From 7713f79b0a5473eb0b8456d36b99ae00815dd8a1 Mon Sep 17 00:00:00 2001 From: Eric Anholt Date: Wed, 27 Mar 2019 17:44:40 -0700 Subject: [PATCH] drm/v3d: Add missing implicit synchronization. It is the expectation of existing userspace (X11 + Mesa, in particular) that jobs submitted to the kernel against a shared BO will get implicitly synchronized by their submission order. If we want to allow clever userspace to disable implicit synchronization, we should do that under its own submit flag (as amdgpu and lima do). Note that we currently only implicitly sync for the rendering pass, not binning -- if you texture-from-pixmap in the binning vertex shader (vertex coordinate generation), you'll miss out on synchronization. Fixes flickering when multiple clients are running in parallel, particularly GL apps and compositors. Signed-off-by: Eric Anholt --- drivers/gpu/drm/v3d/v3d_drv.h | 10 +--- drivers/gpu/drm/v3d/v3d_gem.c | 98 ++++++++++++++++++++++++++++++--- drivers/gpu/drm/v3d/v3d_sched.c | 45 ++------------- 3 files changed, 96 insertions(+), 57 deletions(-) --- a/drivers/gpu/drm/v3d/v3d_drv.h +++ b/drivers/gpu/drm/v3d/v3d_drv.h @@ -186,8 +186,9 @@ struct v3d_job { struct v3d_bo **bo; u32 bo_count; - /* An optional fence userspace can pass in for the job to depend on. */ - struct dma_fence *in_fence; + struct dma_fence **deps; + int deps_count; + int deps_size; /* v3d fence to be signaled by IRQ handler when the job is complete. */ struct dma_fence *irq_fence; @@ -219,11 +220,6 @@ struct v3d_bin_job { struct v3d_render_job { struct v3d_job base; - /* Optional fence for the binner, to depend on before starting - * our job. - */ - struct dma_fence *bin_done_fence; - /* GPU virtual addresses of the start/end of the CL job. */ u32 start, end; --- a/drivers/gpu/drm/v3d/v3d_gem.c +++ b/drivers/gpu/drm/v3d/v3d_gem.c @@ -218,6 +218,71 @@ v3d_unlock_bo_reservations(struct v3d_bo ww_acquire_fini(acquire_ctx); } +static int +v3d_add_dep(struct v3d_job *job, struct dma_fence *fence) +{ + if (!fence) + return 0; + + if (job->deps_size == job->deps_count) { + int new_deps_size = max(job->deps_size * 2, 4); + struct dma_fence **new_deps = + krealloc(job->deps, new_deps_size * sizeof(*new_deps), + GFP_KERNEL); + if (!new_deps) { + dma_fence_put(fence); + return -ENOMEM; + } + + job->deps = new_deps; + job->deps_size = new_deps_size; + } + + job->deps[job->deps_count++] = fence; + + return 0; +} + +/** + * Adds the required implicit fences before executing the job + * + * Userspace (X11 + Mesa) requires that a job submitted against a shared BO + * from one fd will implicitly synchronize against previous jobs submitted + * against that BO from other fds. + * + * Currently we don't bother trying to track the shared BOs, and instead just + * sync everything. However, our synchronization is only for the render pass + * -- the binning stage (VS coordinate calculations) ignores implicit sync, + * since using shared buffers for texture coordinates seems unlikely, and + * implicitly syncing them would break bin/render parallelism. If we want to + * fix that, we should introduce a flag when VS texturing has been used in the + * binning stage, or a set of flags for which BOs are sampled during binning. + */ +static int +v3d_add_implicit_fences(struct v3d_job *job, struct v3d_bo *bo) +{ + int i, ret, nr_fences; + struct dma_fence **fences; + + ret = reservation_object_get_fences_rcu(bo->resv, NULL, + &nr_fences, &fences); + if (ret || !nr_fences) + return ret; + + for (i = 0; i < nr_fences; i++) { + ret = v3d_add_dep(job, fences[i]); + if (ret) + break; + } + + /* Free any remaining fences after error. */ + for (; i < nr_fences; i++) + dma_fence_put(fences[i]); + kfree(fences); + + return ret; +} + /* Takes the reservation lock on all the BOs being referenced, so that * at queue submit time we can update the reservations. * @@ -226,10 +291,11 @@ v3d_unlock_bo_reservations(struct v3d_bo * to v3d, so we don't attach dma-buf fences to them. */ static int -v3d_lock_bo_reservations(struct v3d_bo **bos, - int bo_count, +v3d_lock_bo_reservations(struct v3d_job *job, struct ww_acquire_ctx *acquire_ctx) { + struct v3d_bo **bos = job->bo; + int bo_count = job->bo_count; int contended_lock = -1; int i, ret; @@ -281,6 +347,13 @@ retry: * before we commit the CL to the hardware. */ for (i = 0; i < bo_count; i++) { + ret = v3d_add_implicit_fences(job, bos[i]); + if (ret) { + v3d_unlock_bo_reservations(bos, bo_count, + acquire_ctx); + return ret; + } + ret = reservation_object_reserve_shared(bos[i]->resv); if (ret) { v3d_unlock_bo_reservations(bos, bo_count, @@ -383,7 +456,10 @@ v3d_job_free(struct kref *ref) } kvfree(job->bo); - dma_fence_put(job->in_fence); + for (i = 0; i < job->deps_count; i++) + dma_fence_put(job->deps[i]); + kfree(job->deps); + dma_fence_put(job->irq_fence); dma_fence_put(job->done_fence); @@ -464,15 +540,20 @@ v3d_job_init(struct v3d_dev *v3d, struct struct v3d_job *job, void (*free)(struct kref *ref), u32 in_sync) { + struct dma_fence *in_fence = NULL; int ret; job->v3d = v3d; job->free = free; - ret = drm_syncobj_find_fence(file_priv, in_sync, 0, &job->in_fence); + ret = drm_syncobj_find_fence(file_priv, in_sync, 0, &in_fence); if (ret == -EINVAL) return ret; + ret = v3d_add_dep(job, in_fence); + if (ret) + return ret; + kref_init(&job->refcount); return 0; @@ -590,8 +671,7 @@ v3d_submit_cl_ioctl(struct drm_device *d if (ret) goto fail; - ret = v3d_lock_bo_reservations(render->base.bo, render->base.bo_count, - &acquire_ctx); + ret = v3d_lock_bo_reservations(&render->base, &acquire_ctx); if (ret) goto fail; @@ -601,7 +681,8 @@ v3d_submit_cl_ioctl(struct drm_device *d if (ret) goto fail_unreserve; - render->bin_done_fence = dma_fence_get(bin->base.done_fence); + ret = v3d_add_dep(&render->base, + dma_fence_get(bin->base.done_fence)); } ret = v3d_push_job(v3d_priv, &render->base, V3D_RENDER); @@ -692,8 +773,7 @@ v3d_submit_tfu_ioctl(struct drm_device * } spin_unlock(&file_priv->table_lock); - ret = v3d_lock_bo_reservations(job->base.bo, job->base.bo_count, - &acquire_ctx); + ret = v3d_lock_bo_reservations(&job->base, &acquire_ctx); if (ret) goto fail; --- a/drivers/gpu/drm/v3d/v3d_sched.c +++ b/drivers/gpu/drm/v3d/v3d_sched.c @@ -67,47 +67,10 @@ v3d_job_dependency(struct drm_sched_job struct drm_sched_entity *s_entity) { struct v3d_job *job = to_v3d_job(sched_job); - struct dma_fence *fence; - - fence = job->in_fence; - if (fence) { - job->in_fence = NULL; - return fence; - } - - return NULL; -} -/** - * Returns the fences that the render job depends on, one by one. - * v3d_job_run() won't be called until all of them have been signaled. - */ -static struct dma_fence * -v3d_render_job_dependency(struct drm_sched_job *sched_job, - struct drm_sched_entity *s_entity) -{ - struct v3d_render_job *job = to_render_job(sched_job); - struct dma_fence *fence; - - fence = v3d_job_dependency(sched_job, s_entity); - if (fence) - return fence; - - /* If we had a bin job, the render job definitely depends on - * it. We first have to wait for bin to be scheduled, so that - * its done_fence is created. - */ - fence = job->bin_done_fence; - if (fence) { - job->bin_done_fence = NULL; - return fence; - } - - /* XXX: Wait on a fence for switching the GMP if necessary, - * and then do so. - */ - - return fence; + if (!job->deps_count) + return NULL; + return job->deps[--job->deps_count]; } static struct dma_fence *v3d_bin_job_run(struct drm_sched_job *sched_job) @@ -329,7 +292,7 @@ static const struct drm_sched_backend_op }; static const struct drm_sched_backend_ops v3d_render_sched_ops = { - .dependency = v3d_render_job_dependency, + .dependency = v3d_job_dependency, .run_job = v3d_render_job_run, .timedout_job = v3d_render_job_timedout, .free_job = v3d_job_free,