Files
linux-nvgpu/drivers/gpu/nvgpu/os/linux/driver_common.c
shaochunk 7a65a3e636 gpu: nvgpu: support module parameters for devfreq
For POR like safety GPU, we may specify devfreq
timer and governor to keep GPU running with high
performance.

This change supports module parameters for
specifying devfreq governor and devfreq timer.

safety nvgpu module parameter example:
devfreq_timer="delayed" devfreq_gov="performance"

Regarding devfreq timers, a delayed timer can
ensure that the devfreq monitor polls on time.
However, a deferrable timer might potentially
cause a delay in polling time.

Regarding devfreq governors, the default governor
of nvgpu is nvhost_podgov, which scales the gpu
frequency based on GPU load reported by PMU.
Using the performance governor will keep the GPU
operating at a higher GPU frequency,
providing better performance.

Bug 4084478

Change-Id: I9dfef11648203c6af281e980d3a5790b36742414
Signed-off-by: shaochunk <shaochunk@nvidia.com>
Reviewed-on: https://git-master.nvidia.com/r/c/linux-nvgpu/+/2978810
Reviewed-by: Divya Singhatwaria <dsinghatwari@nvidia.com>
Reviewed-by: Vijayakumar Subbu <vsubbu@nvidia.com>
GVS: Gerrit_Virtual_Submit <buildbot_gerritrpt@nvidia.com>
2023-10-23 05:09:38 -07:00

568 lines
16 KiB
C

/*
* Copyright (c) 2016-2023, NVIDIA CORPORATION. All rights reserved.
*
* This program is free software; you can redistribute it and/or modify it
* under the terms and conditions of the GNU General Public License,
* version 2, as published by the Free Software Foundation.
*
* This program is distributed in the hope it will be useful, but WITHOUT
* ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
* FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License for
* more details.
*
* You should have received a copy of the GNU General Public License
* along with this program. If not, see <http://www.gnu.org/licenses/>.
*/
#include <linux/reboot.h>
#include <nvgpu/errata.h>
#include <linux/dma-mapping.h>
#include <linux/mm.h>
#include <linux/slab.h>
#include <linux/version.h>
#include <linux/pm_runtime.h>
#include <linux/of_platform.h>
#include <uapi/linux/nvgpu.h>
#include <linux/devfreq.h>
#include <nvgpu/defaults.h>
#include <nvgpu/kmem.h>
#include <nvgpu/nvgpu_common.h>
#include <nvgpu/soc.h>
#include <nvgpu/bug.h>
#include <nvgpu/enabled.h>
#include <nvgpu/debug.h>
#include <nvgpu/sizes.h>
#include <nvgpu/gk20a.h>
#include <nvgpu/regops.h>
#include <nvgpu/tsg.h>
#include <nvgpu/gr/gr.h>
#include <nvgpu/cic_rm.h>
#include "platform_gk20a.h"
#include "module.h"
#include "os_linux.h"
#include "sysfs.h"
#include "ioctl.h"
#include "scale.h"
#include "driver_common.h"
#define EMC3D_DEFAULT_RATIO 750
#define NVGPU_BLCG_ENABLEMENT BIT(1)
#define NVGPU_ELCG_ENABLEMENT BIT(3)
#define NVGPU_ELPG_ENABLEMENT BIT(5)
#define NVGPU_FLCG_ENABLEMENT BIT(7)
#define NVGPU_SLCG_ENABLEMENT BIT(9)
#define NVGPU_AELPG_ENABLEMENT BIT(11)
#define NVGPU_MSCG_ENABLEMENT BIT(13)
void nvgpu_kernel_restart(void *cmd)
{
kernel_restart(cmd);
}
void nvgpu_read_support_gpu_tools(struct gk20a *g)
{
struct device_node *np;
int ret = 0;
u32 val = 0;
np = nvgpu_get_node(g);
ret = of_property_read_u32(np, "support-gpu-tools", &val);
if (ret != 0) {
/* The debugger/profiler support should be enabled by default.
