Re: [PATCH Resend 2/8] tools/power x86_energy_perf_policy: enable AMD pstate EPP profile switching on MSR based system
From: Mario Limonciello
Date: Tue Jan 30 2024 - 15:01:27 EST
On 1/29/2024 23:56, Perry Yuan wrote:
From: Perry Yuan <Perry.Yuan@xxxxxxx>
x86_energy_perf_policy adds AMD EPP profiles switch support,the EPP
driver also has four EPP profiles implemented under power
firmware autonomous mode.
(performance, balance-performance, balance-power, power)
User can choose the new profile to use with this utility support while
`amd_pstate_epp` driver instance loaded for the MSR support systems.
for example:
$ sudo x86_energy_perf_policy -d --hwp-epp balance-performance
cpu0: old: HWP_REQ: min 13 max 166 des 0 epp 128 window 0x9 (9*10^0us) use_pkg 0
cpu0: HWP_CAP: low 13 eff 157 guar 0 high 231
cpu0: new: HWP_REQ: min 13 max 166 des 0 epp 128 window 0x9 (9*10^0us) use_pkg 0
cpu1: old: HWP_REQ: min 13 max 166 des 0 epp 128 window 0x9 (9*10^0us) use_pkg 0
cpu1: HWP_CAP: low 13 eff 157 guar 0 high 236
cpu1: new: HWP_REQ: min 13 max 166 des 0 epp 128 window 0x9 (9*10^0us) use_pkg 0
cpu2: old: HWP_REQ: min 13 max 166 des 0 epp 128 window 0x9 (9*10^0us) use_pkg 0
cpu2: HWP_CAP: low 13 eff 157 guar 0 high 211
cpu2: new: HWP_REQ: min 13 max 166 des 0 epp 128 window 0x9 (9*10^0us) use_pkg 0
cpu3: old: HWP_REQ: min 13 max 166 des 0 epp 128 window 0x9 (9*10^0us) use_pkg 0
cpu3: HWP_CAP: low 13 eff 157 guar 0 high 206
cpu3: new: HWP_REQ: min 13 max 166 des 0 epp 128 window 0x9 (9*10^0us) use_pkg 0
Signed-off-by: Perry Yuan <Perry.Yuan@xxxxxxx>
---
.../x86_energy_perf_policy.c | 69 +++++++++++++++----
1 file changed, 54 insertions(+), 15 deletions(-)
diff --git a/tools/power/x86/x86_energy_perf_policy/x86_energy_perf_policy.c b/tools/power/x86/x86_energy_perf_policy/x86_energy_perf_policy.c
index 5daf1c2bb601..da9087873915 100644
--- a/tools/power/x86/x86_energy_perf_policy/x86_energy_perf_policy.c
+++ b/tools/power/x86/x86_energy_perf_policy/x86_energy_perf_policy.c
@@ -902,14 +902,23 @@ void write_hwp_request(int cpu, struct msr_hwp_request *hwp_req, unsigned int ms
hwp_req->hwp_desired, hwp_req->hwp_epp,
hwp_req->hwp_window, hwp_req->hwp_use_pkg);
- msr |= HWP_MIN_PERF(ratio_2_msr_perf(hwp_req->hwp_min));
- msr |= HWP_MAX_PERF(ratio_2_msr_perf(hwp_req->hwp_max));
- msr |= HWP_DESIRED_PERF(ratio_2_msr_perf(hwp_req->hwp_desired));
- msr |= HWP_ENERGY_PERF_PREFERENCE(hwp_req->hwp_epp);
- msr |= HWP_ACTIVITY_WINDOW(hwp_req->hwp_window);
- msr |= HWP_PACKAGE_CONTROL(hwp_req->hwp_use_pkg);
-
- put_msr(cpu, msr_offset, msr);
+ if (genuine_intel) {
+ msr |= HWP_MIN_PERF(ratio_2_msr_perf(hwp_req->hwp_min));
+ msr |= HWP_MAX_PERF(ratio_2_msr_perf(hwp_req->hwp_max));
+ msr |= HWP_DESIRED_PERF(ratio_2_msr_perf(hwp_req->hwp_desired));
+ msr |= HWP_ENERGY_PERF_PREFERENCE(hwp_req->hwp_epp);
+ msr |= HWP_ACTIVITY_WINDOW(hwp_req->hwp_window);
+ msr |= HWP_PACKAGE_CONTROL(hwp_req->hwp_use_pkg);
+ put_msr(cpu, msr_offset, msr);
+ } else if (authentic_amd) {
+ /* AMD EPP need to set desired perf with zero */
+ hwp_req->hwp_desired = 0;
Should this be blocked a level higher up? Like exiting the tool the
function call when calling parse_cmdline_hwp_desired() is used?
