Re: [PATCH 1/5] drm: Add helpers to kick off PSR enable/disable

On Fri, Mar 01, 2019 at 11:00:02AM -0500, Sean Paul wrote:

On Fri, Mar 01, 2019 at 09:18:13AM +0100, Daniel Vetter wrote:

...

On Thu, Feb 28, 2019 at 04:09:30PM -0500, Sean Paul wrote:

...

From: Sean Paul seanpaul@chromium.org

This patch adds a new drm helper library to help drivers implement PSR. Drivers choosing to use it will register connectors with PSR-capable displays connected and will receive callbacks when it's time to enter or exit PSR.

In its current form, it has a timer which will trigger after a driver-specified amount of inactivity. When the timer triggers, the helpers will save the current atomic state and issue a new state which has the PSR-enabled pipes turned off. On the next update, the drm core will poke the PSR helpers to restore the saved state to the driver before servicing said update.

From the driver's perspective, this works like a regular disable/enable cycle. The driver need only check the 'psr_transition' state in connector_state and keep the panel turned on when in .disable(), while everything else will cycle off as normal. If drivers want more control, they can use the psr_transition state to enter a low-power state to minimize PSR exit time.

While this carries the PSR moniker, it is not specific to the DisplayPort technology. This can be used for power savings with other types of self refresh, such as MIPI command mode.

Cc: Zain Wang wzz@rock-chips.com Cc: Tomasz Figa tfiga@chromium.org Signed-off-by: Sean Paul seanpaul@chromium.org

/snip to the good part

...

So for the high-level design, before looking at the implementation details:

• I agree with Jose that a per-crtc state sounds more like what we want, instead of global. Just from a conceptual point.

Agreed this would be better. I was mostly worried about resource sharing between crtcs. ie: If 2 crtcs share a clock, 1 goes to PSR and the other changes the rate to something incompatible. However I suppose the driver would be responsible for sussing this all out, so sure.

It'll be nice to avoid the global lock for psr.

Yeah atomic check must fail in this case. But that's not new, if ACTVE=false then the clock will also be disabled (or should, for a good driver), and atomic_check still must fail.

I don't think there's going to be a semantic difference between psr self-refresh and ACTIVE=false. It might turn up more pre-existing driver bugs, but then we just need to fix those (good igt coverage would be good, maybe we can build on top of kms_frontbuffer_rendering a bit here, without the intel-ism).

...

• We can imo flat-out require that your driver is using the fancy new dirtyfb helpers to remap the dirtyfb call to an atomic update. It's the clean design for sw controlled manual update displays anyway.
• That means your single entry points (instead of trying to catch all of them like above, and missing some like you do) are ->atomic_check and ->atomic_commit. That also gives you the clean hook into helpers.

The reason I did this was because new state was being calculated off the disabled state. So the psr disable commit needed to precede the atomic check. However, with your suggestions below I think we can avoid that :)

...

• The duplicated state only works when no one else is looking, it's very fragile. single-threaded (from a driver pov) suspend/resume is ok, I don't think it's a good idea to expand this.

Yeah, fair. I was trying to stay out of the way of the existing code, but I'll integrate more tightly.

...

Here's what I'd do:

• Stuff a new self_refresh_active state into drm_crtc_state (not sure we should call this psr or self_refresh - latter is imo better name since not associated with DP, but longer).

I struggled with this too, naming is hard.

...

• Use crtc_state->self_refresh_active and crtc_state->active to compute the reported active state to userspace in the get_prop functions. This because doing a mass search&replace over all drivers is too much. Something like

  real_active == active || self_refresh

• In duplicate_state always reset the state active = real_active(); self_refresh = false; to disable self refresh and set active to what userspace the "ACTIVE" property should be.

We'll also have to alter the flags when they don't include DRM_MODE_ATOMIC_ALLOW_MODESET, and async commits would need to go through the synchronous path. Conversely, non-blocking commits will get back to userspace more quickly with this design.

This is the primary reason I kept the psr disable in a separate commit, so the original request could go through as intended (albeit a bit delayed, but there's no getting around that).

