drm/tegra: gr3d: Support generic power domain and runtime PM
authorDmitry Osipenko <digetx@gmail.com>
Tue, 30 Nov 2021 23:23:21 +0000 (02:23 +0300)
committerThierry Reding <treding@nvidia.com>
Thu, 16 Dec 2021 13:07:08 +0000 (14:07 +0100)
Add runtime power management and support generic power domains.

Reviewed-by: Ulf Hansson <ulf.hansson@linaro.org>
Tested-by: Peter Geis <pgwipeout@gmail.com> # Ouya T30
Tested-by: Paul Fertser <fercerpav@gmail.com> # PAZ00 T20
Tested-by: Nicolas Chauvet <kwizart@gmail.com> # PAZ00 T20 and TK1 T124
Tested-by: Matt Merhar <mattmerhar@protonmail.com> # Ouya T30
Signed-off-by: Dmitry Osipenko <digetx@gmail.com>
Signed-off-by: Thierry Reding <treding@nvidia.com>
drivers/gpu/drm/tegra/gr3d.c

index 24442ad..032d713 100644 (file)
@@ -5,32 +5,47 @@
  */
 
 #include <linux/clk.h>
+#include <linux/delay.h>
 #include <linux/host1x.h>
 #include <linux/iommu.h>
 #include <linux/module.h>
 #include <linux/of_device.h>
 #include <linux/platform_device.h>
+#include <linux/pm_domain.h>
+#include <linux/pm_opp.h>
+#include <linux/pm_runtime.h>
 #include <linux/reset.h>
 
+#include <soc/tegra/common.h>
 #include <soc/tegra/pmc.h>
 
 #include "drm.h"
 #include "gem.h"
 #include "gr3d.h"
 
+enum {
+       RST_MC,
+       RST_GR3D,
+       RST_MC2,
+       RST_GR3D2,
+       RST_GR3D_MAX,
+};
+
 struct gr3d_soc {
        unsigned int version;
+       unsigned int num_clocks;
+       unsigned int num_resets;
 };
 
 struct gr3d {
        struct tegra_drm_client client;
        struct host1x_channel *channel;
-       struct clk *clk_secondary;
-       struct clk *clk;
-       struct reset_control *rst_secondary;
-       struct reset_control *rst;
 
        const struct gr3d_soc *soc;
+       struct clk_bulk_data *clocks;
+       unsigned int nclocks;
+       struct reset_control_bulk_data resets[RST_GR3D_MAX];
+       unsigned int nresets;
 
        DECLARE_BITMAP(addr_regs, GR3D_NUM_REGS);
 };
@@ -65,15 +80,22 @@ static int gr3d_init(struct host1x_client *client)
                goto free;
        }
 
+       pm_runtime_enable(client->dev);
+       pm_runtime_use_autosuspend(client->dev);
+       pm_runtime_set_autosuspend_delay(client->dev, 200);
+
        err = tegra_drm_register_client(dev->dev_private, drm);
        if (err < 0) {
                dev_err(client->dev, "failed to register client: %d\n", err);
-               goto detach;
+               goto disable_rpm;
        }
 
        return 0;
 
-detach:
+disable_rpm:
+       pm_runtime_dont_use_autosuspend(client->dev);
+       pm_runtime_force_suspend(client->dev);
+
        host1x_client_iommu_detach(client);
 free:
        host1x_syncpt_put(client->syncpts[0]);
@@ -93,10 +115,15 @@ static int gr3d_exit(struct host1x_client *client)
        if (err < 0)
                return err;
 
+       pm_runtime_dont_use_autosuspend(client->dev);
+       pm_runtime_force_suspend(client->dev);
+
        host1x_client_iommu_detach(client);
        host1x_syncpt_put(client->syncpts[0]);
        host1x_channel_put(gr3d->channel);
 
+       gr3d->channel = NULL;
+
        return 0;
 }
 
@@ -109,10 +136,17 @@ static int gr3d_open_channel(struct tegra_drm_client *client,
                             struct tegra_drm_context *context)
 {
        struct gr3d *gr3d = to_gr3d(client);
+       int err;
+
+       err = pm_runtime_resume_and_get(client->base.dev);
+       if (err)
+               return err;
 
        context->channel = host1x_channel_get(gr3d->channel);
-       if (!context->channel)
+       if (!context->channel) {
+               pm_runtime_put(client->base.dev);
                return -ENOMEM;
+       }
 
        return 0;
 }
@@ -120,6 +154,7 @@ static int gr3d_open_channel(struct tegra_drm_client *client,
 static void gr3d_close_channel(struct tegra_drm_context *context)
 {
        host1x_channel_put(context->channel);
+       pm_runtime_put(context->client->base.dev);
 }
 
