msm: kgsl: Port GPU bus dcvs to kernel 4.4

Port GPU dcvs from kernel 3.18 to kernel 4.4.

CRs-Fixed: 1013343
Change-Id: Ide662b12aa59effa541febcd758426e72b4a1b12
Signed-off-by: Oleg Perelet <operelet@codeaurora.org>
This commit is contained in:
Oleg Perelet 2016-05-10 17:30:07 -07:00 committed by Jeevan Shriram
parent c75e96b609
commit 8df0f9ec1d
7 changed files with 459 additions and 5 deletions

View file

@ -89,6 +89,15 @@ config MSM_BIMC_BWMON
has the capability to raise an IRQ when the count exceeds a
programmable limit.
config DEVFREQ_GOV_QCOM_GPUBW_MON
tristate "GPU BW voting governor"
depends on DEVFREQ_GOV_QCOM_ADRENO_TZ
help
QTI GPU governor for GPU bus bandwidth voting.
This governor works together with QTI Adreno Trustzone governor,
and select bus frequency votes using an "on-demand" alorithm.
This governor is unlikely to be useful for non-QTI devices.
config ARMBW_HWMON
tristate "ARM PMU Bandwidth monitor hardware"
depends on ARCH_QCOM

View file

@ -14,6 +14,8 @@ obj-$(CONFIG_MSM_M4M_HWMON) += m4m-hwmon.o
obj-$(CONFIG_DEVFREQ_GOV_MSM_BW_HWMON) += governor_bw_hwmon.o
obj-$(CONFIG_DEVFREQ_GOV_MSM_CACHE_HWMON) += governor_cache_hwmon.o
obj-$(CONFIG_DEVFREQ_GOV_SPDM_HYP) += governor_spdm_bw_hyp.o
obj-$(CONFIG_DEVFREQ_GOV_QCOM_GPUBW_MON) += governor_gpubw_mon.o
obj-$(CONFIG_DEVFREQ_GOV_QCOM_GPUBW_MON) += governor_bw_vbif.o
obj-$(CONFIG_DEVFREQ_GOV_MEMLAT) += governor_memlat.o
# DEVFREQ Drivers

View file

@ -0,0 +1,44 @@
/* Copyright (c) 2014-2016, The Linux Foundation. All rights reserved.
*
* This program is free software; you can redistribute it and/or modify
* it under the terms of the GNU General Public License version 2 and
* only version 2 as published by the Free Software Foundation.
*
* This program is distributed in the hope that 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.
*
*/
#if !defined(_DEVFREQ_TRACE_H) || defined(TRACE_HEADER_MULTI_READ)
#define _DEVFREQ_TRACE_H
#undef TRACE_SYSTEM
#define TRACE_SYSTEM devfreq
#undef TRACE_INCLUDE_PATH
#define TRACE_INCLUDE_PATH .
#undef TRACE_INCLUDE_FILE
#define TRACE_INCLUDE_FILE devfreq_trace
#include <linux/tracepoint.h>
TRACE_EVENT(devfreq_msg,
TP_PROTO(const char *msg),
TP_ARGS(msg),
TP_STRUCT__entry(
__string(msg, msg)
),
TP_fast_assign(
__assign_str(msg, msg);
),
TP_printk(
"%s", __get_str(msg)
)
);
#endif /* _DEVFREQ_TRACE_H */
/* This part must be outside protection */
#include <trace/define_trace.h>