Hm. In a way we're lying to userspace here. But the lie is intentional I'd say, so what about we make it official in the atomic ioctl? Something like the below:

diff --git a/drivers/gpu/drm/drm_atomic.c b/drivers/gpu/drm/drm_atomic.c index aa0c9af78b71..322bee5ab28c 100644 --- a/drivers/gpu/drm/drm_atomic.c +++ b/drivers/gpu/drm/drm_atomic.c @@ -368,6 +368,18 @@ static int drm_atomic_crtc_check(const struct drm_crtc_state *old_crtc_state, return -EINVAL; }

+ /* + * Transparent self-refresh is a bit a lie to userspace, keep up the + * illusion. We must allow a modeset if one of the crtc is currently in + * self-refresh mode and should be enabled. + */ + if (old_crtc_state->self_refresh && new_crtc_state->active) { + DRM_DEBUG_ATOMIC("[CRTC:%d:%s] self-refresh, transparently allowing modesets\n", + crtc->base.id, crtc->name); + + state->allow_modeset = true; + } + return 0; }

Note how this is in the core atomic ioctl code, not helpers, so mandatory lying for all drivers :-)

igt coverage for this would be really nice, or maybe some in-kernel self-tests ....

...

• self_refresh state in connector_state sounds like a good idea

• Your idle worker would do a normal atomic commit, no evil state duplicate, and then set: self_fresh = active; active = false; Handling races left as an exercise :-)

• Somewhere in the commit helpers put a function to arm your idle work (per-crtc imo simplest).

This should give you a clean lasagne going from core ioctl -> core atomic -> helpers <-> drivers. And not the spaghetti sprawl of finely sprinkling the self refresh helpers all over.

Alright, I'll mock this up and see if it floats. Thanks for the detailed feedback!

Still catching up on mails, hopefully not too much confusion ...

Cheers, Daniel

Sean

...

Aside: Going snowboarding next week, so if this is a bonghits idea I'm afraid will take 1 week until I can spin something new :-)

Cheers, Daniel

...

diff --git a/drivers/gpu/drm/drm_psr_helper.c b/drivers/gpu/drm/drm_psr_helper.c new file mode 100644 index 000000000000..0b57a2a53075 --- /dev/null +++ b/drivers/gpu/drm/drm_psr_helper.c @@ -0,0 +1,343 @@ +/* SPDX-License-Identifier: MIT */ +/*

• • Copyright (C) 2019 Google, Inc.
• • Authors:
• • Sean Paul seanpaul@chromium.org
• */

+#include <drm/drm_atomic.h> +#include <drm/drm_atomic_helper.h> +#include <drm/drm_connector.h> +#include <drm/drm_device.h> +#include <drm/drm_mode_config.h> +#include <drm/drm_modeset_lock.h> +#include <drm/drm_print.h> +#include <drm/drm_psr_helper.h> +#include <linux/bitops.h> +#include <linux/mutex.h> +#include <linux/slab.h> +#include <linux/workqueue.h>

+/**

• • DOC: overview
• • This helper library provides an easy way for drivers to leverage the atomic
• • framework to implement panel self refresh (PSR) support. Drivers are
• • responsible for intializing and cleaning up the PSR helpers on load/unload.
• • When drivers identify a display that supports self refreshing (eDP or MIPI
• • command mode), it should register the affected connector with the PSR
• • helpers.
• • Once a connector is registered, the PSR helpers will monitor activity and
• • call back into the driver to enable/disable PSR as appropriate. The best way
• • to think about this is that it's a DPMS on/off request with a flag set in
• • state that tells you to disable/enable PSR on the panel instead of power-
• • cycling it.
• • Drivers may choose to fully disable their crtc/encoder/bridge hardware, or
• • they can use the "psr_transition" flag in crtc and connector state if they
• • want to enter low power mode without full disable (in case full
• • disable/enable is too slow).
• • PSR will be deactivated if there are any atomic updates, even updates that do
• • not affect the connectors which are self refreshing. Supporting this is
• • possible but non-trivial due to sharing of hardware resources. Similarly, if
• • a crtc is driving multiple connectors, PSR will not be initiated on any of
• • those connectors.
• */

+struct drm_psr_state {

• struct drm_device *dev;
• struct drm_modeset_lock mutex;
• struct delayed_work entry_work;
• struct drm_atomic_state *save_state;
• unsigned int entry_delay_ms;

+};

+static void drm_psr_helper_entry_work(struct work_struct *work) +{

• struct drm_psr_state *psr_state = container_of(to_delayed_work(work),

• 				       struct drm_psr_state,
  

• 				       entry_work);
  

• struct drm_atomic_state *save_state;
• struct drm_device *dev = psr_state->dev;
• struct drm_modeset_acquire_ctx ctx;
• struct drm_atomic_state *state;
• struct drm_connector *conn;
• struct drm_connector_list_iter conn_iter;
• struct drm_connector_state *conn_state;
• struct drm_crtc *crtc;
• struct drm_crtc_state *crtc_state;
• bool commit = false;
• int ret, i;
• DRM_MODESET_LOCK_ALL_BEGIN(dev, ctx, 0, ret);
• ret = drm_modeset_lock(&psr_state->mutex, &ctx);
• if (ret)

• goto out;
  

• /*

• * The only way this can happen is if we schedule the worker while it's
  

• * already running and that timeout subsequently elapses. Since we hold
  

• * the psr_state->mutex when scheduling, we also know where the worker
  

• * is sitting in its execution (hint: look up). In this case, it's
  

• * possible for the entry worker to run twice for the same commit. Since
  

• * the hardware hasn't changed since the last save state, just kick out.
  