* So, set support_gpu_tools to 1 even if the property is missing. */
g->support_gpu_tools = 1;
nvgpu_log_info(g, "GPU tools support enabled by default");
} else {
if (val != 0U) {
g->support_gpu_tools = 1;
} else {
g->support_gpu_tools = 0;
}
}
}
void nvgpu_devfreq_init(struct gk20a *g)
{
struct device *dev = dev_from_gk20a(g);
struct gk20a_platform *platform = dev_get_drvdata(dev);
#if LINUX_VERSION_CODE >= KERNEL_VERSION(5, 9, 0)
/* specify devfreq timer with module parameter */
if (strlen(nvgpu_devfreq_timer) == 0) {
nvgpu_log_info(g, "GPU devfreq uses default timer");
} else if (strncmp(nvgpu_devfreq_timer, "deferrable",
strlen("deferrable")) == 0) {
g->scale_profile->devfreq_profile.timer =
DEVFREQ_TIMER_DEFERRABLE;
nvgpu_log_info(g, "use deferrable devfreq timer");
} else if (strncmp(nvgpu_devfreq_timer, "delayed",
strlen("delayed")) == 0) {
g->scale_profile->devfreq_profile.timer =
DEVFREQ_TIMER_DELAYED;
nvgpu_log_info(g, "use delayed devfreq timer");
} else {
nvgpu_err(g, "specified invalid devfreq timer: %s",
nvgpu_devfreq_timer);
}
#else
nvgpu_log_info(g, "GPU devfreq monitor uses default timer");
#endif
/* specify devfreq governor from module parameter */
#ifdef CONFIG_DEVFREQ_GOV_SIMPLE_ONDEMAND
if (strncmp(nvgpu_devfreq_gov, "simple_ondemand",
strlen("simple_ondemand")) == 0)
platform->devfreq_governor = DEVFREQ_GOV_SIMPLE_ONDEMAND;
#endif
#ifdef CONFIG_DEVFREQ_GOV_PERFORMANCE
if (strncmp(nvgpu_devfreq_gov, "performance",
strlen("performance")) == 0)
platform->devfreq_governor = DEVFREQ_GOV_PERFORMANCE;
#endif
#ifdef CONFIG_DEVFREQ_GOV_POWERSAVE
if (strncmp(nvgpu_devfreq_gov, "powersave", strlen("powersave")) == 0)
platform->devfreq_governor = DEVFREQ_GOV_POWERSAVE;
#endif
#ifdef CONFIG_DEVFREQ_GOV_USERSPACE
if (strncmp(nvgpu_devfreq_gov, "userspace", strlen("userspace")) == 0)
platform->devfreq_governor = DEVFREQ_GOV_USERSPACE;
#endif
#ifdef CONFIG_DEVFREQ_GOV_PASSIVE
if (strncmp(nvgpu_devfreq_gov, "passive", strlen("passive")) == 0)
platform->devfreq_governor = DEVFREQ_GOV_PASSIVE;
#endif
nvgpu_log_info(g, "GPU devfreq governor: %s",
platform->devfreq_governor);
}
static void nvgpu_init_vars(struct gk20a *g)
{
struct nvgpu_os_linux *l = nvgpu_os_linux_from_gk20a(g);
struct device *dev = dev_from_gk20a(g);
struct gk20a_platform *platform = dev_get_drvdata(dev);
init_rwsem(&l->busy_lock);
nvgpu_rwsem_init(&g->deterministic_busy);
nvgpu_spinlock_init(&g->mc.enable_lock);
nvgpu_spinlock_init(&g->power_spinlock);
nvgpu_spinlock_init(&g->mc.intr_lock);
nvgpu_mutex_init(&g->mc.intr_thread_mutex);
nvgpu_mutex_init(&platform->railgate_lock);
nvgpu_mutex_init(&g->dbg_sessions_lock);
nvgpu_mutex_init(&g->power_lock);
nvgpu_mutex_init(&g->static_pg_lock);
nvgpu_mutex_init(&g->clk_arb_enable_lock);
nvgpu_mutex_init(&g->cg_pg_lock);
#if defined(CONFIG_NVGPU_CYCLESTATS)
nvgpu_mutex_init(&g->cs_lock);
#endif
#ifdef CONFIG_NVGPU_TSG_SHARING
nvgpu_mutex_init(&g->ctrl_dev_id_lock);
#endif
/* Init the clock req count to 0 */
nvgpu_atomic_set(&g->clk_arb_global_nr, 0);
/* Atomic set doesn't guarantee a barrier */
nvgpu_smp_wmb();
nvgpu_mutex_init(&l->ctrl_privs_lock);
nvgpu_init_list_node(&l->ctrl_privs);
g->regs_saved = g->regs;
g->bar1_saved = g->bar1;
g->emc3d_ratio = EMC3D_DEFAULT_RATIO;
/* Set DMA parameters to allow larger sgt lists */
dev->dma_parms = &l->dma_parms;
dma_set_max_seg_size(dev, UINT_MAX);
/*
* A default of 16GB is the largest supported DMA size that is
* acceptable to all currently supported Tegra SoCs.
