Skip to contentSkip to conversation

RISC-V: KVM: Avoid redundant page-table allocations in ioremap topup

PATCH In review [email protected] 2 participants 2 messages Jun 10 – Jul 13

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

PATCH 1/1 RISC-V: KVM: Avoid redundant page-table allocations in ioremap topup

From: Fangyu Yu <[email protected]>

kvm_riscv_mmu_ioremap() currently tops up its on-stack page-table cache via kvm_mmu_topup_memory_cache(), which allocates up to KVM_ARCH_NR_OBJS_PER_MEMORY_CACHE (32) objects per topup.

ioremap only consumes non-leaf page-table pages, at most pgd_levels - 1 (1 to 4) per call, and for contiguous mappings within the same huge page the non-leaf pages are allocated once and reused by subsequent pages. Topping up to 32 objects therefore triggers many unnecessary GFP_KERNEL_ACCOUNT allocations on every call, all of which are freed when the function returns. In hot paths (such as vCPU migration), this creates avoidable allocator churn and wastes CPU cycles.

Use __kvm_mmu_topup_memory_cache() with a capacity of pgd_levels so the on-stack cache is sized to the maximum demand of a single mapping. This removes the redundant allocations and reduces per-call overhead without changing behavior.

---
 arch/riscv/kvm/mmu.c | 3 ++-
 1 file changed, 2 insertions(+), 1 deletion(-)
arch/riscv/kvm/mmu.c +2 -1
index 2d3def024270..873a5dde4b92 100644
@@ -39,6 +39,7 @@ int kvm_riscv_mmu_ioremap(struct kvm *kvm, gpa_t gpa, phys_addr_t hpa,
 	pgprot_t prot;
 	unsigned long pfn;
 	phys_addr_t addr, end;
+	unsigned long pgd_levels = kvm->arch.pgd_levels;
 	struct kvm_mmu_memory_cache pcache = {
 		.gfp_custom = (in_atomic) ? GFP_ATOMIC | __GFP_ACCOUNT : 0,
 		.gfp_zero = __GFP_ZERO,
@@ -61,7 +62,7 @@ int kvm_riscv_mmu_ioremap(struct kvm *kvm, gpa_t gpa, phys_addr_t hpa,
 		if (!writable)
 			map.pte = pte_wrprotect(map.pte);
 
-		ret = kvm_mmu_topup_memory_cache(&pcache, kvm->arch.pgd_levels);
+		ret = __kvm_mmu_topup_memory_cache(&pcache, pgd_levels, pgd_levels);
 		if (ret)
 			goto out;
 
Signature · 1 line
2.50.1
maintainer Reviewed-by on v1 1/1
#1 lore raw
Quoted 16 lines · On Wed, Jun 10, 2026 at 3:09 PM <[email protected]> wrote:

From: Fangyu Yu <[email protected]>

kvm_riscv_mmu_ioremap() currently tops up its on-stack page-table cache via kvm_mmu_topup_memory_cache(), which allocates up to KVM_ARCH_NR_OBJS_PER_MEMORY_CACHE (32) objects per topup.

ioremap only consumes non-leaf page-table pages, at most pgd_levels - 1 (1 to 4) per call, and for contiguous mappings within the same huge page the non-leaf pages are allocated once and reused by subsequent pages. Topping up to 32 objects therefore triggers many unnecessary GFP_KERNEL_ACCOUNT allocations on every call, all of which are freed when the function returns. In hot paths (such as vCPU migration), this creates avoidable allocator churn and wastes CPU cycles.

Quoted 6 lines

Use __kvm_mmu_topup_memory_cache() with a capacity of pgd_levels so the on-stack cache is sized to the maximum demand of a single mapping. This removes the redundant allocations and reduces per-call overhead without changing behavior.

LGTM.

Queued this patch as a fix for Linux-7.2-rcX

Thanks,
Anup

Quoted 28 lines
---
 arch/riscv/kvm/mmu.c | 3 ++-
 1 file changed, 2 insertions(+), 1 deletion(-)
arch/riscv/kvm/mmu.c +2 -2
index 2d3def024270..873a5dde4b92 100644
@@ -39,6 +39,7 @@ int kvm_riscv_mmu_ioremap(struct kvm *kvm, gpa_t gpa, phys_addr_t hpa,
        pgprot_t prot;
        unsigned long pfn;
        phys_addr_t addr, end;
+       unsigned long pgd_levels = kvm->arch.pgd_levels;
        struct kvm_mmu_memory_cache pcache = {
                .gfp_custom = (in_atomic) ? GFP_ATOMIC | __GFP_ACCOUNT : 0,
                .gfp_zero = __GFP_ZERO,
@@ -61,7 +62,7 @@ int kvm_riscv_mmu_ioremap(struct kvm *kvm, gpa_t gpa, phys_addr_t hpa,
                if (!writable)
                        map.pte = pte_wrprotect(map.pte);
 
-               ret = kvm_mmu_topup_memory_cache(&pcache, kvm->arch.pgd_levels);
+               ret = __kvm_mmu_topup_memory_cache(&pcache, pgd_levels, pgd_levels);
                if (ret)
                        goto out;
 
--

2.50.1