• */
  

• if (psr_state->save_state)

• goto out;
  

• state = drm_atomic_state_alloc(dev);
• if (!state) {

• ret = -ENOMEM;
  

• goto out;
  

• }
• save_state = drm_atomic_helper_duplicate_state(dev, &ctx);
• /*

• * Now that we have the current the HW state saved, we have to flip the
  

• * psr_transition bit so we know what type of enable we're dealing with
  

• * when coming back on.
  

• *
  

• * NOTE: We don't check conn->capable here since that could change out
  

• * from under us. We'll trust the atomic core to only call enable if
  

• * necessary (ie: only for those connectors/crtcs that currently have
  

• * psr enabled).
  

• */
  

• if (IS_ERR(save_state)) {

• ret = PTR_ERR(save_state);
  

• goto out;
  

• }
• for_each_new_connector_in_state(save_state, conn, conn_state, i) {

• if (!conn_state->crtc)
  

• 	continue;
  

• conn_state->psr_transition = true;
  

• }
• for_each_new_crtc_in_state(save_state, crtc, crtc_state, i) {

• if (!crtc_state->active)
  

• 	continue;
  

• crtc_state->psr_transition = true;
  

• }
• state->acquire_ctx = &ctx;
• drm_connector_list_iter_begin(psr_state->dev, &conn_iter);
• drm_for_each_connector_iter(conn, &conn_iter) {

• if (!conn->psr_capable)
  

• 	continue;
  

• conn_state = drm_atomic_get_connector_state(state, conn);
  

• if (IS_ERR(conn_state)) {
  

• 	ret = PTR_ERR(conn_state);
  

• 	drm_connector_list_iter_end(&conn_iter);
  

• 	goto out_free_state;
  

• }
  

• if (!conn_state->crtc)
  

• 	continue;
  

• crtc_state = drm_atomic_get_crtc_state(state, conn_state->crtc);
  

• if (IS_ERR(crtc_state)) {
  

• 	ret = PTR_ERR(crtc_state);
  

• 	drm_connector_list_iter_end(&conn_iter);
  

• 	goto out_free_state;
  

• }
  

• /* Don't use PSR if the crtc is driving multiple connectors */
  

• if (hweight_long(crtc_state->connector_mask) > 1)
  

• 	continue;
  

• commit = true;
  

• crtc_state->active = false;
  

• crtc_state->psr_transition = true;
  

• conn_state->psr_transition = true;
  

• }
• drm_connector_list_iter_end(&conn_iter);
• /* Nothing to commit, so just exit */
• if (!commit)

• goto out_free_state;
  

• ret = drm_atomic_commit(state);
• if (ret)

• goto out_free_state;
  

• psr_state->save_state = save_state;
• goto out;

+out_free_state:

• drm_atomic_state_put(save_state);
• drm_atomic_state_put(state);

+out:

• DRM_MODESET_LOCK_ALL_END(ctx, ret);

+}

+static int +drm_psr_helper_acquire_modeset_locks(struct drm_atomic_state *state,

• 		     struct drm_modeset_acquire_ctx *ctx)
  

+{

• struct drm_mode_config *config = &state->dev->mode_config;
• struct drm_crtc *crtc;
• struct drm_crtc_state *new_crtc_state;
• struct drm_plane *plane;
• struct drm_plane_state *new_plane_state;
• int ret, i;
• ret = drm_modeset_lock(&config->connection_mutex, ctx);
• if (ret)

• return ret;
  

• for_each_new_plane_in_state(state, plane, new_plane_state, i) {

• ret = drm_modeset_lock(&plane->mutex, ctx);
  

• if (ret)
  

• 	return ret;
  

• }
• for_each_new_crtc_in_state(state, crtc, new_crtc_state, i) {

• ret = drm_modeset_lock(&crtc->mutex, ctx);
  

• if (ret)
  

• 	return ret;
  

• }
• return 0;