*/
if (!platform->dma_mask)
platform->dma_mask = DMA_BIT_MASK(34);
dma_set_mask(dev, platform->dma_mask);
dma_set_coherent_mask(dev, platform->dma_mask);
dma_set_seg_boundary(dev, platform->dma_mask);
nvgpu_init_list_node(&g->profiler_objects);
nvgpu_init_list_node(&g->boardobj_head);
nvgpu_init_list_node(&g->boardobjgrp_head);
nvgpu_set_enabled(g, NVGPU_HAS_SYNCPOINTS, platform->has_syncpoints);
nvgpu_set_enabled(g, NVGPU_SUPPORT_NVS, true);
}
static void nvgpu_init_max_comptag(struct gk20a *g)
{
#ifdef CONFIG_NVGPU_COMPRESSION
#if LINUX_VERSION_CODE >= KERNEL_VERSION(5, 0, 0)
nvgpu_log_info(g, "total ram pages : %lu", totalram_pages());
#else
nvgpu_log_info(g, "total ram pages : %lu", totalram_pages);
#endif
g->max_comptag_mem = totalram_size_in_mb;
#endif
}
static void nvgpu_init_timeout(struct gk20a *g)
{
struct gk20a_platform *platform = dev_get_drvdata(dev_from_gk20a(g));
g->timeouts_disabled_by_user = false;
nvgpu_atomic_set(&g->timeouts_disabled_refcount, 0);
if (nvgpu_platform_is_silicon(g)) {
g->poll_timeout_default = NVGPU_DEFAULT_POLL_TIMEOUT_MS;
g->ch_wdt_init_limit_ms = platform->ch_wdt_init_limit_ms;
} else if (nvgpu_platform_is_fpga(g)) {
g->poll_timeout_default = NVGPU_DEFAULT_FPGA_TIMEOUT_MS;
g->ch_wdt_init_limit_ms = 100U * platform->ch_wdt_init_limit_ms;
} else {
g->poll_timeout_default = (u32)ULONG_MAX;
g->ch_wdt_init_limit_ms = 100U * platform->ch_wdt_init_limit_ms;
}
g->ctxsw_timeout_period_ms = CTXSW_TIMEOUT_PERIOD_MS;
}
static void nvgpu_init_timeslice(struct gk20a *g)
{
g->runlist_interleave = true;
g->tsg_timeslice_low_priority_us =
NVGPU_TSG_TIMESLICE_LOW_PRIORITY_US;
g->tsg_timeslice_medium_priority_us =
NVGPU_TSG_TIMESLICE_MEDIUM_PRIORITY_US;
g->tsg_timeslice_high_priority_us =
NVGPU_TSG_TIMESLICE_HIGH_PRIORITY_US;
g->tsg_timeslice_min_us = NVGPU_TSG_TIMESLICE_MIN_US;
g->tsg_timeslice_max_us = NVGPU_TSG_TIMESLICE_MAX_US;
g->tsg_dbg_timeslice_max_us = NVGPU_TSG_DBG_TIMESLICE_MAX_US_DEFAULT;
}
static void nvgpu_init_pm_vars(struct gk20a *g)
{
struct device *dev = dev_from_gk20a(g);
struct gk20a_platform *platform = dev_get_drvdata(dev);
/* Consider the power setting from module parameter */
platform->enable_blcg = nvgpu_lpwr_enable & NVGPU_BLCG_ENABLEMENT ?
