Re: [PATCH v2 12/16] cpufreq/amd-pstate: Always write EPP value when updating perf
From: Mario Limonciello
Date: Mon Dec 09 2024 - 12:16:21 EST
On 12/9/2024 10:49, Mario Limonciello wrote:
On 12/9/2024 02:42, Gautham R. Shenoy wrote:
Hello Mario,
On Sun, Dec 08, 2024 at 12:30:27AM -0600, Mario Limonciello wrote:
For MSR systems the EPP value is in the same register as perf targets
and so divding them into two separate MSR writes is wasteful.
In msr_update_perf(), update both EPP and perf values in one write to
MSR_AMD_CPPC_REQ, and cache them if successful.
To accomplish this plumb the EPP value into the update_perf call and
modify
all its callers to check the return value.
Reviewed-and-tested-by: Dhananjay Ugwekar <dhananjay.ugwekar@xxxxxxx>
Signed-off-by: Mario Limonciello <mario.limonciello@xxxxxxx>
---
drivers/cpufreq/amd-pstate.c | 71 ++++++++++++++++++++++--------------
1 file changed, 43 insertions(+), 28 deletions(-)
diff --git a/drivers/cpufreq/amd-pstate.c b/drivers/cpufreq/amd-pstate.c
index d21acd961edcd..dd11ba6c00cc3 100644
--- a/drivers/cpufreq/amd-pstate.c
+++ b/drivers/cpufreq/amd-pstate.c
@@ -222,25 +222,36 @@ static s16 shmem_get_epp(struct amd_cpudata
*cpudata)
}
static int msr_update_perf(struct amd_cpudata *cpudata, u32 min_perf,
- u32 des_perf, u32 max_perf, bool fast_switch)
+ u32 des_perf, u32 max_perf, u32 epp, bool fast_switch)
{
+ u64 value;
+
+ value = READ_ONCE(cpudata->cppc_req_cached);
There seems to be a mismatch here between what the API is passing and
parameters and how this function is *not* using them, and instead
using cpudata->cppc_req_cached.
The expectation seems to be that the max_perf, min_perf, des_perf and
epp fields in cpudata->cppc_req_cached would be the same as @des_perf,
@max_perf, @min_perf and @ep, no ?
Or is it that for the MSR update, the value in
cpudata->cppc_req_cached take precedence over the arguments passed ?
Ideally, the "value" should be recomputed here using (@min_perf |
@max_perf | @des_perf | @epp) and that value should be cached as you
are doing below.
Yeah - that's what the next patch does (which I think you probably saw
after you reviewed it).
Do you think maybe I should just squash the two? Or would you be
happier if I re-ordered the two?
FYI - I looked into re-ordering and it's not feasible because you need
EPP plumbed in order to validate the result.
So I'm going to squash the two patches, and I'll do another one that
adjusts tracing locations for your other feedback.
if (fast_switch) {
wrmsrl(MSR_AMD_CPPC_REQ, READ_ONCE(cpudata->cppc_req_cached));
return 0;
+ } else {
+ int ret = wrmsrl_on_cpu(cpudata->cpu, MSR_AMD_CPPC_REQ,
+ READ_ONCE(cpudata->cppc_req_cached));
+ if (ret)
+ return ret;
}
- return wrmsrl_on_cpu(cpudata->cpu, MSR_AMD_CPPC_REQ,
- READ_ONCE(cpudata->cppc_req_cached));
+ WRITE_ONCE(cpudata->cppc_req_cached, value);
Since cppc_req_cached is not changed, why write it again ?
Because of the next patch. It will look at cpudata->cppc_req_cached and
determine if anything changed in it - including EPP.
