| Message ID | 1447795121-27332-1-git-send-email-daniel.vetter@ffwll.ch (mailing list archive) |
|---|---|
| State | New, archived |
| Headers | show |
On Tue, Nov 17, 2015 at 10:18:41PM +0100, Daniel Vetter wrote: > If we can't acquire dev->struct_mutex we need to fail runtime suspend, > at least with the current design. Currently we do that using -EAGAIN, > but that upsets the pm core, resulting in the occasional fail testcase > in our CI with the following dmesg dirt: > > pci_pm_runtime_suspend(): intel_runtime_suspend+0x0/0x240 [i915] returns -11 > > Chris has some ideas to improve this, but for now just shut up the > error. > > Cc: Paulo Zanoni <przanoni@gmail.com> > Cc: Chris Wilson <chris@chris-wilson.co.uk> > Signed-off-by: Daniel Vetter <daniel.vetter@intel.com> > --- > drivers/gpu/drm/i915/i915_drv.c | 6 +++--- > 1 file changed, 3 insertions(+), 3 deletions(-) > > diff --git a/drivers/gpu/drm/i915/i915_drv.c b/drivers/gpu/drm/i915/i915_drv.c > index 5a70aca71d6b..ab8ffbc48e2d 100644 > --- a/drivers/gpu/drm/i915/i915_drv.c > +++ b/drivers/gpu/drm/i915/i915_drv.c > @@ -1497,8 +1497,7 @@ static int intel_runtime_suspend(struct device *device) > * We could deadlock here in case another thread holding struct_mutex > * calls RPM suspend concurrently, since the RPM suspend will wait > * first for this RPM suspend to finish. In this case the concurrent > - * RPM resume will be followed by its RPM suspend counterpart. Still > - * for consistency return -EAGAIN, which will reschedule this suspend. > + * RPM resume will be followed by its RPM suspend counterpart. > */ > if (!mutex_trylock(&dev->struct_mutex)) { > DRM_DEBUG_KMS("device lock contention, deffering suspend\n"); > @@ -1508,7 +1507,8 @@ static int intel_runtime_suspend(struct device *device) > */ > pm_runtime_mark_last_busy(device); > > - return -EAGAIN; > + /* Fail silently to avoid upsetting the pm core. */ > + return 0; So the core will assume we're now suspended and then resume gets called while we're still powered on. Sounds like a bad plan to me. I'm especially worried about VLV here with its GT no wake dance and manual save/restore. > } > /* > * We are safe here against re-faults, since the fault handler takes > -- > 2.5.1 > > _______________________________________________ > Intel-gfx mailing list > Intel-gfx@lists.freedesktop.org > http://lists.freedesktop.org/mailman/listinfo/intel-gfx
On Tue, 2015-11-17 at 23:30 +0200, Ville Syrjälä wrote: > On Tue, Nov 17, 2015 at 10:18:41PM +0100, Daniel Vetter wrote: > > If we can't acquire dev->struct_mutex we need to fail runtime suspend, > > at least with the current design. Currently we do that using -EAGAIN, > > but that upsets the pm core, resulting in the occasional fail testcase > > in our CI with the following dmesg dirt: > > > > pci_pm_runtime_suspend(): intel_runtime_suspend+0x0/0x240 [i915] returns -11 > > > > Chris has some ideas to improve this, but for now just shut up the > > error. > > > > Cc: Paulo Zanoni <przanoni@gmail.com> > > Cc: Chris Wilson <chris@chris-wilson.co.uk> > > Signed-off-by: Daniel Vetter <daniel.vetter@intel.com> > > --- > > drivers/gpu/drm/i915/i915_drv.c | 6 +++--- > > 1 file changed, 3 insertions(+), 3 deletions(-) > > > > diff --git a/drivers/gpu/drm/i915/i915_drv.c b/drivers/gpu/drm/i915/i915_drv.c > > index 5a70aca71d6b..ab8ffbc48e2d 100644 > > --- a/drivers/gpu/drm/i915/i915_drv.c > > +++ b/drivers/gpu/drm/i915/i915_drv.c > > @@ -1497,8 +1497,7 @@ static int intel_runtime_suspend(struct device *device) > > * We could deadlock here in case another thread holding struct_mutex > > * calls RPM suspend concurrently, since the RPM suspend will wait > > * first for this RPM suspend to finish. In this case the concurrent > > - * RPM resume will be followed by its RPM suspend counterpart. Still > > - * for consistency return -EAGAIN, which will reschedule this suspend. > > + * RPM resume will be followed by its RPM suspend counterpart. > > */ > > if (!mutex_trylock(&dev->struct_mutex)) { > > DRM_DEBUG_KMS("device lock contention, deffering suspend\n"); > > @@ -1508,7 +1507,8 @@ static int intel_runtime_suspend(struct device *device) > > */ > > pm_runtime_mark_last_busy(device); > > > > - return -EAGAIN; > > + /* Fail silently to avoid upsetting the pm core. */ > > + return 0; > > So the core will assume we're now suspended and then resume gets called > while we're still powered on. Sounds like a bad plan to me. I'm > especially worried about VLV here with its GT no wake dance and manual > save/restore. Also the PCI core will put the device into D3 if we report success. > > > } > > /* > > * We are safe here against re-faults, since the fault handler takes > > -- > > 2.5.1 > > > > _______________________________________________ > > Intel-gfx mailing list > > Intel-gfx@lists.freedesktop.org > > http://lists.freedesktop.org/mailman/listinfo/intel-gfx >