 static int gr3d_is_addr_reg(struct device *dev, u32 class, u32 offset)
@@ -155,14 +190,20 @@ static const struct tegra_drm_client_ops gr3d_ops = {
 
 static const struct gr3d_soc tegra20_gr3d_soc = {
        .version = 0x20,
+       .num_clocks = 1,
+       .num_resets = 2,
 };
 
 static const struct gr3d_soc tegra30_gr3d_soc = {
        .version = 0x30,
+       .num_clocks = 2,
+       .num_resets = 4,
 };
 
 static const struct gr3d_soc tegra114_gr3d_soc = {
        .version = 0x35,
+       .num_clocks = 1,
+       .num_resets = 2,
 };
 
 static const struct of_device_id tegra_gr3d_match[] = {
@@ -278,69 +319,216 @@ static const u32 gr3d_addr_regs[] = {
        GR3D_GLOBAL_SAMP23SURFADDR(15),
 };
 
-static int gr3d_probe(struct platform_device *pdev)
+static int gr3d_power_up_legacy_domain(struct device *dev, const char *name,
+                                      unsigned int id)
 {
-       struct device_node *np = pdev->dev.of_node;
-       struct host1x_syncpt **syncpts;
-       struct gr3d *gr3d;
+       struct gr3d *gr3d = dev_get_drvdata(dev);
+       struct reset_control *reset;
+       struct clk *clk;
        unsigned int i;
        int err;
 
-       gr3d = devm_kzalloc(&pdev->dev, sizeof(*gr3d), GFP_KERNEL);
-       if (!gr3d)
-               return -ENOMEM;
-
-       gr3d->soc = of_device_get_match_data(&pdev->dev);
+       /*
+        * Tegra20 device-tree doesn't specify 3d clock name and there is only
+        * one clock for Tegra20. Tegra30+ device-trees always specified names
+        * for the clocks.
+        */
+       if (gr3d->nclocks == 1) {
+               if (id == TEGRA_POWERGATE_3D1)
+                       return 0;
+
+               clk = gr3d->clocks[0].clk;
+       } else {
+               for (i = 0; i < gr3d->nclocks; i++) {
+                       if (WARN_ON(!gr3d->clocks[i].id))
+                               continue;
+
+                       if (!strcmp(gr3d->clocks[i].id, name)) {
+                               clk = gr3d->clocks[i].clk;
+                               break;
+                       }
+               }
 
-       syncpts = devm_kzalloc(&pdev->dev, sizeof(*syncpts), GFP_KERNEL);
-       if (!syncpts)
-               return -ENOMEM;
+               if (WARN_ON(i == gr3d->nclocks))
+                       return -EINVAL;
+       }
 
-       gr3d->clk = devm_clk_get(&pdev->dev, NULL);
-       if (IS_ERR(gr3d->clk)) {
-               dev_err(&pdev->dev, "cannot get clock\n");
-               return PTR_ERR(gr3d->clk);
+       /*
+        * We use array of resets, which includes MC resets, and MC
+        * reset shouldn't be asserted while hardware is gated because
+        * MC flushing will fail for gated hardware. Hence for legacy
+        * PD we request the individual reset separately.
+        */
+       reset = reset_control_get_exclusive_released(dev, name);
+       if (IS_ERR(reset))
+               return PTR_ERR(reset);
+
+       err = reset_control_acquire(reset);
+       if (err) {
+               dev_err(dev, "failed to acquire %s reset: %d\n", name, err);
+       } else {
+               err = tegra_powergate_sequence_power_up(id, clk, reset);
+               reset_control_release(reset);
        }
 
-       gr3d->rst = devm_reset_control_get(&pdev->dev, "3d");
-       if (IS_ERR(gr3d->rst)) {
-               dev_err(&pdev->dev, "cannot get reset\n");
-               return PTR_ERR(gr3d->rst);
+       reset_control_put(reset);
+       if (err)
+               return err;
+
+       /*
+        * tegra_powergate_sequence_power_up() leaves clocks enabled,
+        * while GENPD not. Hence keep clock-enable balanced.
+        */
+       clk_disable_unprepare(clk);
+
+       return 0;
+}
+
+static void gr3d_del_link(void *link)
+{
+       device_link_del(link);
+}
+
+static int gr3d_init_power(struct device *dev, struct gr3d *gr3d)
+{
+       static const char * const opp_genpd_names[] = { "3d0", "3d1", NULL };
+       const u32 link_flags = DL_FLAG_STATELESS | DL_FLAG_PM_RUNTIME;
+       struct device **opp_virt_devs, *pd_dev;
+       struct device_link *link;
+       unsigned int i;
+       int err;
+
+       err = of_count_phandle_with_args(dev->of_node, "power-domains",
+                                        "#power-domain-cells");
+       if (err < 0) {
+               if (err != -ENOENT)
+                       return err;
+
+               /*
+                * Older device-trees don't use GENPD. In this case we should
+                * toggle power domain manually.
+                */
+               err = gr3d_power_up_legacy_domain(dev, "3d",
+                                                 TEGRA_POWERGATE_3D);
+               if (err)
+                       return err;
+
+               err = gr3d_power_up_legacy_domain(dev, "3d2",
+                                                 TEGRA_POWERGATE_3D1);
+               if (err)
+                       return err;
+
+               return 0;
        }
 