View file

@ -0,0 +1,144 @@
/*
* Copyright (c) 2014-2016, The Linux Foundation. All rights reserved.
*
* This program is free software; you can redistribute it and/or modify
* it under the terms of the GNU General Public License version 2 and
* only version 2 as published by the Free Software Foundation.
*
* This program is distributed in the hope that 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.
*/
#include <linux/devfreq.h>
#include <linux/module.h>
#include "governor.h"
unsigned long (*extern_get_bw)(void) = NULL;
unsigned long *dev_ab;
static unsigned long dev_ib;
DEFINE_MUTEX(df_lock);
static struct devfreq *df;
/*
* This function is 'get_target_freq' API for the governor.
* It just calls an external function that should be registered
* by KGSL driver to get and return a value for frequency.
*/
static int devfreq_vbif_get_freq(struct devfreq *df,
unsigned long *freq,
u32 *flag)
{
/* If the IB isn't set yet, check if it should be non-zero. */
if (!dev_ib && extern_get_bw) {
dev_ib = extern_get_bw();
if (dev_ab)
*dev_ab = dev_ib / 4;
}
*freq = dev_ib;
return 0;
}
/*
* Registers a function to be used to request a frequency
* value from legacy vbif based bus bandwidth governor.
* This function is called by KGSL driver.
*/
void devfreq_vbif_register_callback(void *p)
{
extern_get_bw = p;
}
int devfreq_vbif_update_bw(unsigned long ib, unsigned long ab)
{
int ret = 0;
mutex_lock(&df_lock);
if (df) {
mutex_lock(&df->lock);
dev_ib = ib;
*dev_ab = ab;
ret = update_devfreq(df);
mutex_unlock(&df->lock);
}
mutex_unlock(&df_lock);
return ret;
}
static int devfreq_vbif_ev_handler(struct devfreq *devfreq,
unsigned int event, void *data)
{
int ret;
struct devfreq_dev_status stat;
switch (event) {
case DEVFREQ_GOV_START:
mutex_lock(&df_lock);
df = devfreq;
if (df->profile->get_dev_status)
ret = df->profile->get_dev_status(df->dev.parent,
&stat);
else
ret = 0;
if (ret || !stat.private_data)
pr_warn("Device doesn't take AB votes!\n");
else
dev_ab = stat.private_data;
mutex_unlock(&df_lock);
ret = devfreq_vbif_update_bw(0, 0);
if (ret) {
pr_err("Unable to update BW! Gov start failed!\n");
return ret;
}
/*
* Normally at this point governors start the polling with
* devfreq_monitor_start(df);
* This governor doesn't poll, but expect external calls
* of its devfreq_vbif_update_bw() function
*/
pr_debug("Enabled MSM VBIF governor\n");
break;
case DEVFREQ_GOV_STOP:
mutex_lock(&df_lock);
df = NULL;
mutex_unlock(&df_lock);
pr_debug("Disabled MSM VBIF governor\n");
break;
}
return 0;
}
static struct devfreq_governor devfreq_vbif = {
.name = "bw_vbif",
.get_target_freq = devfreq_vbif_get_freq,
.event_handler = devfreq_vbif_ev_handler,
};
static int __init devfreq_vbif_init(void)
{
return devfreq_add_governor(&devfreq_vbif);
}
subsys_initcall(devfreq_vbif_init);
static void __exit devfreq_vbif_exit(void)
{
int ret;
ret = devfreq_remove_governor(&devfreq_vbif);
if (ret)
pr_err("%s: failed remove governor %d\n", __func__, ret);
}
module_exit(devfreq_vbif_exit);
MODULE_DESCRIPTION("VBIF based GPU bus BW voting governor");
MODULE_LICENSE("GPL v2");