On Tue, Nov 17, 2015 at 11:35:33PM +0200, Imre Deak wrote: > On Tue, 2015-11-17 at 23:30 +0200, Ville Syrjälä wrote: > > On Tue, Nov 17, 2015 at 10:18:41PM +0100, Daniel Vetter wrote: > > > If we can't acquire dev->struct_mutex we need to fail runtime suspend, > > > at least with the current design. Currently we do that using -EAGAIN, > > > but that upsets the pm core, resulting in the occasional fail testcase > > > in our CI with the following dmesg dirt: > > > > > > pci_pm_runtime_suspend(): intel_runtime_suspend+0x0/0x240 [i915] returns -11 > > > > > > Chris has some ideas to improve this, but for now just shut up the > > > error. > > > > > > Cc: Paulo Zanoni <przanoni@gmail.com> > > > Cc: Chris Wilson <chris@chris-wilson.co.uk> > > > Signed-off-by: Daniel Vetter <daniel.vetter@intel.com> > > > --- > > > drivers/gpu/drm/i915/i915_drv.c | 6 +++--- > > > 1 file changed, 3 insertions(+), 3 deletions(-) > > > > > > diff --git a/drivers/gpu/drm/i915/i915_drv.c b/drivers/gpu/drm/i915/i915_drv.c > > > index 5a70aca71d6b..ab8ffbc48e2d 100644 > > > --- a/drivers/gpu/drm/i915/i915_drv.c > > > +++ b/drivers/gpu/drm/i915/i915_drv.c > > > @@ -1497,8 +1497,7 @@ static int intel_runtime_suspend(struct device *device) > > > * We could deadlock here in case another thread holding struct_mutex > > > * calls RPM suspend concurrently, since the RPM suspend will wait > > > * first for this RPM suspend to finish. In this case the concurrent > > > - * RPM resume will be followed by its RPM suspend counterpart. Still > > > - * for consistency return -EAGAIN, which will reschedule this suspend. > > > + * RPM resume will be followed by its RPM suspend counterpart. > > > */ > > > if (!mutex_trylock(&dev->struct_mutex)) { > > > DRM_DEBUG_KMS("device lock contention, deffering suspend\n"); > > > @@ -1508,7 +1507,8 @@ static int intel_runtime_suspend(struct device *device) > > > */ > > > pm_runtime_mark_last_busy(device); > > > > > > - return -EAGAIN; > > > + /* Fail silently to avoid upsetting the pm core. */ > > > + return 0; > > > > So the core will assume we're now suspended and then resume gets called > > while we're still powered on. Sounds like a bad plan to me. I'm > > especially worried about VLV here with its GT no wake dance and manual > > save/restore. > > Also the PCI core will put the device into D3 if we report success. Oh right. Somehow I remembered that it was for system suspend only, but it's there for runtime pm as well. So I think if we want to hide the dmesg spew temporarily, the only sane option is to reduce the loglevel in __suspend_report_result(). > > > > > > } > > > /* > > > * We are safe here against re-faults, since the fault handler takes > > > -- > > > 2.5.1 > > > > > > _______________________________________________ > > > Intel-gfx mailing list > > > Intel-gfx@lists.freedesktop.org > > > http://lists.freedesktop.org/mailman/listinfo/intel-gfx > >
diff --git a/drivers/gpu/drm/i915/i915_drv.c b/drivers/gpu/drm/i915/i915_drv.c index 5a70aca71d6b..ab8ffbc48e2d 100644 --- a/drivers/gpu/drm/i915/i915_drv.c +++ b/drivers/gpu/drm/i915/i915_drv.c @@ -1497,8 +1497,7 @@ static int intel_runtime_suspend(struct device *device) * We could deadlock here in case another thread holding struct_mutex * calls RPM suspend concurrently, since the RPM suspend will wait * first for this RPM suspend to finish. In this case the concurrent - * RPM resume will be followed by its RPM suspend counterpart. Still - * for consistency return -EAGAIN, which will reschedule this suspend. + * RPM resume will be followed by its RPM suspend counterpart. */ if (!mutex_trylock(&dev->struct_mutex)) { DRM_DEBUG_KMS("device lock contention, deffering suspend\n"); @@ -1508,7 +1507,8 @@ static int intel_runtime_suspend(struct device *device) */ pm_runtime_mark_last_busy(device); - return -EAGAIN; + /* Fail silently to avoid upsetting the pm core. */ + return 0; } /* * We are safe here against re-faults, since the fault handler takes
If we can't acquire dev->struct_mutex we need to fail runtime suspend, at least with the current design. Currently we do that using -EAGAIN, but that upsets the pm core, resulting in the occasional fail testcase in our CI with the following dmesg dirt: pci_pm_runtime_suspend(): intel_runtime_suspend+0x0/0x240 [i915] returns -11 Chris has some ideas to improve this, but for now just shut up the error. Cc: Paulo Zanoni <przanoni@gmail.com> Cc: Chris Wilson <chris@chris-wilson.co.uk> Signed-off-by: Daniel Vetter <daniel.vetter@intel.com> --- drivers/gpu/drm/i915/i915_drv.c | 6 +++--- 1 file changed, 3 insertions(+), 3 deletions(-)