+}

+/**

• • drm_psr_helper_flush - Restore the hardware to pre-PSR state
• • @dev: DRM device
• • @ctx: An acquire context to use for restoring the state
• • This function should be called before every drm_atomic_commit to ensure any
• • connectors that are currently self-refreshing revert back to being bus
• • driven. Drivers may call this function outside of the atomic hooks if
• • they wish to disable PSR pre-emptively (such as upon an input event or when
• • GPU becomes active).
• • If everything is successful, this function will schedule the PSR entry worker
• • to enable PSR after the driver-specified timeout.
• • If the PSR helper is not being used, this is a no-op.
• */

+int drm_psr_helper_flush(struct drm_device *dev,

• 	 struct drm_modeset_acquire_ctx *ctx)
  

+{

• struct drm_mode_config *config = &dev->mode_config;
• struct drm_psr_state *psr_state = config->psr_state;
• int ret;
• if (!psr_state)

• return 0;
  

• ret = drm_modeset_lock(&psr_state->mutex, ctx);
• if (ret)

• return ret;
  

• if (!psr_state->save_state)

• goto out;
  

• ret = drm_psr_helper_acquire_modeset_locks(psr_state->save_state, ctx);
• if (ret)

• goto out;
  

• ret = drm_atomic_helper_commit_duplicated_state(psr_state->save_state,

• 					ctx);
  

+out:

• psr_state->save_state = NULL;
• if (!ret) {

• mod_delayed_work(system_wq, &psr_state->entry_work,
  

• 		 msecs_to_jiffies(psr_state->entry_delay_ms));
  

• }
• return ret;

+} +EXPORT_SYMBOL(drm_psr_helper_flush);

+/**

• • drm_psr_helper_register - Registers a connector with the PSR helpers
• • @connector: the connector which has a PSR-supported display attached
• • Note that this can be called once on initialization for fixed panels, or
• • during enable/hotplug.
• */

+int drm_psr_helper_register(struct drm_connector *connector) +{

• struct drm_mode_config *config = &connector->dev->mode_config;
• struct drm_psr_state *psr_state = config->psr_state;
• /* PSR helpers are uninitialized */
• if (WARN_ON(!psr_state))

• return -EINVAL;
  

• /* Acquired via psr_helper_flush */
• if (!drm_modeset_is_locked(&psr_state->mutex))

• return -EINVAL;
  

• connector->psr_capable = true;
• return 0;

+} +EXPORT_SYMBOL(drm_psr_helper_register);

+/**

• • drm_psr_helper_unregister - Unregisters a connector with the PSR helpers
• • @connector: the connector to unregister
• */

+int drm_psr_helper_unregister(struct drm_connector *connector) +{

• struct drm_mode_config *config = &connector->dev->mode_config;
• struct drm_psr_state *psr_state = config->psr_state;
• /* PSR helpers are uninitialized */
• if (WARN_ON(!psr_state))

• return -EINVAL;
  

• /* Acquired via psr_helper_flush */
• if (!drm_modeset_is_locked(&psr_state->mutex))

• return -EINVAL;
  

• connector->psr_capable = false;
• return 0;

+} +EXPORT_SYMBOL(drm_psr_helper_unregister);

+/**

• • drm_psr_helper_init - Initializes the PSR helpers
• • @dev: DRM device
• • @entry_delay_ms: The amount of time to wait after an atomic commit before

• •     activating PSR
    

• • Drivers using the PSR helpers must call this some time after mode_config
• • is initialized in order to make use of the PSR helpers. Typically
• • entry_delay_ms is a function of how quickly the hardware can enter/exit PSR.
• */

+int drm_psr_helper_init(struct drm_device *dev, unsigned int entry_delay_ms) +{

• struct drm_mode_config *config = &dev->mode_config;
• struct drm_psr_state *psr_state;
• psr_state = kzalloc(sizeof(*psr_state), GFP_KERNEL);
• if (!psr_state)

• return -ENOMEM;
  

• INIT_DELAYED_WORK(&psr_state->entry_work, drm_psr_helper_entry_work);
• drm_modeset_lock_init(&psr_state->mutex);
• psr_state->dev = dev;
• psr_state->entry_delay_ms = entry_delay_ms;
• config->psr_state = psr_state;
• return 0;

+} +EXPORT_SYMBOL(drm_psr_helper_init);

+/**

• • drm_psr_helper_cleanup - De-initializes the PSR helpers
• • @dev: DRM device
• */

