lkml.org 
[lkml]   [2021]   [Mar]   [18]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
Patch in this message
/
SubjectRe: [PATCH v1] memory: tegra20: Add debug statistics
From
Date
On 18/03/2021 12:55, Dmitry Osipenko wrote:
> Add debug statistics collection support. The statistics is available
> via debugfs in '/sys/kernel/debug/mc/stats', it shows percent of memory
> controller utilization for each memory client. This information is
> intended to help with debugging of memory performance issues, it already
> was proven to be useful by helping to improve memory bandwidth management
> of the display driver.
>
> Signed-off-by: Dmitry Osipenko <digetx@gmail.com>
> ---
> drivers/memory/tegra/mc.c | 12 ++
> drivers/memory/tegra/tegra20.c | 342 +++++++++++++++++++++++++++++++++
> include/soc/tegra/mc.h | 7 +
> 3 files changed, 361 insertions(+)
>
> diff --git a/drivers/memory/tegra/mc.c b/drivers/memory/tegra/mc.c
> index a21163ccadc4..85c3364ebf26 100644
> --- a/drivers/memory/tegra/mc.c
> +++ b/drivers/memory/tegra/mc.c
> @@ -827,6 +827,18 @@ static int tegra_mc_probe(struct platform_device *pdev)
> return err;
> }
>
> + mc->debugfs.root = debugfs_create_dir("mc", NULL);
> + if (!mc->debugfs.root)
> + dev_err(&pdev->dev, "failed to create debugfs directory\n");

It's error pointer, not null, but anyway there is no need for handling
debugfs error. See Greg KH's commits like "remove pointless check for
debugfs_create_dir()".

> +
> + if (mc->soc->init) {
> + err = mc->soc->init(mc);
> + if (err < 0)
> + dev_err(&pdev->dev,
> + "failed to register initialize SoC driver: %d\n",

"failed to initialize SoC driver:...."