-       if (of_device_is_compatible(np, "nvidia,tegra30-gr3d")) {
-               gr3d->clk_secondary = devm_clk_get(&pdev->dev, "3d2");
-               if (IS_ERR(gr3d->clk_secondary)) {
-                       dev_err(&pdev->dev, "cannot get secondary clock\n");
-                       return PTR_ERR(gr3d->clk_secondary);
+       /*
+        * The PM domain core automatically attaches a single power domain,
+        * otherwise it skips attaching completely. We have a single domain
+        * on Tegra20 and two domains on Tegra30+.
+        */
+       if (dev->pm_domain)
+               return 0;
+
+       err = devm_pm_opp_attach_genpd(dev, opp_genpd_names, &opp_virt_devs);
+       if (err)
+               return err;
+
+       for (i = 0; opp_genpd_names[i]; i++) {
+               pd_dev = opp_virt_devs[i];
+               if (!pd_dev) {
+                       dev_err(dev, "failed to get %s power domain\n",
+                               opp_genpd_names[i]);
+                       return -EINVAL;
                }
 
-               gr3d->rst_secondary = devm_reset_control_get(&pdev->dev,
-                                                               "3d2");
-               if (IS_ERR(gr3d->rst_secondary)) {
-                       dev_err(&pdev->dev, "cannot get secondary reset\n");
-                       return PTR_ERR(gr3d->rst_secondary);
+               link = device_link_add(dev, pd_dev, link_flags);
+               if (!link) {
+                       dev_err(dev, "failed to link to %s\n", dev_name(pd_dev));
+                       return -EINVAL;
                }
+
+               err = devm_add_action_or_reset(dev, gr3d_del_link, link);
+               if (err)
+                       return err;
        }
 
-       err = tegra_powergate_sequence_power_up(TEGRA_POWERGATE_3D, gr3d->clk,
-                                               gr3d->rst);
+       return 0;
+}
+
+static int gr3d_get_clocks(struct device *dev, struct gr3d *gr3d)
+{
+       int err;
+
+       err = devm_clk_bulk_get_all(dev, &gr3d->clocks);
        if (err < 0) {
-               dev_err(&pdev->dev, "failed to power up 3D unit\n");
+               dev_err(dev, "failed to get clock: %d\n", err);
                return err;
        }
+       gr3d->nclocks = err;
 
-       if (gr3d->clk_secondary) {
-               err = tegra_powergate_sequence_power_up(TEGRA_POWERGATE_3D1,
-                                                       gr3d->clk_secondary,
-                                                       gr3d->rst_secondary);
-               if (err < 0) {
-                       dev_err(&pdev->dev,
-                               "failed to power up secondary 3D unit\n");
-                       return err;
-               }
+       if (gr3d->nclocks != gr3d->soc->num_clocks) {
+               dev_err(dev, "invalid number of clocks: %u\n", gr3d->nclocks);
+               return -ENOENT;
+       }
+
+       return 0;
+}
+
+static int gr3d_get_resets(struct device *dev, struct gr3d *gr3d)
+{
+       int err;
+
+       gr3d->resets[RST_MC].id = "mc";
+       gr3d->resets[RST_MC2].id = "mc2";
+       gr3d->resets[RST_GR3D].id = "3d";
+       gr3d->resets[RST_GR3D2].id = "3d2";
+       gr3d->nresets = gr3d->soc->num_resets;
+
+       err = devm_reset_control_bulk_get_optional_exclusive_released(
+                               dev, gr3d->nresets, gr3d->resets);
+       if (err) {
+               dev_err(dev, "failed to get reset: %d\n", err);
+               return err;
        }
 
