dirty the pte?

It was dirtied before, so it definitely doesn't hurt,

Make pte dirty is necessary(for hardware without Svadu), and here is the first time to make this pte dirty.

Right, we would get a pointless trap otherwise,

Thanks.

maintainer
#5 lore raw
In reply to [email protected]
Quoted 10 lines · On Thu, Aug 7, 2025 at 12:37 PM <[email protected]> wrote:

From: Fangyu Yu <[email protected]>

Currently we use kvm_riscv_gstage_ioremap to map IMSIC gpa to the spa of guest interrupt file within IMSIC.

The PAGE_KERNEL_IO property does not include user mode settings, so when accessing the IMSIC address in the virtual machine, a guest page fault will occur, this is not expected.

Quoted 5 lines

According to the RISC-V Privileged Architecture Spec, for G-stage address translation, all memory accesses are considered to be user-level accesses as though executed in Umode.

Overall, a good fix. Thanks!

The patch subject and description needs improvements. Also, there is no Fixes tag which is required for backporting.

I have taken care of the above things at the time of merging this patch.

Queued this patch as fixes for Linux-6.17

Thanks,
Anup

Quoted 32 lines
---
 arch/riscv/kvm/mmu.c | 6 +++++-
 1 file changed, 5 insertions(+), 1 deletion(-)
arch/riscv/kvm/mmu.c +5 -2
index 1087ea74567b..800064e96ef6 100644
@@ -351,6 +351,7 @@ int kvm_riscv_gstage_ioremap(struct kvm *kvm, gpa_t gpa,
        int ret = 0;
        unsigned long pfn;
        phys_addr_t addr, end;
+       pgprot_t prot;
        struct kvm_mmu_memory_cache pcache = {
                .gfp_custom = (in_atomic) ? GFP_ATOMIC | __GFP_ACCOUNT : 0,
                .gfp_zero = __GFP_ZERO,
@@ -359,8 +360,11 @@ int kvm_riscv_gstage_ioremap(struct kvm *kvm, gpa_t gpa,
        end = (gpa + size + PAGE_SIZE - 1) & PAGE_MASK;
        pfn = __phys_to_pfn(hpa);
 
+       prot = pgprot_noncached(PAGE_WRITE);
+
        for (addr = gpa; addr < end; addr += PAGE_SIZE) {
-               pte = pfn_pte(pfn, PAGE_KERNEL_IO);
+               pte = pfn_pte(pfn, prot);
+               pte = pte_mkdirty(pte);
 
                if (!writable)
                        pte = pte_wrprotect(pte);
--

2.39.3 (Apple Git-146)

#6 lore raw
Quoted 21 lines
Quoted 10 lines

From: Fangyu Yu <[email protected]>

Currently we use kvm_riscv_gstage_ioremap to map IMSIC gpa to the spa of guest interrupt file within IMSIC.

The PAGE_KERNEL_IO property does not include user mode settings, so when accessing the IMSIC address in the virtual machine, a guest page fault will occur, this is not expected.

Quoted 5 lines

According to the RISC-V Privileged Architecture Spec, for G-stage address translation, all memory accesses are considered to be user-level accesses as though executed in Umode.

Overall, a good fix. Thanks!

The patch subject and description needs improvements. Also, there is no Fixes tag which is required for backporting.

Quoted 7 lines

I have taken care of the above things at the time of merging this patch.

Queued this patch as fixes for Linux-6.17

Thanks,
Anup

Thanks for your review.
I will send a v2 patch to fix these comments.

Thanks,
fangyu

Quoted 32 lines
Quoted 32 lines
---
 arch/riscv/kvm/mmu.c | 6 +++++-
 1 file changed, 5 insertions(+), 1 deletion(-)
arch/riscv/kvm/mmu.c +5 -2
index 1087ea74567b..800064e96ef6 100644
@@ -351,6 +351,7 @@ int kvm_riscv_gstage_ioremap(struct kvm *kvm, gpa_t gpa,
        int ret = 0;
        unsigned long pfn;
        phys_addr_t addr, end;
+       pgprot_t prot;
        struct kvm_mmu_memory_cache pcache = {
                .gfp_custom = (in_atomic) ? GFP_ATOMIC | __GFP_ACCOUNT : 0,
                .gfp_zero = __GFP_ZERO,
@@ -359,8 +360,11 @@ int kvm_riscv_gstage_ioremap(struct kvm *kvm, gpa_t gpa,
        end = (gpa + size + PAGE_SIZE - 1) & PAGE_MASK;
        pfn = __phys_to_pfn(hpa);
 
+       prot = pgprot_noncached(PAGE_WRITE);
+
        for (addr = gpa; addr < end; addr += PAGE_SIZE) {
-               pte = pfn_pte(pfn, PAGE_KERNEL_IO);
+               pte = pfn_pte(pfn, prot);
+               pte = pte_mkdirty(pte);
 
                if (!writable)
                        pte = pte_wrprotect(pte);
--

2.39.3 (Apple Git-146)

#7 lore raw
Quoted 24 lines · On Tue, Aug 19, 2025 at 6:56 AM <[email protected]> wrote:
Quoted 24 lines
Quoted 11 lines

From: Fangyu Yu <[email protected]>

Currently we use kvm_riscv_gstage_ioremap to map IMSIC gpa to the spa of guest interrupt file within IMSIC.

The PAGE_KERNEL_IO property does not include user mode settings, so when accessing the IMSIC address in the virtual machine, a guest page fault will occur, this is not expected.

Quoted 5 lines

According to the RISC-V Privileged Architecture Spec, for G-stage address translation, all memory accesses are considered to be user-level accesses as though executed in Umode.

Overall, a good fix. Thanks!

The patch subject and description needs improvements. Also, there is no Fixes tag which is required for backporting.

I have taken care of the above things at the time of merging this patch.

Quoted 8 lines
Quoted 5 lines

Queued this patch as fixes for Linux-6.17

Thanks,
Anup

Thanks for your review.
I will send a v2 patch to fix these comments.

No need, it's already part of my riscv_kvm_fixes branch at:
https//github.com/kvm-riscv/linux.git

Regards,
Anup