drm/nouveau: Don't enabling polling twice on runtime resume

PATCH New Lyude 2 participants 3 messages Jan 12, 2017

Subsystems (3)

DRM drivers DRM misc nouveau

lore

Share & tools

lore

Lyude

author maintainer

2017-01-12 02:25 UTC #0 lore raw

PATCH v2 1/2 drm/nouveau: Don't enabling polling twice on runtime resume

As it turns out, on cards that actually have CRTCs on them we're already calling drm_kms_helper_poll_enable(drm_dev) from nouveau_display_resume() before we call it in nouveau_pmops_runtime_resume(). This leads us to accidentally trying to enable polling twice, which results in a potential deadlock between the RPM locks and drm_dev->mode_config.mutex if we end up trying to enable polling the second time while output_poll_execute is running and holding the mode_config lock. As such, make sure we only enable polling in nouveau_pmops_runtime_resume() if we need to.

This fixes hangs observed on the ThinkPad W541

Signed-off-by Lyude <[email protected]>

Cc Hans de Goede <[email protected]>

Cc Kilian Singer <[email protected]>

Cc Lukas Wunner <[email protected]>

Cc David Airlie <[email protected]>

---
Changes since v1:
- Rebase to work with master

 drivers/gpu/drm/nouveau/nouveau_display.c | 3 ++-
 drivers/gpu/drm/nouveau/nouveau_drm.c     | 5 ++++-
 2 files changed, 6 insertions(+), 2 deletions(-)

drivers/gpu/drm/nouveau/nouveau_display.c +2 -1

index cef08da..6a15776 100644
@@ -411,7 +411,8 @@ nouveau_display_init(struct drm_device *dev)
 		return ret;
 
 	/* enable polling for external displays */
-	drm_kms_helper_poll_enable(dev);
+	if (!dev->mode_config.poll_enabled)
+		drm_kms_helper_poll_enable(dev);
 
 	/* enable hotplug interrupts */
 	list_for_each_entry(connector, &dev->mode_config.connector_list, head) {

drivers/gpu/drm/nouveau/nouveau_drm.c +4 -1

index 59348fc..bc85a45 100644
@@ -773,7 +773,10 @@ nouveau_pmops_runtime_resume(struct device *dev)
 	pci_set_master(pdev);
 
 	ret = nouveau_do_resume(drm_dev, true);
-	drm_kms_helper_poll_enable(drm_dev);
+
+	if (!drm_dev->mode_config.poll_enabled)
+		drm_kms_helper_poll_enable(drm_dev);
+
 	/* do magic */
 	nvif_mask(&device->object, 0x088488, (1 << 25), (1 << 25));
 	vga_switcheroo_set_dynamic_switch(pdev, VGA_SWITCHEROO_ON);

Signature · 1 line

2.9.3

Lyude

author maintainer

2017-01-12 02:25 UTC #1 lore raw

PATCH 2/2 drm/nouveau: Handle fbcon suspend/resume in seperate worker

Resuming from RPM can happen while already holding dev->mode_config.mutex. This means we can't actually handle fbcon in any RPM resume workers, since restoring fbcon requires grabbing dev->mode_config.mutex again. So move the fbcon suspend/resume code into it's own worker, and rely on that instead to avoid deadlocking.

This fixes more deadlocks for runtime suspending the GPU on the ThinkPad W541. Reproduction recipe:

- Get a machine with both optimus and a nvidia card with connectors attached to it
- Wait for the nvidia GPU to suspend
- Attempt to manually reprobe any of the connectors on the nvidia GPU using sysfs
- *deadlock*

Signed-off-by Lyude <[email protected]>

Cc Hans de Goede <[email protected]>

Cc Kilian Singer <[email protected]>

Cc Lukas Wunner <[email protected]>

Cc David Airlie <[email protected]>

---
 drivers/gpu/drm/nouveau/nouveau_drv.h   |  2 ++
 drivers/gpu/drm/nouveau/nouveau_fbcon.c | 43 ++++++++++++++++++++++++++-------
 2 files changed, 36 insertions(+), 9 deletions(-)

drivers/gpu/drm/nouveau/nouveau_drv.h +2 -0

index 8d5ed5b..42c1fa5 100644
@@ -165,6 +165,8 @@ struct nouveau_drm {
 	struct backlight_device *backlight;
 	struct list_head bl_connectors;
 	struct work_struct hpd_work;
+	struct work_struct fbcon_work;
+	int fbcon_new_state;
 #ifdef CONFIG_ACPI
 	struct notifier_block acpi_nb;
 #endif

drivers/gpu/drm/nouveau/nouveau_fbcon.c +34 -9

59 lines not on this page · raw message

Signature · 6 lines

2.9.3

_______________________________________________
Nouveau mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/nouveau

Hans de Goede

2017-01-12 08:27 UTC #2 lore raw

Re: [PATCH 2/2] drm/nouveau: Handle fbcon suspend/resume in seperate worker

Hi,

Good catch (both the previous patch as well as this one). I've one small comment inline:

Quoted 44 lines · Lyude

Resuming from RPM can happen while already holding dev->mode_config.mutex. This means we can't actually handle fbcon in any RPM resume workers, since restoring fbcon requires grabbing dev->mode_config.mutex again. So move the fbcon suspend/resume code into it's own worker, and rely on that instead to avoid deadlocking.
This fixes more deadlocks for runtime suspending the GPU on the ThinkPad W541. Reproduction recipe:
- Get a machine with both optimus and a nvidia card with connectors attached to it
- Wait for the nvidia GPU to suspend
- Attempt to manually reprobe any of the connectors on the nvidia GPU using sysfs
- *deadlock*
Signed-off-by Lyude <[email protected]>
Cc Hans de Goede <[email protected]>
Cc Kilian Singer <[email protected]>
Cc Lukas Wunner <[email protected]>
Cc David Airlie <[email protected]>
---
 drivers/gpu/drm/nouveau/nouveau_drv.h   |  2 ++
 drivers/gpu/drm/nouveau/nouveau_fbcon.c | 43 ++++++++++++++++++++++++++-------
 2 files changed, 36 insertions(+), 9 deletions(-)
drivers/gpu/drm/nouveau/nouveau_drv.h +2 -0
index 8d5ed5b..42c1fa5 100644
@@ -165,6 +165,8 @@ struct nouveau_drm {
 	struct backlight_device *backlight;
 	struct list_head bl_connectors;
 	struct work_struct hpd_work;
+	struct work_struct fbcon_work;
+	int fbcon_new_state;
 #ifdef CONFIG_ACPI
 	struct notifier_block acpi_nb;
 #endif
drivers/gpu/drm/nouveau/nouveau_fbcon.c +0 -0
index 2f2a3dc..87cd30b 100644

drivers/gpu/drm/nouveau/nouveau_fbcon.c static const struct drm_fb_helper_funcs nouveau_fbcon_helper_funcs = { · line 477

Quoted 8 lines

drivers/gpu/drm/nouveau/nouveau_fbcon.c +5 -0

@@ -470,19 +470,43 @@ static const struct drm_fb_helper_funcs nouveau_fbcon_helper_funcs = {
 	.fb_probe = nouveau_fbcon_create,
 };
 
+static void
+nouveau_fbcon_set_suspend_work(struct work_struct *work)
+{
+	struct nouveau_drm *drm = container_of(work, typeof(*drm), fbcon_work);
+	int state = drm->fbcon_new_state;

The compiler may decide to optimize away this variable and simply use drm->fbcon_new_state in the if-s below, which is racy. I would fix this by making drm->fbcon_new_state an atomic_t and using

atomic_read(&drm->fbcon_new_state) here.
(and atomic_set below).

Regards,

Hans

Quoted 53 lines

drivers/gpu/drm/nouveau/nouveau_fbcon.c +29 -9

+
+	if (state == FBINFO_STATE_RUNNING)
+		pm_runtime_get_sync(drm->dev->dev);
+
+	console_lock();
+	if (state == FBINFO_STATE_RUNNING)
+		nouveau_fbcon_accel_restore(drm->dev);
+	drm_fb_helper_set_suspend(&drm->fbcon->helper, state);
+	if (state != FBINFO_STATE_RUNNING)
+		nouveau_fbcon_accel_save_disable(drm->dev);
+	console_unlock();
+
+	if (state == FBINFO_STATE_RUNNING) {
+		pm_runtime_mark_last_busy(drm->dev->dev);
+		pm_runtime_put_sync(drm->dev->dev);
+	}
+}
+
 void
 nouveau_fbcon_set_suspend(struct drm_device *dev, int state)
 {
 	struct nouveau_drm *drm = nouveau_drm(dev);
-	if (drm->fbcon) {
-		console_lock();
-		if (state == FBINFO_STATE_RUNNING)
-			nouveau_fbcon_accel_restore(dev);
-		drm_fb_helper_set_suspend(&drm->fbcon->helper, state);
-		if (state != FBINFO_STATE_RUNNING)
-			nouveau_fbcon_accel_save_disable(dev);
-		console_unlock();
-	}
+
+	if (!drm->fbcon)
+		return;
+
+	drm->fbcon_new_state = state;
+	/* Since runtime resume can happen as a result of a sysfs operation,
+	 * it's possible we already have the console locked. So handle fbcon
+	 * init/deinit from a seperate work thread
+	 */
+	schedule_work(&drm->fbcon_work);
 }
 
 int

@@ -502,6 +526,7 @@ nouveau_fbcon_init(struct drm_device *dev)
 		return -ENOMEM;
 
 	drm->fbcon = fbcon;
+	INIT_WORK(&drm->fbcon_work, nouveau_fbcon_set_suspend_work);
 
 	drm_fb_helper_prepare(dev, &fbcon->helper, &nouveau_fbcon_helper_funcs);