nvgpu_lpwr_enable & (NVGPU_BLCG_ENABLEMENT >> 1) :
platform->enable_blcg;
platform->enable_elcg = nvgpu_lpwr_enable & NVGPU_ELCG_ENABLEMENT ?
nvgpu_lpwr_enable & (NVGPU_ELCG_ENABLEMENT >> 1) :
platform->enable_elcg;
platform->enable_elpg = nvgpu_lpwr_enable & NVGPU_ELPG_ENABLEMENT ?
nvgpu_lpwr_enable & (NVGPU_ELPG_ENABLEMENT >> 1) :
platform->enable_elpg;
platform->enable_flcg = nvgpu_lpwr_enable & NVGPU_FLCG_ENABLEMENT ?
nvgpu_lpwr_enable & (NVGPU_FLCG_ENABLEMENT >> 1) :
platform->enable_flcg;
platform->enable_slcg = nvgpu_lpwr_enable & NVGPU_SLCG_ENABLEMENT ?
nvgpu_lpwr_enable & (NVGPU_SLCG_ENABLEMENT >> 1) :
platform->enable_slcg;
platform->enable_aelpg = nvgpu_lpwr_enable & NVGPU_AELPG_ENABLEMENT ?
nvgpu_lpwr_enable & (NVGPU_AELPG_ENABLEMENT >> 1) :
platform->enable_aelpg;
platform->enable_mscg = nvgpu_lpwr_enable & NVGPU_MSCG_ENABLEMENT ?
nvgpu_lpwr_enable & (NVGPU_MSCG_ENABLEMENT >> 1) :
platform->enable_mscg;
/*
* Set up initial power settings. For non-slicon platforms, disable
* power features and for silicon platforms, read from platform data
*/
g->flcg_enabled =
nvgpu_platform_is_silicon(g) ? platform->enable_flcg : false;
g->slcg_enabled =
nvgpu_platform_is_silicon(g) ? platform->enable_slcg : false;
g->blcg_enabled =
nvgpu_platform_is_silicon(g) ? platform->enable_blcg : false;
g->elcg_enabled =
nvgpu_platform_is_silicon(g) ? platform->enable_elcg : false;
/* disable devfreq for pre-silicon */
if (!nvgpu_platform_is_silicon(g)) {
platform->devfreq_governor = NULL;
#if LINUX_VERSION_CODE >= KERNEL_VERSION(5, 10, 0)
platform->qos_min_notify = NULL;
platform->qos_max_notify = NULL;
#else
platform->qos_notify = NULL;
#endif
}
nvgpu_set_enabled(g, NVGPU_GPU_CAN_ELCG,
nvgpu_platform_is_silicon(g) ? platform->can_elcg : false);
nvgpu_set_enabled(g, NVGPU_GPU_CAN_FLCG,
nvgpu_platform_is_silicon(g) ? platform->can_flcg : false);
nvgpu_set_enabled(g, NVGPU_GPU_CAN_SLCG,
nvgpu_platform_is_silicon(g) ? platform->can_slcg : false);
nvgpu_set_enabled(g, NVGPU_GPU_CAN_BLCG,
nvgpu_platform_is_silicon(g) ? platform->can_blcg : false);
g->aggressive_sync_destroy_thresh = platform->aggressive_sync_destroy_thresh;
#ifdef CONFIG_NVGPU_SUPPORT_CDE
g->has_cde = platform->has_cde;
#endif
g->ptimer_src_freq = platform->ptimer_src_freq;
if (nvgpu_is_hypervisor_mode(g)) {
nvgpu_set_enabled(g, NVGPU_CAN_RAILGATE, false);
platform->can_railgate_init = false;
/* Disable frequency scaling for hypervisor platforms */
platform->devfreq_governor = NULL;
#if LINUX_VERSION_CODE >= KERNEL_VERSION(5, 10, 0)
platform->qos_min_notify = NULL;
platform->qos_max_notify = NULL;
#else
platform->qos_notify = NULL;
#endif
} else {
/* Always enable railgating on simulation platform */
platform->can_railgate_init = nvgpu_platform_is_simulation(g) ?
true : platform->can_railgate_init;
/*
* Disable railgating if GPU power domain node is not defined
* in the DT as bpmp will not powergate/ungate the GPU on
* suspend/resume and can lead to ACR failure on resume
* as it expects GPU to be reset on every resume.