> + err);
> + }
> +
> err = tegra_mc_reset_setup(mc);
> if (err < 0)
> dev_err(&pdev->dev, "failed to register reset controller: %d\n",
> diff --git a/drivers/memory/tegra/tegra20.c b/drivers/memory/tegra/tegra20.c
> index 29ecf02805a0..513c07104296 100644
> --- a/drivers/memory/tegra/tegra20.c
> +++ b/drivers/memory/tegra/tegra20.c
> @@ -3,6 +3,8 @@
> * Copyright (C) 2012 NVIDIA CORPORATION. All rights reserved.
> */
>
> +#include <linux/bitfield.h>
> +#include <linux/delay.h>
> #include <linux/of_device.h>
> #include <linux/slab.h>
> #include <linux/string.h>
> @@ -11,6 +13,77 @@
>
> #include "mc.h"
>
> +#define MC_STAT_CONTROL 0x90
> +#define MC_STAT_EMC_CLOCK_LIMIT 0xa0
> +#define MC_STAT_EMC_CLOCKS 0xa4
> +#define MC_STAT_EMC_CONTROL_0 0xa8
> +#define MC_STAT_EMC_CONTROL_1 0xac
> +#define MC_STAT_EMC_COUNT_0 0xb8
> +#define MC_STAT_EMC_COUNT_1 0xbc
> +
> +#define MC_STAT_CONTROL_CLIENT_ID GENMASK(13, 8)
> +#define MC_STAT_CONTROL_EVENT GENMASK(23, 16)
> +#define MC_STAT_CONTROL_PRI_EVENT GENMASK(25, 24)
> +#define MC_STAT_CONTROL_FILTER_CLIENT_ENABLE GENMASK(26, 26)
> +#define MC_STAT_CONTROL_FILTER_PRI GENMASK(29, 28)
> +
> +#define MC_STAT_CONTROL_PRI_EVENT_HP 0
> +#define MC_STAT_CONTROL_PRI_EVENT_TM 1
> +#define MC_STAT_CONTROL_PRI_EVENT_BW 2
> +
> +#define MC_STAT_CONTROL_FILTER_PRI_DISABLE 0
> +#define MC_STAT_CONTROL_FILTER_PRI_NO 1
> +#define MC_STAT_CONTROL_FILTER_PRI_YES 2
> +
> +#define MC_STAT_CONTROL_EVENT_QUALIFIED 0
> +#define MC_STAT_CONTROL_EVENT_ANY_READ 1
> +#define MC_STAT_CONTROL_EVENT_ANY_WRITE 2
> +#define MC_STAT_CONTROL_EVENT_RD_WR_CHANGE 3
> +#define MC_STAT_CONTROL_EVENT_SUCCESSIVE 4
> +#define MC_STAT_CONTROL_EVENT_ARB_BANK_AA 5
> +#define MC_STAT_CONTROL_EVENT_ARB_BANK_BB 6
> +#define MC_STAT_CONTROL_EVENT_PAGE_MISS 7
> +#define MC_STAT_CONTROL_EVENT_AUTO_PRECHARGE 8
> +
> +#define EMC_GATHER_RST (0 << 8)
> +#define EMC_GATHER_CLEAR (1 << 8)
> +#define EMC_GATHER_DISABLE (2 << 8)
> +#define EMC_GATHER_ENABLE (3 << 8)
> +
> +#define MC_STAT_SAMPLE_TIME_USEC 16000
> +
> +/* we store collected statistics as a fixed point values */
> +#define MC_FX_FRAC_SCALE 100
> +
> +struct tegra20_mc_stat_gather {
> + unsigned int pri_filter;
> + unsigned int pri_event;
> + unsigned int result;
> + unsigned int client;
> + unsigned int event;
> + bool client_enb;
> +};
> +
> +struct tegra20_mc_stat {
> + struct tegra20_mc_stat_gather gather0;
> + struct tegra20_mc_stat_gather gather1;
> + unsigned int sample_time_usec;
> + struct tegra_mc *mc;
> +};
> +
> +struct tegra20_mc_client_stat {
> + unsigned int events;
> + unsigned int arb_high_prio;
> + unsigned int arb_timeout;
> + unsigned int arb_bandwidth;
> + unsigned int rd_wr_change;
> + unsigned int successive;
> + unsigned int page_miss;
> + unsigned int auto_precharge;
> + unsigned int arb_bank_aa;
> + unsigned int arb_bank_bb;
> +};
> +
> static const struct tegra_mc_client tegra20_mc_clients[] = {
> {
> .id = 0x00,
> @@ -356,6 +429,274 @@ static const struct tegra_mc_icc_ops tegra20_mc_icc_ops = {
> .set = tegra20_mc_icc_set,
> };
>
> +static u32 tegra20_mc_stat_gather_control(struct tegra20_mc_stat_gather *g)