+ msr |= AMD_CPPC_MIN_PERF(hwp_req->hwp_min);
+ msr |= AMD_CPPC_MAX_PERF(hwp_req->hwp_max);
+ msr |= AMD_CPPC_DES_PERF(hwp_req->hwp_desired);
Given the msr is already initialized to 0, this is unnecessary isn't it?
+ msr |= AMD_CPPC_ENERGY_PERF_PREF(hwp_req->hwp_epp);
+ amd_put_msr(cpu, msr_offset, (unsigned int)msr);
+ }
}
static int get_epb(int cpu)
@@ -1157,8 +1166,12 @@ int update_hwp_request(int cpu)
{
struct msr_hwp_request req;
struct msr_hwp_cap cap;
+ int msr_offset = 0;
- int msr_offset = MSR_HWP_REQUEST;
+ if (genuine_intel)
+ msr_offset = MSR_HWP_REQUEST;
+ else if (authentic_amd)
+ msr_offset = MSR_AMD_CPPC_REQ;
read_hwp_request(cpu, &req, msr_offset);
if (debug)
@@ -1181,7 +1194,11 @@ int update_hwp_request(int cpu)
req.hwp_use_pkg = req_update.hwp_use_pkg;
- read_hwp_cap(cpu, &cap, MSR_HWP_CAPABILITIES);
+ if (genuine_intel)
+ read_hwp_cap(cpu, &cap, MSR_HWP_CAPABILITIES);
+ else if (authentic_amd)
+ read_hwp_cap(cpu, &cap, MSR_AMD_CPPC_CAP1);
+
if (debug)
print_hwp_cap(cpu, &cap, "");
@@ -1203,8 +1220,12 @@ int update_hwp_request_pkg(int pkg)
struct msr_hwp_request req;
struct msr_hwp_cap cap;
int cpu = first_cpu_in_pkg[pkg];
+ int msr_offset = 0;
- int msr_offset = MSR_HWP_REQUEST_PKG;
+ if (genuine_intel)
+ msr_offset = MSR_HWP_REQUEST_PKG;
+ else if (authentic_amd)
+ msr_offset = MSR_AMD_CPPC_REQ;
read_hwp_request(cpu, &req, msr_offset);
if (debug)
@@ -1225,7 +1246,11 @@ int update_hwp_request_pkg(int pkg)
if (update_hwp_epp)
req.hwp_epp = req_update.hwp_epp;
- read_hwp_cap(cpu, &cap, MSR_HWP_CAPABILITIES);
+ if (genuine_intel)
+ read_hwp_cap(cpu, &cap, MSR_HWP_CAPABILITIES);
+ else if (authentic_amd)
+ read_hwp_cap(cpu, &cap, MSR_AMD_CPPC_CAP1);
+
if (debug)
print_hwp_cap(cpu, &cap, "");
@@ -1246,13 +1271,27 @@ int update_hwp_request_pkg(int pkg)
int enable_hwp_on_cpu(int cpu)
{
unsigned long long msr;
+ int ret;
+
+ if (genuine_intel) {
+ get_msr(cpu, MSR_PM_ENABLE, &msr);
+ put_msr(cpu, MSR_PM_ENABLE, 1);
+ } else if (authentic_amd) {
+ ret = amd_get_msr(cpu, MSR_AMD_CPPC_ENABLE, (unsigned long *)(&msr));
+ if (ret < 0)
+ errx(-1, "failed to get msr with return %d", ret);
- get_msr(cpu, MSR_PM_ENABLE, &msr);
- put_msr(cpu, MSR_PM_ENABLE, 1);
+ ret = amd_put_msr(cpu, MSR_AMD_CPPC_ENABLE, 1);
+ if (ret < 0)
+ errx(-1, "failed to put msr with return %d", ret);
+ }
- if (verbose)
+ if (verbose && genuine_intel)
printf("cpu%d: MSR_PM_ENABLE old: %d new: %d\n", cpu, (unsigned int) msr, 1);
+ if (verbose && authentic_amd)
+ printf("cpu%d: MSR_AMD_CPPC_ENABLE old: %d new: %d\n", cpu, (unsigned int) msr, 1);
+
return 0;
}