> The deeper the idle state, the lower the power consumption, but the longer
> the resume time. Some service are delay sensitive and very except the low
> resume time, like interrupt packet receiving mode.
>
> And the "/sys/devices/system/cpu/cpuX/power/pm_qos_resume_latency_us" sysfs
> interface is used to set and get the resume latency limit on the cpuX for
> userspace. Each cpuidle governor in Linux select which idle state to enter
> based on this CPU resume latency in their idle task.
>
> The per-CPU PM QoS API can be used to control this CPU's idle state
> selection and limit just enter the shallowest idle state to low the delay
> after sleep by setting strict resume latency (zero value).
>
> Signed-off-by: Huisong Li <lihuis...@huawei.com>
> Acked-by: Morten Brørup <m...@smartsharesystems.com>
> ---
> doc/guides/prog_guide/power_man.rst | 24 ++++++
> doc/guides/rel_notes/release_24_11.rst | 5 ++
> lib/power/meson.build | 2 +
> lib/power/rte_power_qos.c | 111 +++++++++++++++++++++++++
> lib/power/rte_power_qos.h | 73 ++++++++++++++++
> lib/power/version.map | 4 +
> 6 files changed, 219 insertions(+)
> create mode 100644 lib/power/rte_power_qos.c
> create mode 100644 lib/power/rte_power_qos.h
>
> diff --git a/doc/guides/prog_guide/power_man.rst
> b/doc/guides/prog_guide/power_man.rst
> index f6674efe2d..faa32b4320 100644
> --- a/doc/guides/prog_guide/power_man.rst
> +++ b/doc/guides/prog_guide/power_man.rst
> @@ -249,6 +249,30 @@ Get Num Pkgs
> Get Num Dies
> Get the number of die's on a given package.
>
> +
> +PM QoS
> +------
> +
> +The deeper the idle state, the lower the power consumption, but the longer
> +the resume time. Some service are delay sensitive and very except the low
> +resume time, like interrupt packet receiving mode.
> +
> +And the "/sys/devices/system/cpu/cpuX/power/pm_qos_resume_latency_us" sysfs
> +interface is used to set and get the resume latency limit on the cpuX for
> +userspace. Each cpuidle governor in Linux select which idle state to enter
> +based on this CPU resume latency in their idle task.
> +
> +The per-CPU PM QoS API can be used to set and get the CPU resume latency
> based
> +on this sysfs.
> +
> +The ``rte_power_qos_set_cpu_resume_latency()`` function can control the CPU's
> +idle state selection in Linux and limit just to enter the shallowest idle
> state
> +to low the delay of resuming service after sleeping by setting strict resume
> +latency (zero value).
> +
> +The ``rte_power_qos_get_cpu_resume_latency()`` function can get the resume
> +latency on specified CPU.
> +
> References
> ----------
>
> diff --git a/doc/guides/rel_notes/release_24_11.rst
> b/doc/guides/rel_notes/release_24_11.rst
> index 0ff70d9057..bd72d0a595 100644
> --- a/doc/guides/rel_notes/release_24_11.rst
> +++ b/doc/guides/rel_notes/release_24_11.rst
> @@ -55,6 +55,11 @@ New Features
> Also, make sure to start the actual text at the margin.
> =======================================================
>
> +* **Introduce per-CPU PM QoS interface.**
> +
> + * Add per-CPU PM QoS interface to low the delay after sleep by controlling
> + CPU idle state selection.