*/
if (!of_property_read_bool(dev->of_node, "power-domains")) {
platform->can_railgate_init = false;
}
nvgpu_set_enabled(g, NVGPU_CAN_RAILGATE,
platform->can_railgate_init);
}
#ifdef CONFIG_NVGPU_STATIC_POWERGATE
g->can_tpc_pg = platform->can_tpc_pg;
g->can_gpc_pg = platform->can_gpc_pg;
g->can_fbp_pg = platform->can_fbp_pg;
#endif
g->ldiv_slowdown_factor = platform->ldiv_slowdown_factor_init;
/* if default delay is not set, set default delay to 500msec */
if (platform->railgate_delay_init)
g->railgate_delay = platform->railgate_delay_init;
else
g->railgate_delay = NVGPU_DEFAULT_RAILGATE_IDLE_TIMEOUT;
g->support_ls_pmu = support_gk20a_pmu(dev_from_gk20a(g));
if (g->support_ls_pmu) {
if (nvgpu_is_hypervisor_mode(g)) {
g->elpg_enabled = false;
g->aelpg_enabled = false;
g->can_elpg = false;
} else {
g->elpg_enabled =
nvgpu_platform_is_silicon(g) ? platform->enable_elpg : false;
g->aelpg_enabled =
nvgpu_platform_is_silicon(g) ? platform->enable_aelpg : false;
g->can_elpg =
nvgpu_platform_is_silicon(g) ? platform->can_elpg_init : false;
}
g->mscg_enabled =
nvgpu_platform_is_silicon(g) ? platform->enable_mscg : false;
if (nvgpu_is_enabled(g, NVGPU_SUPPORT_MIG)) {
g->can_elpg = false;
}
nvgpu_set_enabled(g, NVGPU_PMU_PERFMON, platform->enable_perfmon);
/* ELPG feature enable is SW pre-requisite for ELPG_MS */
if (g->elpg_enabled) {
nvgpu_set_enabled(g, NVGPU_ELPG_MS_ENABLED,
platform->enable_elpg_ms);
g->elpg_ms_enabled = platform->enable_elpg_ms;
}
}
nvgpu_set_enabled(g, NVGPU_SUPPORT_ASPM, !platform->disable_aspm);
#ifdef CONFIG_NVGPU_SIM
if (nvgpu_is_enabled(g, NVGPU_IS_FMODEL)) {
nvgpu_set_enabled(g, NVGPU_PMU_PSTATE, false);
} else
#endif
{
nvgpu_set_enabled(g, NVGPU_PMU_PSTATE, platform->pstate);
}
}
static void nvgpu_init_vbios_vars(struct gk20a *g)
{
struct gk20a_platform *platform = dev_get_drvdata(dev_from_gk20a(g));
nvgpu_set_enabled(g, NVGPU_PMU_RUN_PREOS, platform->run_preos);
}
static void nvgpu_init_ltc_vars(struct gk20a *g)
{
struct gk20a_platform *platform = dev_get_drvdata(dev_from_gk20a(g));
g->ltc_streamid = platform->ltc_streamid;
}
static void nvgpu_init_mm_vars(struct gk20a *g)
{
struct gk20a_platform *platform = dev_get_drvdata(dev_from_gk20a(g));
g->mm.disable_bigpage = platform->disable_bigpage;
nvgpu_set_enabled(g, NVGPU_MM_HONORS_APERTURE,
platform->honors_aperture);
nvgpu_set_enabled(g, NVGPU_MM_UNIFIED_MEMORY,
platform->unified_memory);
nvgpu_set_enabled(g, NVGPU_MM_UNIFY_ADDRESS_SPACES,
platform->unify_address_spaces);
nvgpu_set_errata(g, NVGPU_ERRATA_MM_FORCE_128K_PMU_VM,
platform->force_128K_pmu_vm);
nvgpu_mutex_init(&g->mm.tlb_lock);
}
int nvgpu_probe(struct gk20a *g,
