similar to: [PATCH 0/3] drm/nouveau: Add support for fence FDs

From: Thierry Reding <treding at nvidia.com> Hi, This series implements a new IOCTL to submit push buffers that can optionally return a sync FD or sync object to userspace. This is useful in cases where userspace wants to synchronize operations between the GPU and another driver (such as KMS for display). Among other things this allows extensions such as eglDupNativeFenceFDANDROID to be

Hi guys, I'd like to start a new thread about explicit fence synchronization. This time with a Nouveau twist. :-) First, let me define what I understand by implicit/explicit sync: Implicit synchronization * Fences are attached to buffers * Kernel manages fences automatically based on buffer read/write access Explicit synchronization * Fences are passed around independently * Kernel takes

From: Dave Airlie <airlied at redhat.com> This is the 3rd iteration of this after talking to Ben and Danilo, I think this makes sense now. The fence sync logic doesn't handle a fence sync across devices as it tries to write to a channel offset from one device into the fence bo from a different device, which won't work so well. This patch fixes that to avoid using the sync path in

This series applies on top of the driver-core-next branch of git://git.kernel.org/pub/scm/linux/kernel/git/gregkh/driver-core.git Before converting ttm to the new fence interface I had to fix some drivers to require a reservation before poking with fence_obj. After flipping the switch RCU becomes available instead, and the extra reservations can be dropped again. :-) I've done at least basic

This series depends on the previously posted reservation api patches. 2 of them are not yet in for-next-fences branch of git://git.linaro.org/people/sumit.semwal/linux-3.x.git The missing patches are still in my vmwgfx_wip branch at git://people.freedesktop.org/~mlankhorst/linux All ttm drivers are converted to the fence api, fence_lock is removed and rcu is used in its place. qxl is the first

fence_is_signaled callback should support being run in atomic context, but not in irq context. Signed-off-by: Maarten Lankhorst <maarten.lankhorst at canonical.com> --- include/linux/fence.h | 23 +++++++++++++++++++---- 1 file changed, 19 insertions(+), 4 deletions(-) diff --git a/include/linux/fence.h b/include/linux/fence.h index d174585b874b..c1a4519ba2f5 100644 ---

On Thu, Jan 09, 2025 at 10:55:53AM +1000, Dave Airlie wrote: > From: Dave Airlie <airlied at redhat.com> > > This is the 3rd iteration of this after talking to Ben and > Danilo, I think this makes sense now. > > The fence sync logic doesn't handle a fence sync across devices > as it tries to write to a channel offset from one device into > the fence bo from a

nouveau_fence_emit() can fail before and after initializing the dma-fence and hence before and after initializing the dma-fence' kref. In order to avoid nouveau_fence_emit() potentially failing before dma-fence initialization pass the channel to nouveau_fence_new() already and perform the required check before even allocating the fence. While at it, restore the original behavior of

nouveau_bo_wait will make the GPU channel wait for fence if possible, otherwise falling back to waiting with the CPU using ttm_bo_wait. The nouveau_fence_sync function currently returns -ENOSYS, and is the focus of the next patch. Signed-off-by: Luca Barbieri <luca at luca-barbieri.com> --- drivers/gpu/drm/nouveau/nouveau_bo.c | 68 ++++++++++++++++++++++++++++++-

We can't copy_*_user while holding reservations, that will (soon even for nouveau) lead to deadlocks. And it breaks the cross-driver contract around dma_resv. Fix this by adding a slowpath for when we need relocations, and by pushing the writeback of the new presumed offsets to the very end. Aside from "it compiles" entirely untested unfortunately. Signed-off-by: Daniel Vetter

- Depth and stencil buffers are supposed to be large enough in general. Signed-off-by: Maarten Maathuis <madman2003 at gmail.com> --- drivers/gpu/drm/nouveau/nouveau_bo.c | 9 ++++----- 1 files changed, 4 insertions(+), 5 deletions(-) diff --git a/drivers/gpu/drm/nouveau/nouveau_bo.c b/drivers/gpu/drm/nouveau/nouveau_bo.c index e342a41..9fc4bd6 100644 ---

Changes in v2: - Addressed review comments nouveau_bo_wait will make the GPU channel wait for fence if possible, otherwise falling back to waiting with the CPU using ttm_bo_wait. The nouveau_fence_sync function currently returns -ENOSYS, and is the focus of the next patch. Signed-off-by: Luca Barbieri <luca at luca-barbieri.com> --- drivers/gpu/drm/nouveau/nouveau_bo.c | 68

From: Maarten Lankhorst <maarten.lankhorst at canonical.com> Commit de7b7d59d54852c introduced tiled GART, but a linear copy is still performed. This may result in errors on eviction, fix it by checking tiling from memtype. Signed-off-by: Maarten Lankhorst <maarten.lankhorst at canonical.com> Cc: stable at vger.kernel.org #3.10+ --- drivers/gpu/drm/nouveau/nouveau_bo.c | 33

We can't copy_*_user while holding reservations, that will (soon even for nouveau) lead to deadlocks. And it breaks the cross-driver contract around dma_resv. Fix this by adding a slowpath for when we need relocations, and by pushing the writeback of the new presumed offsets to the very end. Aside from "it compiles" entirely untested unfortunately. Signed-off-by: Daniel Vetter

We can't copy_*_user while holding reservations, that will (soon even for nouveau) lead to deadlocks. And it breaks the cross-driver contract around dma_resv. Fix this by adding a slowpath for when we need relocations, and by pushing the writeback of the new presumed offsets to the very end. Aside from "it compiles" entirely untested unfortunately. Signed-off-by: Daniel Vetter

From: Thierry Reding <treding at nvidia.com> This set of patches adds support for fences to Tegra DRM and complements the fence FD support for Nouveau. Technically this isn't necessary for a fence-based synchronization loop with Nouveau because the KMS core takes care of all that, but engines behind host1x can use the IOCTL extensions provided here to emit fence FDs that in turn can be

Hi there, Resending these now that they've had some more polish and testing, and I heard that Ben's vacation is over :-) These patches work as a starting point for more explicit error mechanisms and better robustness. At the moment, when a job hangs or faults, it seems that nouveau doesn't quite know how to handle the situation and often results in a hang. Some of these situations

From: Dave Airlie <airlied at redhat.com> fences are signalled on nvidia hw using non-stall interrupts. non-stall interrupts are not latched from my reading. When nouveau emits a fence, it requests a NON_STALL signalling, but it only calls the interface to allow the non-stall irq to happen after it has already emitted the fence. A recent change eacabb546271 ("nouveau: push event

On Tue, Jan 23, 2024 at 05:25:38PM +1000, Dave Airlie wrote: > From: Dave Airlie <airlied at redhat.com> > > fences are signalled on nvidia hw using non-stall interrupts. > > non-stall interrupts are not latched from my reading. > > When nouveau emits a fence, it requests a NON_STALL signalling, > but it only calls the interface to allow the non-stall irq to happen

From: Ben Skeggs <bskeggs at redhat.com> Signed-off-by: Ben Skeggs <bskeggs at redhat.com> --- drivers/gpu/drm/nouveau/nouveau_channel.c | 2 -- drivers/gpu/drm/nouveau/nouveau_drv.h | 1 - drivers/gpu/drm/nouveau/nouveau_fence.c | 28 ++++------------------------ drivers/gpu/drm/nouveau/nv04_graph.c | 1 - 4 files changed, 4 insertions(+), 28 deletions(-) diff