Re: [PATCH v7 1/3] drm/fence: add in-fences support
On Tue, Nov 08, 2016 at 03:27:14PM +, Brian Starkey wrote: > Hi Gustavo, > > On Tue, Nov 08, 2016 at 03:54:48PM +0900, Gustavo Padovan wrote: > > From: Gustavo Padovan > > > > There is now a new property called IN_FENCE_FD attached to every plane > > state that receives sync_file fds from userspace via the atomic commit > > IOCTL. > > > > The fd is then translated to a fence (that may be a fence_array > > subclass or just a normal fence) and then used by DRM to fence_wait() for > > all fences in the sync_file to signal. So it only commits when all > > framebuffers are ready to scanout. > > > > v2: Comments by Daniel Vetter: > > - remove set state->fence = NULL in destroy phase > > - accept fence -1 as valid and just return 0 > > - do not call fence_get() - sync_file_fences_get() already calls it > > - fence_put() if state->fence is already set, in case userspace > > set the property more than once. > > > > v3: WARN_ON if fence is set but state has no FB > > > > v4: Comment from Maarten Lankhorst > > - allow set fence with no related fb > > > > v5: rename FENCE_FD to IN_FENCE_FD > > > > v6: Comments by Daniel Vetter: > > - rename plane_state->in_fence back to "fence" > > - re-introduce WARN_ON if fence set but no fb > > > > - rebase after fence -> dma_fence rename > > > > Signed-off-by: Gustavo Padovan > > --- > > drivers/gpu/drm/Kconfig | 1 + > > drivers/gpu/drm/drm_atomic.c| 14 ++ > > drivers/gpu/drm/drm_atomic_helper.c | 3 +++ > > drivers/gpu/drm/drm_crtc.c | 6 ++ > > drivers/gpu/drm/drm_plane.c | 1 + > > include/drm/drm_crtc.h | 5 + > > 6 files changed, 30 insertions(+) > > > > diff --git a/drivers/gpu/drm/Kconfig b/drivers/gpu/drm/Kconfig > > index 483059a..43cb33d 100644 > > --- a/drivers/gpu/drm/Kconfig > > +++ b/drivers/gpu/drm/Kconfig > > @@ -12,6 +12,7 @@ menuconfig DRM > > select I2C > > select I2C_ALGOBIT > > select DMA_SHARED_BUFFER > > + select SYNC_FILE > > help > > Kernel-level support for the Direct Rendering Infrastructure (DRI) > > introduced in XFree86 4.0. If you say Y here, you need to select > > diff --git a/drivers/gpu/drm/drm_atomic.c b/drivers/gpu/drm/drm_atomic.c > > index 5e73954..d1ae463 100644 > > --- a/drivers/gpu/drm/drm_atomic.c > > +++ b/drivers/gpu/drm/drm_atomic.c > > @@ -30,6 +30,7 @@ > > #include > > #include > > #include > > +#include > > > > #include "drm_crtc_internal.h" > > > > @@ -686,6 +687,17 @@ int drm_atomic_plane_set_property(struct drm_plane > > *plane, > > drm_atomic_set_fb_for_plane(state, fb); > > if (fb) > > drm_framebuffer_unreference(fb); > > + } else if (property == config->prop_in_fence_fd) { > > + if (U642I64(val) == -1) > > + return 0; > > + > > Sorry, I'm a bit late to bring this up but what's the expected > behaviour of a commit which sets OUT_FENCE twice (first with a valid > fence, then with -1 later in the list)? Maybe you need to actually > clear state->fence for -1. Good point, we should probaly reject that. Since ->fence should autoclear (I totally missed that one) we should check for an existing fence and fail if it's set already. > > + if (state->fence) > > + dma_fence_put(state->fence); > > + > > + state->fence = sync_file_get_fence(val); > > + if (!state->fence) > > + return -EINVAL; > > + > > } else if (property == config->prop_crtc_id) { > > struct drm_crtc *crtc = drm_crtc_find(dev, val); > > return drm_atomic_set_crtc_for_plane(state, crtc); > > @@ -747,6 +759,8 @@ drm_atomic_plane_get_property(struct drm_plane *plane, > > > > if (property == config->prop_fb_id) { > > *val = (state->fb) ? state->fb->base.id : 0; > > + } else if (property == config->prop_in_fence_fd) { > > + *val = -1; > > } else if (property == config->prop_crtc_id) { > > *val = (state->crtc) ? state->crtc->base.id : 0; > > } else if (property == config->prop_crtc_x) { > > diff --git a/drivers/gpu/drm/drm_atomic_helper.c > > b/drivers/gpu/drm/drm_atomic_helper.c > > index 75ad01d..26a067f 100644 > > --- a/drivers/gpu/drm/drm_atomic_helper.c > > +++ b/drivers/gpu/drm/drm_atomic_helper.c > > @@ -3114,6 +3114,9 @@ void __drm_atomic_helper_plane_destroy_state(struct > > drm_plane_state *state) > > { > > if (state->fb) > > drm_framebuffer_unreference(state->fb); > > + > > + if (state->fence) > > + dma_fence_put(state->fence); > > } > > EXPORT_SYMBOL(__drm_atomic_helper_plane_destroy_state); > > > > It looks like you need to add something in > __drm_atomic_helper_plane_duplicate_state() to either get a reference > on the fence or set state->fence = NULL, because the memcpy() there > will copy the pointer. Clearing to NULL would be best, since that matches the mea
Re: [PATCH v7 1/3] drm/fence: add in-fences support
Hi Gustavo, On Tue, Nov 08, 2016 at 03:54:48PM +0900, Gustavo Padovan wrote: From: Gustavo Padovan There is now a new property called IN_FENCE_FD attached to every plane state that receives sync_file fds from userspace via the atomic commit IOCTL. The fd is then translated to a fence (that may be a fence_array subclass or just a normal fence) and then used by DRM to fence_wait() for all fences in the sync_file to signal. So it only commits when all framebuffers are ready to scanout. v2: Comments by Daniel Vetter: - remove set state->fence = NULL in destroy phase - accept fence -1 as valid and just return 0 - do not call fence_get() - sync_file_fences_get() already calls it - fence_put() if state->fence is already set, in case userspace set the property more than once. v3: WARN_ON if fence is set but state has no FB v4: Comment from Maarten Lankhorst - allow set fence with no related fb v5: rename FENCE_FD to IN_FENCE_FD v6: Comments by Daniel Vetter: - rename plane_state->in_fence back to "fence" - re-introduce WARN_ON if fence set but no fb - rebase after fence -> dma_fence rename Signed-off-by: Gustavo Padovan --- drivers/gpu/drm/Kconfig | 1 + drivers/gpu/drm/drm_atomic.c| 14 ++ drivers/gpu/drm/drm_atomic_helper.c | 3 +++ drivers/gpu/drm/drm_crtc.c | 6 ++ drivers/gpu/drm/drm_plane.c | 1 + include/drm/drm_crtc.h | 5 + 6 files changed, 30 insertions(+) diff --git a/drivers/gpu/drm/Kconfig b/drivers/gpu/drm/Kconfig index 483059a..43cb33d 100644 --- a/drivers/gpu/drm/Kconfig +++ b/drivers/gpu/drm/Kconfig @@ -12,6 +12,7 @@ menuconfig DRM select I2C select I2C_ALGOBIT select DMA_SHARED_BUFFER + select SYNC_FILE help Kernel-level support for the Direct Rendering Infrastructure (DRI) introduced in XFree86 4.0. If you say Y here, you need to select diff --git a/drivers/gpu/drm/drm_atomic.c b/drivers/gpu/drm/drm_atomic.c index 5e73954..d1ae463 100644 --- a/drivers/gpu/drm/drm_atomic.c +++ b/drivers/gpu/drm/drm_atomic.c @@ -30,6 +30,7 @@ #include #include #include +#include #include "drm_crtc_internal.h" @@ -686,6 +687,17 @@ int drm_atomic_plane_set_property(struct drm_plane *plane, drm_atomic_set_fb_for_plane(state, fb); if (fb) drm_framebuffer_unreference(fb); + } else if (property == config->prop_in_fence_fd) { + if (U642I64(val) == -1) + return 0; + Sorry, I'm a bit late to bring this up but what's the expected behaviour of a commit which sets OUT_FENCE twice (first with a valid fence, then with -1 later in the list)? Maybe