> +
>
> Removed Items
> -------------
> diff --git a/lib/power/meson.build b/lib/power/meson.build
> index b8426589b2..8222e178b0 100644
> --- a/lib/power/meson.build
> +++ b/lib/power/meson.build
> @@ -23,12 +23,14 @@ sources = files(
> 'rte_power.c',
> 'rte_power_uncore.c',
> 'rte_power_pmd_mgmt.c',
> + 'rte_power_qos.c',
> )
> headers = files(
> 'rte_power.h',
> 'rte_power_guest_channel.h',
> 'rte_power_pmd_mgmt.h',
> 'rte_power_uncore.h',
> + 'rte_power_qos.h',
> )
> if cc.has_argument('-Wno-cast-qual')
> cflags += '-Wno-cast-qual'
> diff --git a/lib/power/rte_power_qos.c b/lib/power/rte_power_qos.c
> new file mode 100644
> index 0000000000..8eb26cd41a
> --- /dev/null
> +++ b/lib/power/rte_power_qos.c
> @@ -0,0 +1,111 @@
> +/* SPDX-License-Identifier: BSD-3-Clause
> + * Copyright(c) 2024 HiSilicon Limited
> + */
> +
> +#include <errno.h>
> +#include <stdlib.h>
> +#include <string.h>
> +
> +#include <rte_lcore.h>
> +#include <rte_log.h>
> +
> +#include "power_common.h"
> +#include "rte_power_qos.h"
> +
> +#define PM_QOS_SYSFILE_RESUME_LATENCY_US \
> + "/sys/devices/system/cpu/cpu%u/power/pm_qos_resume_latency_us"
> +
> +#define PM_QOS_CPU_RESUME_LATENCY_BUF_LEN 32
> +
> +int
> +rte_power_qos_set_cpu_resume_latency(uint16_t lcore_id, int latency)
> +{
> + char buf[PM_QOS_CPU_RESUME_LATENCY_BUF_LEN];
> + FILE *f;
> + int ret;
> +
> + if (!rte_lcore_is_enabled(lcore_id)) {
> + POWER_LOG(ERR, "lcore id %u is not enabled", lcore_id);
> + return -EINVAL;
> + }
> +
> + if (latency < 0) {
> + POWER_LOG(ERR, "latency should be greater than and equal to 0");
> + return -EINVAL;
> + }
> +
> + ret = open_core_sysfs_file(&f, "w", PM_QOS_SYSFILE_RESUME_LATENCY_US,
> lcore_id);
That was already brought by Morten:
lcore_id is not always equal to cpu_core_id (cpu affinity).
Looking through power library it is not specific to that particular patch,
but sort of common limitation (bug?) in rte_power lib.
> + if (ret != 0) {
> + POWER_LOG(ERR, "Failed to open
> "PM_QOS_SYSFILE_RESUME_LATENCY_US, lcore_id);
> + return ret;
> + }
> +
> + /*
> + * Based on the sysfs interface pm_qos_resume_latency_us under
> + * @PM_QOS_SYSFILE_RESUME_LATENCY_US directory in kernel, their meaning
> + * is as follows for different input string.
> + * 1> the resume latency is 0 if the input is "n/a".
> + * 2> the resume latency is no constraint if the input is "0".
> + * 3> the resume latency is the actual value to be set.
> + */
> + if (latency == 0)
> + snprintf(buf, sizeof(buf), "%s", "n/a");
> + else if (latency == RTE_POWER_QOS_RESUME_LATENCY_NO_CONSTRAINT)
> + snprintf(buf, sizeof(buf), "%u", 0);
> + else
> + snprintf(buf, sizeof(buf), "%u", latency);
> +
> + ret = write_core_sysfs_s(f, buf);
> + if (ret != 0)
> + POWER_LOG(ERR, "Failed to write
> "PM_QOS_SYSFILE_RESUME_LATENCY_US, lcore_id);
> +
> + fclose(f);
> +
> + return ret;
> +}
> +
> +int
> +rte_power_qos_get_cpu_resume_latency(uint16_t lcore_id)
> +{
> + char buf[PM_QOS_CPU_RESUME_LATENCY_BUF_LEN];
> + int latency = -1;
> + FILE *f;
> + int ret;
> +
> + if (!rte_lcore_is_enabled(lcore_id)) {
> + POWER_LOG(ERR, "lcore id %u is not enabled", lcore_id);
> + return -EINVAL;
> + }
> +
> + ret = open_core_sysfs_file(&f, "r", PM_QOS_SYSFILE_RESUME_LATENCY_US,
> lcore_id);
> + if (ret != 0) {
> + POWER_LOG(ERR, "Failed to open
> "PM_QOS_SYSFILE_RESUME_LATENCY_US, lcore_id);
> + return ret;
> + }
> +
> + ret = read_core_sysfs_s(f, buf, sizeof(buf));
> + if (ret != 0) {
> + POWER_LOG(ERR, "Failed to read
> "PM_QOS_SYSFILE_RESUME_LATENCY_US, lcore_id);
> + goto out;
> + }
> +
> + /*
> + * Based on the sysfs interface pm_qos_resume_latency_us under
> + * @PM_QOS_SYSFILE_RESUME_LATENCY_US directory in kernel, their meaning
> + * is as follows for different output string.
