From: Ashish Kalra rmpopt_scan_range() walked a single global [rmpopt_pa_start, rmpopt_pa_end] range, which is only correct while every core shares one window. Read the executing core's own RMPOPT_BASE instead and walk only that window, so each core optimizes the window its own RMPOPT_BASE points at. The window is described entirely by the RMPOPT_BASE MSR: RMPOPT_BASE_ADDR (bits 51:30) is the 1 GB-aligned base and RMPOPT_TABLE_SIZE (bits 22:1) is the window size in GB. Walk [base, base + size), clamped to the top of populated memory. Signed-off-by: Ashish Kalra --- arch/x86/include/asm/msr-index.h | 2 ++ arch/x86/virt/svm/sev.c | 29 +++++++++++++++-------------- 2 files changed, 17 insertions(+), 14 deletions(-) diff --git a/arch/x86/include/asm/msr-index.h b/arch/x86/include/asm/msr-index.h index c0914829a925..6d453bb13621 100644 --- a/arch/x86/include/asm/msr-index.h +++ b/arch/x86/include/asm/msr-index.h @@ -773,6 +773,8 @@ #define MSR_AMD64_RMPOPT_BASE 0xc0010139 #define MSR_AMD64_RMPOPT_ENABLE_BIT 0 #define MSR_AMD64_RMPOPT_ENABLE BIT_ULL(MSR_AMD64_RMPOPT_ENABLE_BIT) +#define MSR_AMD64_RMPOPT_TABLE_SIZE GENMASK_ULL(22, 1) +#define MSR_AMD64_RMPOPT_BASE_ADDR GENMASK_ULL(51, 30) #define MSR_SVSM_CAA 0xc001f000 diff --git a/arch/x86/virt/svm/sev.c b/arch/x86/virt/svm/sev.c index 4de9681cc2f8..a2756b18d5f7 100644 --- a/arch/x86/virt/svm/sev.c +++ b/arch/x86/virt/svm/sev.c @@ -8,6 +8,7 @@ * */ +#include #include #include #include @@ -125,8 +126,6 @@ static void *rmp_bookkeeping __ro_after_init; static u64 probed_rmp_base, probed_rmp_size; -static u64 rmpopt_pa_start, rmpopt_pa_end; - enum rmpopt_op_type { RMPOPT_OP_VERIFY_AND_REPORT_STATUS, RMPOPT_OP_REPORT_STATUS @@ -618,11 +617,18 @@ static void rmpopt(u64 pa) : "memory", "cc"); } +/* A core can only RMPOPT within its own window, [base, base + size). */ static void rmpopt_scan_range(void *arg) { - u64 pa; + u64 msr, base, end, pa; + unsigned int size_gb; + + rdmsrq(MSR_AMD64_RMPOPT_BASE, msr); + base = msr & MSR_AMD64_RMPOPT_BASE_ADDR; + size_gb = FIELD_GET(MSR_AMD64_RMPOPT_TABLE_SIZE, msr); + end = min_t(u64, base + (u64)size_gb * SZ_1G, ALIGN(PFN_PHYS(max_pfn), SZ_1G)); - for (pa = rmpopt_pa_start; pa < rmpopt_pa_end; pa += SZ_1G) + for (pa = base; pa < end; pa += SZ_1G) rmpopt(pa); } @@ -650,12 +656,6 @@ static int __init rmpopt_init(void) INIT_DELAYED_WORK(&rmpopt_delayed_work, do_rmpopt_work); - /* The optimization range is fixed at boot; compute it once. */ - rmpopt_pa_start = ALIGN_DOWN(PFN_PHYS(min_low_pfn), SZ_1G); - rmpopt_pa_end = ALIGN(PFN_PHYS(max_pfn), SZ_1G); - if ((rmpopt_pa_end - rmpopt_pa_start) > SZ_2T) - rmpopt_pa_end = rmpopt_pa_start + SZ_2T; - pr_info("RMPOPT optimizations enabled\n"); rmpopt_soft_init = true; @@ -665,7 +665,7 @@ static int __init rmpopt_init(void) void snp_enable_rmpopt(void) { - u64 base; + u64 base, pa_start; int cpu; if (!cc_platform_has(CC_ATTR_HOST_SEV_SNP)) @@ -678,14 +678,15 @@ void snp_enable_rmpopt(void) return; /* - * Per-CPU RMPOPT tables cover at most 2 TB. Program each core's - * RMPOPT_BASE with the start of RAM to optimize up to 2 TB. + * A core's RMPOPT_BASE covers at most 2 TB. Program every core with + * the start of RAM to optimize up to 2 TB. */ + pa_start = ALIGN_DOWN(PFN_PHYS(min_low_pfn), SZ_1G); rdmsrq(MSR_AMD64_RMPOPT_BASE, base); if (!(base & MSR_AMD64_RMPOPT_ENABLE)) for_each_cpu(cpu, cpu_primary_thread_mask) wrmsrq_on_cpu(cpu, MSR_AMD64_RMPOPT_BASE, - rmpopt_pa_start | MSR_AMD64_RMPOPT_ENABLE); + pa_start | MSR_AMD64_RMPOPT_ENABLE); mod_delayed_work(rmpopt_wq, &rmpopt_delayed_work, msecs_to_jiffies(RMPOPT_WORK_TIMEOUT)); -- 2.43.0