View file

@ -0,0 +1,255 @@
/* Copyright (c) 2014-2016, The Linux Foundation. All rights reserved.
*
* This program is free software; you can redistribute it and/or modify
* it under the terms of the GNU General Public License version 2 and
* only version 2 as published by the Free Software Foundation.
*
* This program is distributed in the hope that 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.
*
*/
#include <linux/devfreq.h>
#include <linux/module.h>
#include <linux/msm_adreno_devfreq.h>
#include <linux/slab.h>
#include "devfreq_trace.h"
#include "governor.h"
#define MIN_BUSY 1000
#define LONG_FLOOR 50000
#define HIST 5
#define TARGET 80
#define CAP 75
/* AB vote is in multiple of BW_STEP Mega bytes */
#define BW_STEP 160
static void _update_cutoff(struct devfreq_msm_adreno_tz_data *priv,
unsigned int norm_max)
{
int i;
priv->bus.max = norm_max;
for (i = 0; i < priv->bus.num; i++) {
priv->bus.up[i] = priv->bus.p_up[i] * norm_max / 100;
priv->bus.down[i] = priv->bus.p_down[i] * norm_max / 100;
}
}
static int devfreq_gpubw_get_target(struct devfreq *df,
unsigned long *freq,
u32 *flag)
{
struct devfreq_msm_adreno_tz_data *priv = df->data;
struct msm_busmon_extended_profile *bus_profile = container_of(
(df->profile),
struct msm_busmon_extended_profile,
profile);
struct devfreq_dev_status stats;
struct xstats b;
int result;
int level = 0;
int act_level;
int norm_cycles;
int gpu_percent;
/*
* Normalized AB should at max usage be the gpu_bimc frequency in MHz.
* Start with a reasonable value and let the system push it up to max.
*/
static int norm_ab_max = 300;
int norm_ab;
unsigned long ab_mbytes = 0;
if (priv == NULL)
return 0;
stats.private_data = &b;
result = df->profile->get_dev_status(df->dev.parent, &stats);
*freq = stats.current_frequency;
priv->bus.total_time += stats.total_time;
priv->bus.gpu_time += stats.busy_time;
priv->bus.ram_time += b.ram_time;
priv->bus.ram_wait += b.ram_wait;
level = devfreq_get_freq_level(df, stats.current_frequency);
if (priv->bus.total_time < LONG_FLOOR)
return result;
norm_cycles = (unsigned int)(priv->bus.ram_time + priv->bus.ram_wait) /
(unsigned int) priv->bus.total_time;
gpu_percent = (100 * (unsigned int)priv->bus.gpu_time) /
(unsigned int) priv->bus.total_time;
/*
* If there's a new high watermark, update the cutoffs and send the
* FAST hint. Otherwise check the current value against the current
* cutoffs.
*/
if (norm_cycles > priv->bus.max) {
_update_cutoff(priv, norm_cycles);
bus_profile->flag = DEVFREQ_FLAG_FAST_HINT;
} else {
/* GPU votes for IB not AB so don't under vote the system */
norm_cycles = (100 * norm_cycles) / TARGET;
act_level = priv->bus.index[level] + b.mod;
act_level = (act_level < 0) ? 0 : act_level;
act_level = (act_level >= priv->bus.num) ?
(priv->bus.num - 1) : act_level;
if (norm_cycles > priv->bus.up[act_level] &&
gpu_percent > CAP)
bus_profile->flag = DEVFREQ_FLAG_FAST_HINT;
else if (norm_cycles < priv->bus.down[act_level] && level)
bus_profile->flag = DEVFREQ_FLAG_SLOW_HINT;
}
/* Calculate the AB vote based on bus width if defined */
if (priv->bus.width) {
norm_ab = (unsigned int)priv->bus.ram_time /
(unsigned int) priv->bus.total_time;
/* Calculate AB in Mega Bytes and roundup in BW_STEP */
ab_mbytes = (norm_ab * priv->bus.width * 1000000ULL) >> 20;
bus_profile->ab_mbytes = roundup(ab_mbytes, BW_STEP);
} else if (bus_profile->flag) {
/* Re-calculate the AB percentage for a new IB vote */
norm_ab = (unsigned int)priv->bus.ram_time /
(unsigned int) priv->bus.total_time;
if (norm_ab > norm_ab_max)
norm_ab_max = norm_ab;
bus_profile->percent_ab = (100 * norm_ab) / norm_ab_max;
}
priv->bus.total_time = 0;
priv->bus.gpu_time = 0;
priv->bus.ram_time = 0;
priv->bus.ram_wait = 0;
return result;
}
static int gpubw_start(struct devfreq *devfreq)
{
struct devfreq_msm_adreno_tz_data *priv;
struct msm_busmon_extended_profile *bus_profile = container_of(
(devfreq->profile),
struct msm_busmon_extended_profile,
profile);
unsigned int t1, t2 = 2 * HIST;
int i, bus_size;
devfreq->data = bus_profile->private_data;
priv = devfreq->data;
bus_size = sizeof(u32) * priv->bus.num;
priv->bus.up = kzalloc(bus_size, GFP_KERNEL);
priv->bus.down = kzalloc(bus_size, GFP_KERNEL);
priv->bus.p_up = kzalloc(bus_size, GFP_KERNEL);
priv->bus.p_down = kzalloc(bus_size, GFP_KERNEL);
if (priv->bus.up == NULL || priv->bus.down == NULL ||
priv->bus.p_up == NULL || priv->bus.p_down == NULL)
return -ENOMEM;
/* Set up the cut-over percentages for the bus calculation. */
for (i = 0; i < priv->bus.num; i++) {
t1 = (u32)(100 * priv->bus.ib[i]) /
(u32)priv->bus.ib[priv->bus.num - 1];
priv->bus.p_up[i] = t1 - HIST;
priv->bus.p_down[i] = t2 - 2 * HIST;
t2 = t1;
}
/* Set the upper-most and lower-most bounds correctly. */
priv->bus.p_down[0] = 0;
priv->bus.p_down[1] = (priv->bus.p_down[1] > (2 * HIST)) ?
priv->bus.p_down[1] : (2 * HIST);
if (priv->bus.num >= 1)
priv->bus.p_up[priv->bus.num - 1] = 100;
_update_cutoff(priv, priv->bus.max);
return 0;
}
static int gpubw_stop(struct devfreq *devfreq)
{
struct devfreq_msm_adreno_tz_data *priv = devfreq->data;
if (priv) {
kfree(priv->bus.up);
kfree(priv->bus.down);
kfree(priv->bus.p_up);
kfree(priv->bus.p_down);
}
devfreq->data = NULL;
return 0;
}
static int devfreq_gpubw_event_handler(struct devfreq *devfreq,
unsigned int event, void *data)
{
int result = 0;
unsigned long freq;
mutex_lock(&devfreq->lock);
freq = devfreq->previous_freq;
switch (event) {
case DEVFREQ_GOV_START:
result = gpubw_start(devfreq);
break;
case DEVFREQ_GOV_STOP:
result = gpubw_stop(devfreq);
break;
case DEVFREQ_GOV_RESUME:
/* TODO ..... */
/* ret = update_devfreq(devfreq); */
break;
case DEVFREQ_GOV_SUSPEND:
{
struct devfreq_msm_adreno_tz_data *priv = devfreq->data;
priv->bus.total_time = 0;
priv->bus.gpu_time = 0;
priv->bus.ram_time = 0;
}
break;
default:
result = 0;
break;
}
mutex_unlock(&devfreq->lock);
return result;
}
static struct devfreq_governor devfreq_gpubw = {
.name = "gpubw_mon",
.get_target_freq = devfreq_gpubw_get_target,
.event_handler = devfreq_gpubw_event_handler,
};
static int __init devfreq_gpubw_init(void)
{
return devfreq_add_governor(&devfreq_gpubw);
}
subsys_initcall(devfreq_gpubw_init);
static void __exit devfreq_gpubw_exit(void)
{
int ret;
ret = devfreq_remove_governor(&devfreq_gpubw);
if (ret)
pr_err("%s: failed remove governor %d\n", __func__, ret);
}
module_exit(devfreq_gpubw_exit);
MODULE_DESCRIPTION("GPU bus bandwidth voting driver. Uses VBIF counters");
MODULE_LICENSE("GPL v2");