+ WRITE_ONCE(cpudata->epp_cached, epp);
+
+ return 0;
}
DEFINE_STATIC_CALL(amd_pstate_update_perf, msr_update_perf);
static inline int amd_pstate_update_perf(struct amd_cpudata *cpudata,
u32 min_perf, u32 des_perf,
- u32 max_perf, bool fast_switch)
+ u32 max_perf, u32 epp,
+ bool fast_switch)
{
return static_call(amd_pstate_update_perf)(cpudata, min_perf,
des_perf,
- max_perf, fast_switch);
+ max_perf, epp, fast_switch);
}
static int msr_set_epp(struct amd_cpudata *cpudata, u32 epp)
@@ -459,12 +470,19 @@ static inline int amd_pstate_init_perf(struct
amd_cpudata *cpudata)
return static_call(amd_pstate_init_perf)(cpudata);
}
-static int shmem_update_perf(struct amd_cpudata *cpudata,
- u32 min_perf, u32 des_perf,
- u32 max_perf, bool fast_switch)
+static int shmem_update_perf(struct amd_cpudata *cpudata, u32 min_perf,
+ u32 des_perf, u32 max_perf, u32 epp, bool fast_switch)
{
struct cppc_perf_ctrls perf_ctrls;
+ if (cppc_state == AMD_PSTATE_ACTIVE) {
+ int ret = shmem_set_epp(cpudata, epp);
+
+ if (ret)
+ return ret;
+ WRITE_ONCE(cpudata->epp_cached, epp);
+ }
+
perf_ctrls.max_perf = max_perf;
perf_ctrls.min_perf = min_perf;
perf_ctrls.desired_perf = des_perf;
@@ -545,10 +563,10 @@ static void amd_pstate_update(struct
amd_cpudata *cpudata, u32 min_perf,
WRITE_ONCE(cpudata->cppc_req_cached, value);
- amd_pstate_update_perf(cpudata, min_perf, des_perf,
- max_perf, fast_switch);
+ amd_pstate_update_perf(cpudata, min_perf, des_perf, max_perf, 0,
fast_switch);
cpufreq_policy_put:
+
cpufreq_cpu_put(policy);
}
@@ -1545,6 +1563,7 @@ static int amd_pstate_epp_update_limit(struct
cpufreq_policy *policy)
{
struct amd_cpudata *cpudata = policy->driver_data;
u64 value;
+ u32 epp;
amd_pstate_update_min_max_limit(policy);
@@ -1557,23 +1576,19 @@ static int amd_pstate_epp_update_limit(struct
cpufreq_policy *policy)
value |= FIELD_PREP(AMD_CPPC_MIN_PERF_MASK, cpudata-
>min_limit_perf);
if (cpudata->policy == CPUFREQ_POLICY_PERFORMANCE)
- WRITE_ONCE(cpudata->epp_cached, 0);
- value |= FIELD_PREP(AMD_CPPC_EPP_PERF_MASK, cpudata->epp_cached);
-
- WRITE_ONCE(cpudata->cppc_req_cached, value);
+ epp = 0;
+ else
+ epp = READ_ONCE(cpudata->epp_cached);
if (trace_amd_pstate_epp_perf_enabled()) {
- trace_amd_pstate_epp_perf(cpudata->cpu, cpudata->highest_perf,
- cpudata->epp_cached,
+ trace_amd_pstate_epp_perf(cpudata->cpu, cpudata-
>highest_perf, epp,
cpudata->min_limit_perf,
cpudata->max_limit_perf,
policy->boost_enabled);
}
- amd_pstate_update_perf(cpudata, cpudata->min_limit_perf, 0U,
- cpudata->max_limit_perf, false);
-
- return amd_pstate_set_epp(cpudata, READ_ONCE(cpudata->epp_cached));
+ return amd_pstate_update_perf(cpudata, cpudata->min_limit_perf, 0U,
+ cpudata->max_limit_perf, epp, false);
}
static int amd_pstate_epp_set_policy(struct cpufreq_policy *policy)
@@ -1602,7 +1617,7 @@ static int amd_pstate_epp_set_policy(struct
cpufreq_policy *policy)
return 0;
}
-static void amd_pstate_epp_reenable(struct amd_cpudata *cpudata)
+static int amd_pstate_epp_reenable(struct amd_cpudata *cpudata)
{
u64 max_perf;
int ret;
@@ -1620,17 +1635,19 @@ static void amd_pstate_epp_reenable(struct
amd_cpudata *cpudata)
max_perf, cpudata->boost_state);
}
- amd_pstate_update_perf(cpudata, 0, 0, max_perf, false);
- amd_pstate_set_epp(cpudata, cpudata->epp_cached);
+ return amd_pstate_update_perf(cpudata, 0, 0, max_perf, cpudata-
>epp_cached, false);
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
On an MSR based system, none of the values passed here will be used,
and instead the value in cpudata->cppc_req_cached will be used, no?
Currently; yes. After the next patch that changes.
}
static int amd_pstate_epp_cpu_online(struct cpufreq_policy *policy)
{
struct amd_cpudata *cpudata = policy->driver_data;
+ int ret;
pr_debug("AMD CPU Core %d going online\n", cpudata->cpu);
- amd_pstate_epp_reenable(cpudata);
+ ret = amd_pstate_epp_reenable(cpudata);
+ if (ret)
+ return ret;
cpudata->suspended = false;
return 0;
@@ -1654,10 +1671,8 @@ static int amd_pstate_epp_cpu_offline(struct
cpufreq_policy *policy)
min_perf, min_perf, policy->boost_enabled);
}
- amd_pstate_update_perf(cpudata, min_perf, 0, min_perf, false);
- amd_pstate_set_epp(cpudata, AMD_CPPC_EPP_BALANCE_POWERSAVE);
-
- return 0;
+ return amd_pstate_update_perf(cpudata, min_perf, 0, min_perf,
+ AMD_CPPC_EPP_BALANCE_POWERSAVE, false);
}
static int amd_pstate_epp_suspend(struct cpufreq_policy *policy)
--
2.43.0
--
Thanks and Regards
gautham.