you need to actually clear state->fence for -1. + if (state->fence) + dma_fence_put(state->fence); + + state->fence = sync_file_get_fence(val); + if (!state->fence) + return -EINVAL; + } else if (property == config->prop_crtc_id) { struct drm_crtc *crtc = drm_crtc_find(dev, val); return drm_atomic_set_crtc_for_plane(state, crtc); @@ -747,6 +759,8 @@ drm_atomic_plane_get_property(struct drm_plane *plane, if (property == config->prop_fb_id) { *val = (state->fb) ? state->fb->base.id : 0; + } else if (property == config->prop_in_fence_fd) { + *val = -1; } else if (property == config->prop_crtc_id) { *val = (state->crtc) ? state->crtc->base.id : 0; } else if (property == config->prop_crtc_x) { diff --git a/drivers/gpu/drm/drm_atomic_helper.c b/drivers/gpu/drm/drm_atomic_helper.c index 75ad01d..26a067f 100644 --- a/drivers/gpu/drm/drm_atomic_helper.c +++ b/drivers/gpu/drm/drm_atomic_helper.c @@ -3114,6 +3114,9 @@ void __drm_atomic_helper_plane_destroy_state(struct drm_plane_state *state) { if (state->fb) drm_framebuffer_unreference(state->fb); + + if (state->fence) + dma_fence_put(state->fence); } EXPORT_SYMBOL(__drm_atomic_helper_plane_destroy_state); It looks like you need to add something in __drm_atomic_helper_plane_duplicate_state() to either get a reference on the fence or set state->fence = NULL, because the memcpy() there will copy the pointer. Cheers, -Brian diff --git a/drivers/gpu/drm/drm_crtc.c b/drivers/gpu/drm/drm_crtc.c index 13441e2..7878bfd 100644 --- a/drivers/gpu/drm/drm_crtc.c +++ b/drivers/gpu/drm/drm_crtc.c @@ -397,6 +397,12 @@ static int drm_mode_create_standard_properties(struct drm_device *dev) return -ENOMEM; dev->mode_config.prop_fb_id = prop; + prop = drm_property_create_signed_range(dev, DRM_MODE_PROP_ATOMIC, + "IN_FENCE_FD", -1, INT_MAX); + if (!prop) + return -ENOMEM; + dev->mode_config.prop_in_fence_fd = prop; +
Re: [PATCH v7 1/3] drm/fence: add in-fences support
On Tue, Nov 08, 2016 at 03:54:48PM +0900, Gustavo Padovan wrote: > From: Gustavo Padovan > > There is now a new property called IN_FENCE_FD attached to every plane > state that receives sync_file fds from userspace via the atomic commit > IOCTL. > > The fd is then translated to a fence (that may be a fence_array > subclass or just a normal fence) and then used by DRM to fence_wait() for > all fences in the sync_file to signal. So it only commits when all > framebuffers are ready to scanout. > > v2: Comments by Daniel Vetter: > - remove set state->fence = NULL in destroy phase > - accept fence -1 as valid and just return 0 > - do not call fence_get() - sync_file_fences_get() already calls it > - fence_put() if state->fence is already set, in case userspace > set the property more than once. > > v3: WARN_ON if fence is set but state has no FB > > v4: Comment from Maarten Lankhorst > - allow set fence with no related fb > > v5: rename FENCE_FD to IN_FENCE_FD > > v6: Comments by Daniel Vetter: > - rename plane_state->in_fence back to "fence" > - re-introduce WARN_ON if fence set but no fb > > - rebase after fence -> dma_fence rename > > Signed-off-by: Gustavo Padovan > --- > drivers/gpu/drm/Kconfig | 1 + > drivers/gpu/drm/drm_atomic.c| 14 ++ > drivers/gpu/drm/drm_atomic_helper.c | 3 +++ > drivers/gpu/drm/drm_crtc.c | 6 ++ > drivers/gpu/drm/drm_plane.c | 1 + > include/drm/drm_crtc.h | 5 + > 6 files changed, 30 insertions(+) > > diff --git a/drivers/gpu/drm/Kconfig b/drivers/gpu/drm/Kconfig > index 483059a..43cb33d 100644 > --- a/drivers/gpu/drm/Kconfig > +++ b/drivers/gpu/drm/Kconfig > @@ -12,6 +12,7 @@ menuconfig DRM > select I2C > select I2C_ALGOBIT > select DMA_SHARED_BUFFER > + select SYNC_FILE > help > Kernel-level support for the Direct Rendering Infrastructure (DRI) > introduced in XFree86 4.0. If you say Y here, you need to select > diff --git a/drivers/gpu/drm/drm_atomic.c b/drivers/gpu/drm/drm_atomic.c > index 5e73954..d1ae463 100644 > --- a/drivers/gpu/drm/drm_atomic.c > +++ b/drivers/gpu/drm/drm_atomic.c > @@ -30,6 +30,7 @@ > #include > #include > #include > +#include > > #include "drm_crtc_internal.h" > > @@ -686,6 +687,17 @@ int drm_atomic_plane_set_property(struct drm_plane > *plane, > drm_atomic_set_fb_for_plane(state, fb); > if (fb) > drm_framebuffer_unreference(fb); > + } else if (property == config->prop_in_fence_fd) { > + if (U642I64(val) == -1) > + return 0; > + > + if (state->fence) > + dma_fence_put(state->fence); > + > + state->fence = sync_file_get_fence(val); > + if (!state->fence) > + return -EINVAL; Might have missed it, but your igt only seems to test for -1 (i.e. no fence), not for an entirely invalid fence. I think we need another testcase which opens some file (/dev/null maybe) and tries to use that as fence. Another corner case is error unrolling. I think we need 3 cases for full coverage: - Early error unrolling while setting properties. Best way imo would be 1) set an in-fence (correctly), then 2) set an invalid property (just set the CRTC plane property to something invalid like ~0ULL). - Late unrolling (i.e. atomic_check fails). I think the reliable way here to provoke this on all drivers would be: Atomic commit with in-fence and otherwise everything the same (i.e. known to work), but changing the crtc's ACTIVE property from false to true without setting ALLOW_MODESET. - And then also a successful TEST_ONLY commit. Same is required for out-fences ofc too. Patch itself looks good, and I couldn't poke any holes into it. Reviewed-by: Daniel Vetter > + > } else if (property == config->prop_crtc_id) { > struct drm_crtc *crtc = drm_crtc_find(dev, val); > return drm_atomic_set_crtc_for_plane(state, crtc); > @@ -747,6 +759,8 @@ drm_atomic_plane_get_property(struct drm_plane *plane, > > if (property == config->prop_fb_id) { > *val = (state->fb) ? state->fb->base.id : 0; > + } else if (property == config->prop_in_fence_fd) { > + *val = -1; > } else if (property == config->prop_crtc_id) { > *val = (state->crtc) ? state->crtc->base.id : 0; > } else if (property == config->prop_crtc_x) { > diff --git a/drivers/gpu/drm/drm_atomic_helper.c > b/drivers/gpu/drm/drm_atomic_helper.c > index 75ad01d..26a067f 100644 > --- a/drivers/gpu/drm/drm_atomic_helper.c > +++ b/drivers/gpu/drm/drm_atomic_helper.c > @@ -3114,6 +3114,9 @@ void __drm_atomic_helper_plane_destroy_state(struct > drm_plane_state *state) > { > if (state->fb) > drm_framebuffer_unreference(state->fb); > + > + if (state->fence) > +
[PATCH v7 1/3] drm/fence: add in-fences support
From: Gustavo Padovan There is now a new property called IN_FENCE_FD attached to every plane state that receives sync_file fds from userspace via the atomic commit IOCTL. The fd is then translated to a fence (that may be a fence_array subclass or just a normal fence) and then used by DRM to fence_wait() for all fences in the sync_file to signal. So it only commits when all framebuffers are ready to scanout. v2: Comments by Daniel Vetter: - remove set state->fence = NULL in destroy phase - accept fence -1 as valid and just return 0 - do not call fence_get() - sync_file_fences_get() already calls it - fence_put() if state->fence is already set, in case userspace set the property more