"g" looks like pointer to const here

> +{
> + u32 control;
> +
> + control = FIELD_PREP(MC_STAT_CONTROL_EVENT, g->event);
> + control |= FIELD_PREP(MC_STAT_CONTROL_CLIENT_ID, g->client);
> + control |= FIELD_PREP(MC_STAT_CONTROL_PRI_EVENT, g->pri_event);
> + control |= FIELD_PREP(MC_STAT_CONTROL_FILTER_PRI, g->pri_filter);
> + control |= FIELD_PREP(MC_STAT_CONTROL_FILTER_CLIENT_ENABLE, g->client_enb);
> +
> + return control;
> +}
> +
> +static void tegra20_mc_stat_gather(struct tegra20_mc_stat *stat)
> +{
> + u32 clocks, count0, count1, control_0, control_1;
> + struct tegra_mc *mc = stat->mc;
> +
> + control_0 = tegra20_mc_stat_gather_control(&stat->gather0);
> + control_1 = tegra20_mc_stat_gather_control(&stat->gather1);
> +
> + /*
> + * Reset statistic gathers state, select bandwidth mode for the
> + * statistics collection mode and set clocks counter saturation
> + * limit to maximum.
> + */
> + mc_writel(mc, 0x00000000, MC_STAT_CONTROL);
> + mc_writel(mc, control_0, MC_STAT_EMC_CONTROL_0);
> + mc_writel(mc, control_1, MC_STAT_EMC_CONTROL_1);
> + mc_writel(mc, 0xffffffff, MC_STAT_EMC_CLOCK_LIMIT);
> +
> + mc_writel(mc, EMC_GATHER_ENABLE, MC_STAT_CONTROL);
> + fsleep(stat->sample_time_usec);
> + mc_writel(mc, EMC_GATHER_DISABLE, MC_STAT_CONTROL);
> +
> + count0 = mc_readl(mc, MC_STAT_EMC_COUNT_0);
> + count1 = mc_readl(mc, MC_STAT_EMC_COUNT_1);
> + clocks = mc_readl(mc, MC_STAT_EMC_CLOCKS);
> + clocks = max(clocks / 100 / MC_FX_FRAC_SCALE, 1u);
> +
> + stat->gather0.result = DIV_ROUND_UP(count0, clocks);
> + stat->gather1.result = DIV_ROUND_UP(count1, clocks);
> +}
> +
> +static void tegra20_mc_stat_events(struct tegra_mc *mc,
> + const struct tegra_mc_client *client0,
> + const struct tegra_mc_client *client1,
> + unsigned int pri_filter,
> + unsigned int pri_event,
> + unsigned int event,
> + unsigned int *result0,
> + unsigned int *result1)
> +{
> + struct tegra20_mc_stat stat = {};
> +
> + stat.gather0.client = client0 ? client0->id : 0;
> + stat.gather0.pri_filter = pri_filter;
> + stat.gather0.client_enb = !!client0;
> + stat.gather0.pri_event = pri_event;
> + stat.gather0.event = event;
> +
> + stat.gather1.client = client1 ? client1->id : 0;
> + stat.gather1.pri_filter = pri_filter;
> + stat.gather1.client_enb = !!client1;
> + stat.gather1.pri_event = pri_event;
> + stat.gather1.event = event;
> +
> + stat.sample_time_usec = MC_STAT_SAMPLE_TIME_USEC;
> + stat.mc = mc;
> +
> + tegra20_mc_stat_gather(&stat);
> +
> + *result0 = stat.gather0.result;
> + *result1 = stat.gather1.result;
> +}
> +
> +static void tegra20_mc_collect_stats(struct tegra_mc *mc,
> + struct tegra20_mc_client_stat *stats)
> +{
> + const struct tegra_mc_client *client0, *client1;
> + unsigned int i, clienta, clientb;

Define the clienta and clientb inside the loop, to reduce the scope.
Unless it is used between the loop iterations?

> +
> + /* collect memory controller utilization percent for each client */
> + for (i = 0; i < mc->soc->num_clients; i += 2) {
> + client0 = &mc->soc->clients[i];
> + client1 = &mc->soc->clients[i + 1];
> +
> + if (i + 1 == mc->soc->num_clients)
> + client1 = NULL;
> +
> + tegra20_mc_stat_events(mc, client0, client1,
> + MC_STAT_CONTROL_FILTER_PRI_DISABLE,
> + MC_STAT_CONTROL_PRI_EVENT_HP,
> + MC_STAT_CONTROL_EVENT_QUALIFIED,
> + &stats[i + 0].events,
> + &stats[i + 1].events);
> + }
> +
> + /* collect more info from active clients */
> + for (i = 0; i < mc->soc->num_clients; i++) {
> + clientb = mc->soc->num_clients;
> +
> + for (client0 = NULL; i < mc->soc->num_clients; i++) {
> + if (stats[i].events) {
> + client0 = &mc->soc->clients[i];
> + clienta = i++;
> + break;
> + }
> + }

Could all clients have 0 events ending with "clienta" being undefined?
"client0" would be non-NULL because of loop before.

> +
> + for (client1 = NULL; i < mc->soc->num_clients; i++) {
> + if (stats[i].events) {
> + client1 = &mc->soc->clients[i];
> + clientb = i;
> + break;
> + }
> + }
> +
> + if (!client0 && !client1)
> + break;
> +
> + tegra20_mc_stat_events(mc, client0, client1,
> + MC_STAT_CONTROL_FILTER_PRI_YES,
> + MC_STAT_CONTROL_PRI_EVENT_HP,
> + MC_STAT_CONTROL_EVENT_QUALIFIED,
> + &stats[clienta].arb_high_prio,
> + &stats[clientb].arb_high_prio);
> +


Best regards,
Krzysztof

\
 
 \ /
  Last update: 2021-03-18 16:24    [W:0.065 / U:0.112 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site