arm-trusted-firmware/plat/common/plat_psci_common.c
Boyan Karatotev abf6666e26 perf(psci): get PMF timestamps with no cache flushes if possible
Whenever we have HW_ASSISTED_COHERENCY, caches are enabled early and we
let the cores do the cache maintenance on our behalf. This is true for
the PSCI stat timestamp capture and used to be the case. However, a
model bug required us to do the cache maintenance manually. That has
been fixed so we can revert back.

Change-Id: Id315a8fea500fb5e2433d3786b2be5a9084300a7
Signed-off-by: Boyan Karatotev <boyan.karatotev@arm.com>
2025-02-25 08:51:24 +00:00

179 lines
4.9 KiB
C

/*
* Copyright (c) 2016-2025, Arm Limited and Contributors. All rights reserved.
* Copyright (c) 2020, NVIDIA Corporation. All rights reserved.
*
* SPDX-License-Identifier: BSD-3-Clause
*/
#include <assert.h>
#include <arch.h>
#include <lib/pmf/pmf.h>
#include <lib/psci/psci.h>
#include <lib/utils_def.h>
#include <plat/common/platform.h>
#if ENABLE_PSCI_STAT && ENABLE_PMF
#pragma weak plat_psci_stat_accounting_start
#pragma weak plat_psci_stat_accounting_stop
#pragma weak plat_psci_stat_get_residency
/* Maximum time-stamp value read from architectural counters */
#ifdef __aarch64__
#define MAX_TS UINT64_MAX
#else
#define MAX_TS UINT32_MAX
#endif
/* Following are used as ID's to capture time-stamp */
#define PSCI_STAT_ID_ENTER_LOW_PWR 0
#define PSCI_STAT_ID_EXIT_LOW_PWR 1
#define PSCI_STAT_TOTAL_IDS 2
#if HW_ASSISTED_COHERENCY
#define CACHE_MAINTENANCE_ATTR PMF_NO_CACHE_MAINT
#else
#define CACHE_MAINTENANCE_ATTR PMF_CACHE_MAINT
#endif
PMF_DECLARE_CAPTURE_TIMESTAMP(psci_svc)
PMF_DECLARE_GET_TIMESTAMP(psci_svc)
PMF_REGISTER_SERVICE(psci_svc, PMF_PSCI_STAT_SVC_ID, PSCI_STAT_TOTAL_IDS,
PMF_STORE_ENABLE)
/*
* This function calculates the stats residency in microseconds,
* taking in account the wrap around condition.
*/
static u_register_t calc_stat_residency(unsigned long long pwrupts,
unsigned long long pwrdnts)
{
/* The divisor to use to convert raw timestamp into microseconds. */
u_register_t residency_div;
u_register_t res;
/*
* Calculate divisor so that it can be directly used to
* convert time-stamp into microseconds.
*/
residency_div = read_cntfrq_el0() / MHZ_TICKS_PER_SEC;
assert(residency_div > 0U);
if (pwrupts < pwrdnts)
res = MAX_TS - pwrdnts + pwrupts;
else
res = pwrupts - pwrdnts;
return res / residency_div;
}
/*
* Capture timestamp before entering a low power state.
* Cache maintenance may be needed when reading these timestamps.
*/
void plat_psci_stat_accounting_start(
__unused const psci_power_state_t *state_info)
{
assert(state_info != NULL);
PMF_CAPTURE_TIMESTAMP(psci_svc, PSCI_STAT_ID_ENTER_LOW_PWR,
CACHE_MAINTENANCE_ATTR);
}
/*
* Capture timestamp after exiting a low power state.
* Cache maintenance may be needed when reading these timestamps.
*/
void plat_psci_stat_accounting_stop(
__unused const psci_power_state_t *state_info)
{
assert(state_info != NULL);
PMF_CAPTURE_TIMESTAMP(psci_svc, PSCI_STAT_ID_EXIT_LOW_PWR,
CACHE_MAINTENANCE_ATTR);
}
/*
* Calculate the residency for the given level and power state
* information.
*/
u_register_t plat_psci_stat_get_residency(unsigned int lvl,
const psci_power_state_t *state_info,
unsigned int last_cpu_idx)
{
unsigned long long pwrup_ts = 0, pwrdn_ts = 0;
unsigned int pmf_flags;
assert((lvl >= PSCI_CPU_PWR_LVL) && (lvl <= PLAT_MAX_PWR_LVL));
assert(last_cpu_idx <= PLATFORM_CORE_COUNT);
if (lvl == PSCI_CPU_PWR_LVL)
assert(last_cpu_idx == plat_my_core_pos());
#if HW_ASSISTED_COHERENCY
/* HW coherency allows for the capture and access to happen with caches
* ON. So these timestamps don't need cache maintenance */
pmf_flags = PMF_NO_CACHE_MAINT;
#else
/*
* If power down is requested, then timestamp capture will
* be with caches OFF. Hence we have to do cache maintenance
* when reading the timestamp.
*/
plat_local_state_t state;
assert(state_info != NULL);
state = state_info->pwr_domain_state[PSCI_CPU_PWR_LVL];
if (is_local_state_off(state) != 0) {
pmf_flags = PMF_CACHE_MAINT;
} else {
assert(is_local_state_retn(state) == 1);
pmf_flags = PMF_NO_CACHE_MAINT;
}
#endif
PMF_GET_TIMESTAMP_BY_INDEX(psci_svc,
PSCI_STAT_ID_ENTER_LOW_PWR,
last_cpu_idx,
pmf_flags,
pwrdn_ts);
PMF_GET_TIMESTAMP_BY_INDEX(psci_svc,
PSCI_STAT_ID_EXIT_LOW_PWR,
plat_my_core_pos(),
pmf_flags,
pwrup_ts);
return calc_stat_residency(pwrup_ts, pwrdn_ts);
}
#endif /* ENABLE_PSCI_STAT && ENABLE_PMF */
/*
* The PSCI generic code uses this API to let the platform participate in state
* coordination during a power management operation. It compares the platform
* specific local power states requested by each cpu for a given power domain
* and returns the coordinated target power state that the domain should
* enter. A platform assigns a number to a local power state. This default
* implementation assumes that the platform assigns these numbers in order of
* increasing depth of the power state i.e. for two power states X & Y, if X < Y
* then X represents a shallower power state than Y. As a result, the
* coordinated target local power state for a power domain will be the minimum
* of the requested local power states.
*/
plat_local_state_t plat_get_target_pwr_state(unsigned int lvl,
const plat_local_state_t *states,
unsigned int ncpu)
{
plat_local_state_t target = PLAT_MAX_OFF_STATE, temp;
const plat_local_state_t *st = states;
unsigned int n = ncpu;
assert(ncpu > 0U);
do {
temp = *st;
st++;
if (temp < target)
target = temp;
n--;
} while (n > 0U);
return target;
}