const char *debugfs_symlink)
{
struct device *dev = dev_from_gk20a(g);
struct gk20a_platform *platform = dev_get_drvdata(dev);
struct nvgpu_os_linux *l = nvgpu_os_linux_from_gk20a(g);
int err = 0;
struct device_node *np = dev->of_node;
bool disable_l3_alloc = false;
err = nvgpu_cic_rm_setup(g);
if (err != 0) {
nvgpu_err(g, "CIC-RM setup failed");
return err;
}
err = nvgpu_cic_rm_init_vars(g);
if (err != 0) {
nvgpu_err(g, "CIC-RM init vars failed");
(void) nvgpu_cic_rm_remove(g);
return err;
}
nvgpu_init_vars(g);
nvgpu_init_max_comptag(g);
nvgpu_init_timeout(g);
nvgpu_init_timeslice(g);
nvgpu_init_pm_vars(g);
nvgpu_init_vbios_vars(g);
nvgpu_init_ltc_vars(g);
err = nvgpu_init_soc_vars(g);
if (err) {
nvgpu_err(g, "init soc vars failed");
return err;
}
/* Initialize the platform interface. */
err = platform->probe(dev);
if (err) {
if (err == -EPROBE_DEFER)
nvgpu_info(g, "platform probe failed");
else
nvgpu_err(g, "platform probe failed");
return err;
}
disable_l3_alloc = of_property_read_bool(np, "disable_l3_alloc");
if (disable_l3_alloc) {
nvgpu_log_info(g, "L3 alloc is disabled");
nvgpu_set_enabled(g, NVGPU_DISABLE_L3_SUPPORT, true);
}
nvgpu_init_mm_vars(g);
err = gk20a_power_node_init(dev);
if (err) {
nvgpu_err(g, "power_node creation failed");
return err;
}
/* Read the DT 'support-gpu-tools' property before creating
* user nodes (via gk20a_user_nodes_init().
*/
nvgpu_read_support_gpu_tools(g);
/*
* TODO: While removing the legacy nodes the following condition
* need to be removed.
*/
if (platform->platform_chip_id == TEGRA_210) {
err = gk20a_user_nodes_init(dev);
if (err)
return err;
l->dev_nodes_created = true;
}
/*
* Note that for runtime suspend to work the clocks have to be setup
* which happens in the probe call above. Hence the driver resume
* is done here and not in gk20a_pm_init.
*/
pm_runtime_get_sync(dev);
if (platform->late_probe) {
err = platform->late_probe(dev);
if (err) {
nvgpu_err(g, "late probe failed");
return err;
}
}
pm_runtime_put_sync_autosuspend(dev);
nvgpu_create_sysfs(dev);
gk20a_debug_init(g, debugfs_symlink);
#ifdef CONFIG_NVGPU_DEBUGGER
g->dbg_regops_tmp_buf = nvgpu_kzalloc(g, SZ_4K);
if (!g->dbg_regops_tmp_buf) {
nvgpu_err(g, "couldn't allocate regops tmp buf");
return -ENOMEM;
}
g->dbg_regops_tmp_buf_ops =
SZ_4K / sizeof(g->dbg_regops_tmp_buf[0]);
#endif
g->remove_support = gk20a_remove_support;
nvgpu_ref_init(&g->refcount);
return 0;
}
static void nvgpu_free_gk20a(struct gk20a *g)
{
struct nvgpu_os_linux *l = nvgpu_os_linux_from_gk20a(g);
g->probe_done = false;
kfree(l);
}
void nvgpu_init_gk20a(struct gk20a *g)
{
g->gfree = nvgpu_free_gk20a;
}