On Tue, Jul 17, 2018 at 09:33:52AM +0200, Lukas Wunner wrote:

On Thu, Jul 12, 2018 at 01:02:53PM -0400, Lyude Paul wrote:

...

--- a/drivers/gpu/drm/nouveau/dispnv50/disp.c +++ b/drivers/gpu/drm/nouveau/dispnv50/disp.c @@ -1878,7 +1878,7 @@ nv50_disp_atomic_commit(struct drm_device *dev, nv50_disp_atomic_commit_tail(state);

drm_for_each_crtc(crtc, dev) {

• if (crtc->state->enable) {
  

• if (crtc->state->active) {
  if (!drm->have_disp_power_ref) {
  	drm->have_disp_power_ref = true;
  	return 0;
  

Somewhat tangential comment on this older patch, since you continue to dig around in the runtime PM area:

Whenever a crtc is activated or deactivated in nouveau, we iterate over all crtcs and acquire a runtime PM if a crtc is active and previously there was no active one, or we drop a ref if none is active and previously there was an active one.

For a while now I've been thinking that it would be more straightforward to acquire a ref whenever a crtc is activated and drop one when a crtc is deactivated, i.e. hold one ref for every active crtc. That way the have_disp_power_ref variable as well as the iteration logic could be removed, leading to a simplification. Just a suggestion anyway.

The current code looks somewhat busted anyway. First problem is that it's accessing crtc->state without the appropriate locks held (unless something always pulls in all crtcs to every commit?). Second issue is that the rpm_put() is called without waiting for nonblocking commits to have finished so it looks like you can currentlly remove the power before the hardware has been properly shut down.

On Thu, Jul 12, 2018 at 01:02:54PM -0400, Lyude Paul wrote:

This both uses the legacy modesetting structures in a racy manner, and additionally also doesn't even check the right variable (enabled != the CRTC is actually turned on for atomic).

This fixes issues on my P50 regarding the dedicated GPU not entering runtime suspend.

Signed-off-by: Lyude Paul lyude@redhat.com Cc: stable@vger.kernel.org

On both patch 2&3:

Reviewed-by: Daniel Vetter daniel.vetter@ffwll.ch

->enable vs. ->active is probably the biggest source of pain in atomic, and beyond typing even more kerneldoc that will be ignored (there's another series doing exactly that on the list) I have no idea what to do. 90% rule is to look at ->enable in atomic_check code (since DPMS changes should always work) and ->active in atomic_commit code.

Wrt the legacy state: For the legacy pointers we can set them to NULL for atomic, and Ville has done that. That's real effective at stopping drivers from looking at the wrong thing. But for the others like this one here I dunno what to do to effectively hide them from atomic drivers.

</rant>

Cheers, Daniel

---

drivers/gpu/drm/nouveau/nouveau_drm.c | 11 ----------- 1 file changed, 11 deletions(-)

diff --git a/drivers/gpu/drm/nouveau/nouveau_drm.c b/drivers/gpu/drm/nouveau/nouveau_drm.c index 0f668e275ee1..c7ec86d6c3c9 100644 --- a/drivers/gpu/drm/nouveau/nouveau_drm.c +++ b/drivers/gpu/drm/nouveau/nouveau_drm.c @@ -881,22 +881,11 @@ nouveau_pmops_runtime_resume(struct device *dev) static int nouveau_pmops_runtime_idle(struct device *dev) {

• struct pci_dev *pdev = to_pci_dev(dev);

• struct drm_device *drm_dev = pci_get_drvdata(pdev);

• struct nouveau_drm *drm = nouveau_drm(drm_dev);

• struct drm_crtc *crtc;

• if (!nouveau_pmops_runtime()) { pm_runtime_forbid(dev); return -EBUSY; }

• list_for_each_entry(crtc, &drm->dev->mode_config.crtc_list, head) {

• if (crtc->enabled) {
  

• 	DRM_DEBUG_DRIVER("failing to power off - crtc active\n");
  

• 	return -EBUSY;
  

• }
  

• } pm_runtime_mark_last_busy(dev); pm_runtime_autosuspend(dev); /* we don't want the main rpm_idle to call suspend - we want to autosuspend */

-- 2.17.1

---

dri-devel mailing list dri-devel@lists.freedesktop.org https://lists.freedesktop.org/mailman/listinfo/dri-devel