Re: [Intel-gfx] [PATCH v2] drm/i915: remove user GTT mappings early during runtime suspend
On 07/05/2014 17:57, Imre Deak wrote: Currently user space can access GEM buffers mapped to GTT through existing mappings concurrently while the platform specific suspend handlers are running. Since these handlers may change the HW state in a way that would break such accesses, remove the mappings before calling the handlers. Spotted by Ville. Also Chris pointed out that the lists that i915_gem_release_all_mmaps() walks through need dev-struct_mutex, so take this lock. There is a potential deadlock against a concurrent RPM resume, resolve this by aborting and rescheduling the suspend (Daniel). v2: - take struct_mutex around i915_gem_release_all_mmaps() (Chris, Daniel) Signed-off-by: Imre Deak imre.d...@intel.com --- drivers/gpu/drm/i915/i915_drv.c | 27 +-- 1 file changed, 25 insertions(+), 2 deletions(-) diff --git a/drivers/gpu/drm/i915/i915_drv.c b/drivers/gpu/drm/i915/i915_drv.c index 4024e16..0c9858c 100644 --- a/drivers/gpu/drm/i915/i915_drv.c +++ b/drivers/gpu/drm/i915/i915_drv.c @@ -36,6 +36,7 @@ #include linux/console.h #include linux/module.h +#include linux/pm_runtime.h #include drm/drm_crtc_helper.h static struct drm_driver driver; @@ -1315,6 +1316,30 @@ static int intel_runtime_suspend(struct device *device) DRM_DEBUG_KMS(Suspending device\n); /* +* 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. +*/ + if (!mutex_trylock(dev-struct_mutex)) { + DRM_DEBUG_KMS(device lock contention, deffering suspend\n); + /* +* Bump the expiration timestamp, otherwise the suspend won't +* be rescheduled. +*/ + pm_runtime_mark_last_busy(device); + + return -EAGAIN; + } + /* +* We are safe here against re-faults, since the fault handler takes +* an RPM reference. +*/ + i915_gem_release_all_mmaps(dev_priv); + mutex_unlock(dev-struct_mutex); + + /* * rps.work can't be rearmed here, since we get here only after making * sure the GPU is idle and the RPS freq is set to the minimum. See * intel_mark_idle(). @@ -1340,8 +1365,6 @@ static int intel_runtime_suspend(struct device *device) return ret; } - i915_gem_release_all_mmaps(dev_priv); - del_timer_sync(dev_priv-gpu_error.hangcheck_timer); dev_priv-pm.suspended = true; Looks good to me. Reviewed-by: Robert Beckett robert.beck...@intel.com ___ Intel-gfx mailing list Intel-gfx@lists.freedesktop.org http://lists.freedesktop.org/mailman/listinfo/intel-gfx
[Intel-gfx] [PATCH v2] drm/i915: remove user GTT mappings early during runtime suspend
Currently user space can access GEM buffers mapped to GTT through existing mappings concurrently while the platform specific suspend handlers are running. Since these handlers may change the HW state in a way that would break such accesses, remove the mappings before calling the handlers. Spotted by Ville. Also Chris pointed out that the lists that i915_gem_release_all_mmaps() walks through need dev-struct_mutex, so take this lock. There is a potential deadlock against a concurrent RPM resume, resolve this by aborting and rescheduling the suspend (Daniel). v2: - take struct_mutex around i915_gem_release_all_mmaps() (Chris, Daniel) Signed-off-by: Imre Deak imre.d...