than once. v3: WARN_ON if fence is set but state has no FB v4: Comment from Maarten Lankhorst - allow set fence with no related fb v5: rename FENCE_FD to IN_FENCE_FD v6: Comments by Daniel Vetter: - rename plane_state->in_fence back to "fence" - re-introduce WARN_ON if fence set but no fb - rebase after fence -> dma_fence rename Signed-off-by: Gustavo Padovan --- drivers/gpu/drm/Kconfig | 1 + drivers/gpu/drm/drm_atomic.c| 14 ++ drivers/gpu/drm/drm_atomic_helper.c | 3 +++ drivers/gpu/drm/drm_crtc.c | 6 ++ drivers/gpu/drm/drm_plane.c | 1 + include/drm/drm_crtc.h | 5 + 6 files changed, 30 insertions(+) diff --git a/drivers/gpu/drm/Kconfig b/drivers/gpu/drm/Kconfig index 483059a..43cb33d 100644 --- a/drivers/gpu/drm/Kconfig +++ b/drivers/gpu/drm/Kconfig @@ -12,6 +12,7 @@ menuconfig DRM select I2C select I2C_ALGOBIT select DMA_SHARED_BUFFER + select SYNC_FILE help Kernel-level support for the Direct Rendering Infrastructure (DRI) introduced in XFree86 4.0. If you say Y here, you need to select diff --git a/drivers/gpu/drm/drm_atomic.c b/drivers/gpu/drm/drm_atomic.c index 5e73954..d1ae463 100644 --- a/drivers/gpu/drm/drm_atomic.c +++ b/drivers/gpu/drm/drm_atomic.c @@ -30,6 +30,7 @@ #include #include #include +#include #include "drm_crtc_internal.h" @@ -686,6 +687,17 @@ int drm_atomic_plane_set_property(struct drm_plane *plane, drm_atomic_set_fb_for_plane(state, fb); if (fb) drm_framebuffer_unreference(fb); + } else if (property == config->prop_in_fence_fd) { + if (U642I64(val) == -1) + return 0; + + if (state->fence) + dma_fence_put(state->fence); + + state->fence = sync_file_get_fence(val); + if (!state->fence) + return -EINVAL; + } else if (property == config->prop_crtc_id) { struct drm_crtc *crtc = drm_crtc_find(dev, val); return drm_atomic_set_crtc_for_plane(state, crtc); @@ -747,6 +759,8 @@ drm_atomic_plane_get_property(struct drm_plane *plane, if (property == config->prop_fb_id) { *val = (state->fb) ? state->fb->base.id : 0; + } else if (property == config->prop_in_fence_fd) { + *val = -1; } else if (property == config->prop_crtc_id) { *val = (state->crtc) ? state->crtc->base.id : 0; } else if (property == config->prop_crtc_x) { diff --git a/drivers/gpu/drm/drm_atomic_helper.c b/drivers/gpu/drm/drm_atomic_helper.c index 75ad01d..26a067f 100644 --- a/drivers/gpu/drm/drm_atomic_helper.c +++ b/drivers/gpu/drm/drm_atomic_helper.c @@ -3114,6 +3114,9 @@ void __drm_atomic_helper_plane_destroy_state(struct drm_plane_state *state) { if (state->fb) drm_framebuffer_unreference(state->fb); + + if (state->fence) + dma_fence_put(state->fence); } EXPORT_SYMBOL(__drm_atomic_helper_plane_destroy_state); diff --git a/drivers/gpu/drm/drm_crtc.c b/drivers/gpu/drm/drm_crtc.c index 13441e2..7878bfd 100644 --- a/drivers/gpu/drm/drm_crtc.c +++ b/drivers/gpu/drm/drm_crtc.c @@ -397,6 +397,12 @@ static int drm_mode_create_standard_properties(struct drm_device *dev) return -ENOMEM; dev->mode_config.prop_fb_id = prop; + prop = drm_property_create_signed_range(dev, DRM_MODE_PROP_ATOMIC, + "IN_FENCE_FD", -1, INT_MAX); + if (!prop) + return -ENOMEM; + dev->mode_config.prop_in_fence_fd = prop; + prop = drm_property_create_object(dev, DRM_MODE_PROP_ATOMIC, "CRTC_ID", DRM_MODE_OBJECT_CRTC); if (!prop) diff --git a/drivers/gpu/drm/drm_plane.c b/drivers/gpu/drm/drm_plane.c index 249c0ae..3957ef8 100644 --- a/drivers/gpu/drm/drm_plane.c +++ b/drivers/gpu/drm/drm_plane.c @@ -137,6 +137,7 @@ int drm_universal_plane_init(struct drm_device *dev, struct drm_plane *plane, if (drm_core_check_feature(dev, DRIVER_ATOMIC)) { drm_object_attach_property(&pla