> + * 1> the resume latency is 0 if the output is "n/a".
> + * 2> the resume latency is no constraint if the output is "0".
> + * 3> the resume latency is the actual value in used for other string.
> + */
> + if (strcmp(buf, "n/a") == 0)
> + latency = 0;
> + else {
> + latency = strtoul(buf, NULL, 10);
> + latency = latency == 0 ?
> RTE_POWER_QOS_RESUME_LATENCY_NO_CONSTRAINT : latency;
> + }
> +
> +out:
> + fclose(f);
> +
> + return latency != -1 ? latency : ret;
> +}
> diff --git a/lib/power/rte_power_qos.h b/lib/power/rte_power_qos.h
> new file mode 100644
> index 0000000000..990c488373
> --- /dev/null
> +++ b/lib/power/rte_power_qos.h
> @@ -0,0 +1,73 @@
> +/* SPDX-License-Identifier: BSD-3-Clause
> + * Copyright(c) 2024 HiSilicon Limited
> + */
> +
> +#ifndef RTE_POWER_QOS_H
> +#define RTE_POWER_QOS_H
> +
> +#include <stdint.h>
> +
> +#include <rte_compat.h>
> +
> +#ifdef __cplusplus
> +extern "C" {
> +#endif
> +
> +/**
> + * @file rte_power_qos.h
> + *
> + * PM QoS API.
> + *
> + * The CPU-wide resume latency limit has a positive impact on this CPU's idle
> + * state selection in each cpuidle governor.
> + * Please see the PM QoS on CPU wide in the following link:
> + *
> https://www.kernel.org/doc/html/latest/admin-guide/abi-testing.html?highlight=pm_qos_resume_latency_us#abi-sys-devices-
> power-pm-qos-resume-latency-us
> + *
> + * The deeper the idle state, the lower the power consumption, but the
> + * longer the resume time. Some service are delay sensitive and very except
> the
> + * low resume time, like interrupt packet receiving mode.
> + *
> + * In these case, per-CPU PM QoS API can be used to control this CPU's idle
> + * state selection and limit just enter the shallowest idle state to low the
> + * delay after sleep by setting strict resume latency (zero value).
> + */
> +
> +#define RTE_POWER_QOS_STRICT_LATENCY_VALUE 0
> +#define RTE_POWER_QOS_RESUME_LATENCY_NO_CONSTRAINT ((int)(UINT32_MAX >>
> 1))
> +
> +/**
> + * @warning
> + * @b EXPERIMENTAL: this API may change without prior notice.
> + *
> + * @param lcore_id
> + * target logical core id
> + *
> + * @param latency
> + * The latency should be greater than and equal to zero in microseconds
> unit.
> + *
> + * @return
> + * 0 on success. Otherwise negative value is returned.
> + */
> +__rte_experimental
> +int rte_power_qos_set_cpu_resume_latency(uint16_t lcore_id, int latency);
> +
> +/**
> + * @warning
> + * @b EXPERIMENTAL: this API may change without prior notice.
> + *
> + * Get the current resume latency of this logical core.
> + * The default value in kernel is @see
> RTE_POWER_QOS_RESUME_LATENCY_NO_CONSTRAINT
> + * if don't set it.
> + *
> + * @return
> + * Negative value on failure.
> + * >= 0 means the actual resume latency limit on this core.
> + */
> +__rte_experimental
> +int rte_power_qos_get_cpu_resume_latency(uint16_t lcore_id);
> +
> +#ifdef __cplusplus
> +}
> +#endif
> +
> +#endif /* RTE_POWER_QOS_H */
> diff --git a/lib/power/version.map b/lib/power/version.map
> index c9a226614e..08f178a39d 100644
> --- a/lib/power/version.map
> +++ b/lib/power/version.map
> @@ -51,4 +51,8 @@ EXPERIMENTAL {
> rte_power_set_uncore_env;
> rte_power_uncore_freqs;
> rte_power_unset_uncore_env;
> +
> + # added in 24.11
> + rte_power_qos_get_cpu_resume_latency;
> + rte_power_qos_set_cpu_resume_latency;
> };
> --
> 2.22.0