@intel.com --- drivers/gpu/drm/i915/i915_drv.c | 27 +-- 1 file changed, 25 insertions(+), 2 deletions(-) diff --git a/drivers/gpu/drm/i915/i915_drv.c b/drivers/gpu/drm/i915/i915_drv.c index 4024e16..0c9858c 100644 --- a/drivers/gpu/drm/i915/i915_drv.c +++ b/drivers/gpu/drm/i915/i915_drv.c @@ -36,6 +36,7 @@ #include linux/console.h #include linux/module.h +#include linux/pm_runtime.h #include drm/drm_crtc_helper.h static struct drm_driver driver; @@ -1315,6 +1316,30 @@ static int intel_runtime_suspend(struct device *device) DRM_DEBUG_KMS(Suspending device\n); /* +* 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. +*/ + if (!mutex_trylock(dev-struct_mutex)) { + DRM_DEBUG_KMS(device lock contention, deffering suspend\n); + /* +* Bump the expiration timestamp, otherwise the suspend won't +* be rescheduled. +*/ + pm_runtime_mark_last_busy(device); + + return -EAGAIN; + } + /* +* We are safe here against re-faults, since the fault handler takes +* an RPM reference. +*/ + i915_gem_release_all_mmaps(dev_priv); + mutex_unlock(dev-struct_mutex); + + /* * rps.work can't be rearmed here, since we get here only after making * sure the GPU is idle and the RPS freq is set to the minimum. See * intel_mark_idle(). @@ -1340,8 +1365,6 @@ static int intel_runtime_suspend(struct device *device) return ret; } - i915_gem_release_all_mmaps(dev_priv); - del_timer_sync(dev_priv-gpu_error.hangcheck_timer); dev_priv-pm.suspended = true; -- 1.8.4 ___ Intel-gfx mailing list Intel-gfx@lists.freedesktop.org http://lists.freedesktop.org/mailman/listinfo/intel-gfx
Re: [Intel-gfx] [PATCH v2] drm/i915: remove user GTT mappings early during runtime suspend
On Wed, 2014-05-07 at 19:57 +0300, Imre Deak wrote: Currently user space can access GEM buffers mapped to GTT through existing mappings concurrently while the platform specific suspend handlers are running. Since these handlers may change the HW state in a way that would break such accesses, remove the mappings before calling the handlers. Spotted by Ville. Also Chris pointed out that the lists that i915_gem_release_all_mmaps() walks through need dev-struct_mutex, so take this lock. There is a potential deadlock against a concurrent RPM resume, resolve this by aborting and rescheduling the suspend (Daniel). v2: - take struct_mutex around i915_gem_release_all_mmaps() (Chris, Daniel) Signed-off-by: Imre Deak imre.d...@intel.com --- drivers/gpu/drm/i915/i915_drv.c | 27 +-- 1 file changed, 25 insertions(+), 2 deletions(-) diff --git a/drivers/gpu/drm/i915/i915_drv.c b/drivers/gpu/drm/i915/i915_drv.c index 4024e16..0c9858c 100644 --- a/drivers/gpu/drm/i915/i915_drv.c +++ b/drivers/gpu/drm/i915/i915_drv.c @@ -36,6 +36,7 @@ #include linux/console.h #include linux/module.h +#include linux/pm_runtime.h #include drm/drm_crtc_helper.h static struct drm_driver driver; @@ -1315,6 +1316,30 @@ static int intel_runtime_suspend(struct device *device) DRM_DEBUG_KMS(Suspending device\n); /* + * We could deadlock here in case another thread holding struct_mutex + * calls RPM suspend concurrently, since the RPM suspend will wait resume^ resume^ + * 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. + */ + if (!mutex_trylock(dev-struct_mutex)) { + DRM_DEBUG_KMS(device lock contention, deffering suspend\n); + /* + * Bump the expiration timestamp, otherwise the suspend won't + * be rescheduled. + */ + pm_runtime_mark_last_busy(device); + + return -EAGAIN; + } + /* + * We are safe here against re-faults, since the fault handler takes + * an RPM reference. + */ + i915_gem_release_all_mmaps(dev_priv); + mutex_unlock(dev-struct_mutex); + + /* * rps.work can't be rearmed here, since we get here only after making * sure the GPU is idle and the RPS freq is set to the minimum. See * intel_mark_idle(). @@ -1340,8 +1365,6 @@ static int intel_runtime_suspend(struct device *device) return ret; } - i915_gem_release_all_mmaps(dev_priv); - del_timer_sync(dev_priv-gpu_error.hangcheck_timer); dev_priv-pm.suspended = true; signature.asc Description: This is a digitally signed message part ___ Intel-gfx mailing list Intel-gfx@lists.freedesktop.org http://lists.freedesktop.org/mailman/listinfo/intel-gfx