+void drm_psr_helper_cleanup(struct drm_device *dev) +{

• struct drm_mode_config *config = &dev->mode_config;
• cancel_delayed_work_sync(&config->psr_state->entry_work);
• kfree(config->psr_state);
• config->psr_state = NULL;

+} +EXPORT_SYMBOL(drm_psr_helper_cleanup); diff --git a/include/drm/drm_connector.h b/include/drm/drm_connector.h index c8061992d6cb..9612b3d56f30 100644 --- a/include/drm/drm_connector.h +++ b/include/drm/drm_connector.h @@ -501,6 +501,17 @@ struct drm_connector_state { /** @tv: TV connector state */ struct drm_tv_connector_state tv;

• /**

• * @psr_transition:
  

• *
  

• * Used by the PSR helpers to denote when a PSR transition is occuring.
  

• * If your connector is PSR-capable, register it with the helpers and
  

• * check this flag in .enable() and .disable(). If it is true, instead
  

• * of shutting off the panel, put it in or take it out of self
  

• * refreshing mode.
  

• */
  

• bool psr_transition;
• /**
  • @picture_aspect_ratio: Connector property to control the
  • HDMI infoframe aspect ratio setting.

@@ -993,6 +1004,17 @@ struct drm_connector { */ struct drm_display_info display_info;

• /**

• * @psr_capable:
  

• *
  

• * Set by the driver via drm_psr_helper_register(). Signals that this
  

• * connector (and associated pipe) is PSR capable and should be put in
  

• * low-power mode when it is inactive.
  

• *
  

• * Protected by &drm_mode_config.psr_state.mutex
  

• */
  

• bool psr_capable;
• /** @funcs: connector control functions */ const struct drm_connector_funcs *funcs;

diff --git a/include/drm/drm_crtc.h b/include/drm/drm_crtc.h index f7c3022dbdf4..10acd4fc0991 100644 --- a/include/drm/drm_crtc.h +++ b/include/drm/drm_crtc.h @@ -299,6 +299,17 @@ struct drm_crtc_state { */ bool vrr_enabled;

• /**

• * @psr_transition:
  

• *
  

• * Used by the PSR helpers to denote when a PSR transition is occuring.
  

• * This will be set on enable/disable callbacks when PSR is being
  

• * enabled or disabled. In some cases, it may not be desirable to fully
  

• * shut off the crtc during PSR. CRTC's can inspect this flag and
  

• * determine the best course of action.
  

• */
  

• bool psr_transition;
• /**
  • @event:

diff --git a/include/drm/drm_mode_config.h b/include/drm/drm_mode_config.h index 7f60e8eb269a..371b80d090ab 100644 --- a/include/drm/drm_mode_config.h +++ b/include/drm/drm_mode_config.h @@ -37,6 +37,7 @@ struct drm_atomic_state; struct drm_mode_fb_cmd2; struct drm_format_info; struct drm_display_mode; +struct drm_psr_state;

/**

• struct drm_mode_config_funcs - basic driver provided mode setting functions

@@ -900,6 +901,11 @@ struct drm_mode_config { */ struct drm_atomic_state *suspend_state;

• /**

• * @psr_state: Holds the state for the psr helper
  

• */
  

• struct drm_psr_state *psr_state;
• const struct drm_mode_config_helper_funcs *helper_private;

};

diff --git a/include/drm/drm_psr_helper.h b/include/drm/drm_psr_helper.h new file mode 100644 index 000000000000..972c4ec98d05 --- /dev/null +++ b/include/drm/drm_psr_helper.h @@ -0,0 +1,24 @@ +/* SPDX-License-Identifier: MIT */ +/*

• • Copyright (C) 2019 Google, Inc.
• • Authors:
• • Sean Paul seanpaul@chromium.org
• */

+#ifndef DRM_PSR_HELPER_H_ +#define DRM_PSR_HELPER_H_

+struct drm_connector; +struct drm_device; +struct drm_psr_state; +struct drm_modeset_acquire_ctx;

+int drm_psr_helper_flush(struct drm_device *dev,

• 	 struct drm_modeset_acquire_ctx *ctx);
  

+int drm_psr_helper_register(struct drm_connector *connector); +int drm_psr_helper_unregister(struct drm_connector *connector);

+int drm_psr_helper_init(struct drm_device *dev, unsigned int entry_delay_ms); +void drm_psr_helper_cleanup(struct drm_device *dev);

+#endif

Sean Paul, Software Engineer, Google / Chromium OS

-- Daniel Vetter Software Engineer, Intel Corporation http://blog.ffwll.ch

-- Sean Paul, Software Engineer, Google / Chromium OS