Re: [PATCH v14 09/13] tsc: Use the GUEST_TSC_FREQ MSR for discovering TSC frequency

From: Xiaoyao Li
Date: Tue Oct 29 2024 - 05:15:28 EST


On 10/29/2024 11:56 AM, Nikunj A. Dadhania wrote:


On 10/29/2024 8:32 AM, Xiaoyao Li wrote:
On 10/28/2024 1:34 PM, Nikunj A Dadhania wrote:
Calibrating the TSC frequency using the kvmclock is not correct for
SecureTSC enabled guests. Use the platform provided TSC frequency via the
GUEST_TSC_FREQ MSR (C001_0134h).

Signed-off-by: Nikunj A Dadhania <nikunj@xxxxxxx>
---
  arch/x86/include/asm/sev.h |  2 ++
  arch/x86/coco/sev/core.c   | 16 ++++++++++++++++
  arch/x86/kernel/tsc.c      |  5 +++++
  3 files changed, 23 insertions(+)

diff --git a/arch/x86/include/asm/sev.h b/arch/x86/include/asm/sev.h
index d27c4e0f9f57..9ee63ddd0d90 100644
--- a/arch/x86/include/asm/sev.h
+++ b/arch/x86/include/asm/sev.h
@@ -536,6 +536,7 @@ static inline int handle_guest_request(struct snp_msg_desc *mdesc, u64 exit_code
  }
    void __init snp_secure_tsc_prepare(void);
+void __init snp_secure_tsc_init(void);
    #else    /* !CONFIG_AMD_MEM_ENCRYPT */
  @@ -584,6 +585,7 @@ static inline int handle_guest_request(struct snp_msg_desc *mdesc, u64 exit_code
                         u32 resp_sz) { return -ENODEV; }
    static inline void __init snp_secure_tsc_prepare(void) { }
+static inline void __init snp_secure_tsc_init(void) { }
    #endif    /* CONFIG_AMD_MEM_ENCRYPT */
  diff --git a/arch/x86/coco/sev/core.c b/arch/x86/coco/sev/core.c
index 140759fafe0c..0be9496b8dea 100644
--- a/arch/x86/coco/sev/core.c
+++ b/arch/x86/coco/sev/core.c
@@ -3064,3 +3064,19 @@ void __init snp_secure_tsc_prepare(void)
        pr_debug("SecureTSC enabled");
  }
+
+static unsigned long securetsc_get_tsc_khz(void)
+{
+    unsigned long long tsc_freq_mhz;
+
+    setup_force_cpu_cap(X86_FEATURE_TSC_KNOWN_FREQ);
+    rdmsrl(MSR_AMD64_GUEST_TSC_FREQ, tsc_freq_mhz);
+
+    return (unsigned long)(tsc_freq_mhz * 1000);
+}
+
+void __init snp_secure_tsc_init(void)
+{
+    x86_platform.calibrate_cpu = securetsc_get_tsc_khz;
+    x86_platform.calibrate_tsc = securetsc_get_tsc_khz;
+}
diff --git a/arch/x86/kernel/tsc.c b/arch/x86/kernel/tsc.c
index dfe6847fd99e..730cbbd4554e 100644
--- a/arch/x86/kernel/tsc.c
+++ b/arch/x86/kernel/tsc.c
@@ -30,6 +30,7 @@
  #include <asm/i8259.h>
  #include <asm/topology.h>
  #include <asm/uv/uv.h>
+#include <asm/sev.h>
    unsigned int __read_mostly cpu_khz;    /* TSC clocks / usec, not used here */
  EXPORT_SYMBOL(cpu_khz);
@@ -1514,6 +1515,10 @@ void __init tsc_early_init(void)
      /* Don't change UV TSC multi-chassis synchronization */
      if (is_early_uv_system())
          return;
+
+    if (cc_platform_has(CC_ATTR_GUEST_SNP_SECURE_TSC))
+        snp_secure_tsc_init();

IMHO, it isn't the good place to call snp_secure_tsc_init() to update the callbacks here.

It's better to be called in some snp init functions.

As part of setup_arch(), init_hypervisor_platform() gets called and all the PV clocks
are registered and initialized as part of init_platform callback. Once the hypervisor
platform is initialized, tsc_early_init() is called. SEV SNP guest can be running on
any hypervisor, so the call back needs to be updated either in tsc_early_init() or
init_hypervisor_platform(), as the change is TSC related, I have updated it here.

I think it might be due to

1. it lacks a central place for SNP related stuff, like tdx_early_init()
2. even we have some place of 1), the callbacks will be overwrote in init_hypervisor_platform() by specific PV ops.

However, I don't think it's good practice to update it tsc.c. The reason why callback is used is that arch/hypervisor specific code can implement
and overwrite with it's own implementation in its own file.

Back to your case, I think a central snp init function would be helpful, and we can introduce a new flag to skip the overwrite of tsc/cpu calibration for hypervisor when the flag is set.



      if (!determine_cpu_tsc_frequencies(true))
          return;
      tsc_enable_sched_clock();


Regards,
Nikunj