+       if (WARN_ON(!gr3d->resets[RST_GR3D].rstc) ||
+           WARN_ON(!gr3d->resets[RST_GR3D2].rstc && gr3d->nresets == 4))
+               return -ENOENT;
+
+       return 0;
+}
+
+static int gr3d_probe(struct platform_device *pdev)
+{
+       struct host1x_syncpt **syncpts;
+       struct gr3d *gr3d;
+       unsigned int i;
+       int err;
+
+       gr3d = devm_kzalloc(&pdev->dev, sizeof(*gr3d), GFP_KERNEL);
+       if (!gr3d)
+               return -ENOMEM;
+
+       platform_set_drvdata(pdev, gr3d);
+
+       gr3d->soc = of_device_get_match_data(&pdev->dev);
+
+       syncpts = devm_kzalloc(&pdev->dev, sizeof(*syncpts), GFP_KERNEL);
+       if (!syncpts)
+               return -ENOMEM;
+
+       err = gr3d_get_clocks(&pdev->dev, gr3d);
+       if (err)
+               return err;
+
+       err = gr3d_get_resets(&pdev->dev, gr3d);
+       if (err)
+               return err;
+
+       err = gr3d_init_power(&pdev->dev, gr3d);
+       if (err)
+               return err;
+
        INIT_LIST_HEAD(&gr3d->client.base.list);
        gr3d->client.base.ops = &gr3d_client_ops;
        gr3d->client.base.dev = &pdev->dev;
@@ -352,6 +540,10 @@ static int gr3d_probe(struct platform_device *pdev)
        gr3d->client.version = gr3d->soc->version;
        gr3d->client.ops = &gr3d_ops;
 
+       err = devm_tegra_core_dev_init_opp_table_common(&pdev->dev);
+       if (err)
+               return err;
+
        err = host1x_client_register(&gr3d->client.base);
        if (err < 0) {
                dev_err(&pdev->dev, "failed to register host1x client: %d\n",
@@ -363,8 +555,6 @@ static int gr3d_probe(struct platform_device *pdev)
        for (i = 0; i < ARRAY_SIZE(gr3d_addr_regs); i++)
                set_bit(gr3d_addr_regs[i], gr3d->addr_regs);
 
-       platform_set_drvdata(pdev, gr3d);
-
        return 0;
 }
 
@@ -380,23 +570,80 @@ static int gr3d_remove(struct platform_device *pdev)
                return err;
        }
 
-       if (gr3d->clk_secondary) {
-               reset_control_assert(gr3d->rst_secondary);
-               tegra_powergate_power_off(TEGRA_POWERGATE_3D1);
-               clk_disable_unprepare(gr3d->clk_secondary);
+       return 0;
+}
+
+static int __maybe_unused gr3d_runtime_suspend(struct device *dev)
+{
+       struct gr3d *gr3d = dev_get_drvdata(dev);
+       int err;
+
+       host1x_channel_stop(gr3d->channel);
+
+       err = reset_control_bulk_assert(gr3d->nresets, gr3d->resets);
+       if (err) {
+               dev_err(dev, "failed to assert reset: %d\n", err);
+               return err;
        }
 
-       reset_control_assert(gr3d->rst);
-       tegra_powergate_power_off(TEGRA_POWERGATE_3D);
-       clk_disable_unprepare(gr3d->clk);
+       usleep_range(10, 20);
+
+       /*
+        * Older device-trees don't specify MC resets and power-gating can't
+        * be done safely in that case. Hence we will keep the power ungated
+        * for older DTBs. For newer DTBs, GENPD will perform the power-gating.
+        */
+
+       clk_bulk_disable_unprepare(gr3d->nclocks, gr3d->clocks);
+       reset_control_bulk_release(gr3d->nresets, gr3d->resets);
 
        return 0;
 }
 
+static int __maybe_unused gr3d_runtime_resume(struct device *dev)
+{
+       struct gr3d *gr3d = dev_get_drvdata(dev);
+       int err;
+
+       err = reset_control_bulk_acquire(gr3d->nresets, gr3d->resets);
+       if (err) {
+               dev_err(dev, "failed to acquire reset: %d\n", err);
+               return err;
+       }
+
+       err = clk_bulk_prepare_enable(gr3d->nclocks, gr3d->clocks);
+       if (err) {
+               dev_err(dev, "failed to enable clock: %d\n", err);
+               goto release_reset;
+       }
+
+       err = reset_control_bulk_deassert(gr3d->nresets, gr3d->resets);
+       if (err) {
+               dev_err(dev, "failed to deassert reset: %d\n", err);
+               goto disable_clk;
+       }
+
+       return 0;
+
+disable_clk:
+       clk_bulk_disable_unprepare(gr3d->nclocks, gr3d->clocks);
+release_reset:
+       reset_control_bulk_release(gr3d->nresets, gr3d->resets);
+
+       return err;
+}
+
+static const struct dev_pm_ops tegra_gr3d_pm = {
+       SET_RUNTIME_PM_OPS(gr3d_runtime_suspend, gr3d_runtime_resume, NULL)
+       SET_SYSTEM_SLEEP_PM_OPS(pm_runtime_force_suspend,
+                               pm_runtime_force_resume)
+};
+
 struct platform_driver tegra_gr3d_driver = {
        .driver = {
                .name = "tegra-gr3d",
                .of_match_table = tegra_gr3d_match,
+               .pm = &tegra_gr3d_pm,
        },
        .probe = gr3d_probe,
        .remove = gr3d_remove,