View file

@ -116,7 +116,7 @@ static void _record_pwrevent(struct kgsl_device *device,
}
}
#ifdef CONFIG_DEVFREQ_GOV_MSM_GPUBW_MON
#ifdef CONFIG_DEVFREQ_GOV_QCOM_GPUBW_MON
#include <soc/qcom/devfreq_devbw.h>
/**
@ -199,7 +199,7 @@ static unsigned int _adjust_pwrlevel(struct kgsl_pwrctrl *pwr, int level,
return level;
}
#ifdef CONFIG_DEVFREQ_GOV_MSM_GPUBW_MON
#ifdef CONFIG_DEVFREQ_GOV_QCOM_GPUBW_MON
static void kgsl_pwrctrl_vbif_update(unsigned long ab)
{
/* ask a governor to vote on behalf of us */
@ -1398,7 +1398,7 @@ static void kgsl_pwrctrl_clk(struct kgsl_device *device, int state,
}
}
#ifdef CONFIG_DEVFREQ_GOV_MSM_GPUBW_MON
#ifdef CONFIG_DEVFREQ_GOV_QCOM_GPUBW_MON
static void kgsl_pwrctrl_suspend_devbw(struct kgsl_pwrctrl *pwr)
{
if (pwr->devbw)
@ -1596,7 +1596,7 @@ void kgsl_deep_nap_timer(unsigned long data)
}
}
#ifdef CONFIG_DEVFREQ_GOV_MSM_GPUBW_MON
#ifdef CONFIG_DEVFREQ_GOV_QCOM_GPUBW_MON
static int kgsl_pwrctrl_vbif_init(void)
{
devfreq_vbif_register_callback(kgsl_get_bw);

View file

@ -71,7 +71,7 @@ struct msm_busmon_extended_profile {
struct devfreq_dev_profile profile;
};
#ifdef CONFIG_DEVFREQ_GOV_MSM_GPUBW_MON
#ifdef CONFIG_DEVFREQ_GOV_QCOM_GPUBW_MON
int devfreq_vbif_update_bw(unsigned long ib, unsigned long ab);
int devfreq_vbif_register_callback(void *);
#endif