diff mbox series

[v2,04/23] drm/msm: Decouple vma tracking from obj lock

Message ID 20230320144356.803762-5-robdclark@gmail.com (mailing list archive)
State Not Applicable
Headers show
Series drm/msm+PM+icc: Make job_run() reclaim-safe | expand

Commit Message

Rob Clark March 20, 2023, 2:43 p.m. UTC
From: Rob Clark <robdclark@chromium.org>

We need to use the inuse count to track that a BO is pinned until
we have the hw_fence.  But we want to remove the obj lock from the
job_run() path as this could deadlock against reclaim/shrinker
(because it is blocking the hw_fence from eventually being signaled).
So split that tracking out into a per-vma lock with narrower scope.

Signed-off-by: Rob Clark <robdclark@chromium.org>
---
 drivers/gpu/drm/msm/msm_gem.h        |  1 +
 drivers/gpu/drm/msm/msm_gem_vma.c    | 44 ++++++++++++++++++++++++----
 drivers/gpu/drm/msm/msm_ringbuffer.c |  2 +-
 3 files changed, 40 insertions(+), 7 deletions(-)
diff mbox series

Patch

diff --git a/drivers/gpu/drm/msm/msm_gem.h b/drivers/gpu/drm/msm/msm_gem.h
index d3219c523034..1929f09c5b0d 100644
--- a/drivers/gpu/drm/msm/msm_gem.h
+++ b/drivers/gpu/drm/msm/msm_gem.h
@@ -59,6 +59,7 @@  struct msm_fence_context;
 
 struct msm_gem_vma {
 	struct drm_mm_node node;
+	spinlock_t lock;
 	uint64_t iova;
 	struct msm_gem_address_space *aspace;
 	struct list_head list;    /* node in msm_gem_object::vmas */
diff --git a/drivers/gpu/drm/msm/msm_gem_vma.c b/drivers/gpu/drm/msm/msm_gem_vma.c
index 2827679dc39a..98287ed99960 100644
--- a/drivers/gpu/drm/msm/msm_gem_vma.c
+++ b/drivers/gpu/drm/msm/msm_gem_vma.c
@@ -40,19 +40,28 @@  msm_gem_address_space_get(struct msm_gem_address_space *aspace)
 
 bool msm_gem_vma_inuse(struct msm_gem_vma *vma)
 {
+	bool ret = true;
+
+	spin_lock(&vma->lock);
+
 	if (vma->inuse > 0)
-		return true;
+		goto out;
 
 	while (vma->fence_mask) {
 		unsigned idx = ffs(vma->fence_mask) - 1;
 
 		if (!msm_fence_completed(vma->fctx[idx], vma->fence[idx]))
-			return true;
+			goto out;
 
 		vma->fence_mask &= ~BIT(idx);
 	}
 
-	return false;
+	ret = false;
+
+out:
+	spin_unlock(&vma->lock);
+
+	return ret;
 }
 
 /* Actually unmap memory for the vma */
@@ -73,8 +82,7 @@  void msm_gem_vma_purge(struct msm_gem_vma *vma)
 	vma->mapped = false;
 }
 
-/* Remove reference counts for the mapping */
-void msm_gem_vma_unpin(struct msm_gem_vma *vma)
+static void vma_unpin_locked(struct msm_gem_vma *vma)
 {
 	if (GEM_WARN_ON(!vma->inuse))
 		return;
@@ -82,13 +90,23 @@  void msm_gem_vma_unpin(struct msm_gem_vma *vma)
 		vma->inuse--;
 }
 
+/* Remove reference counts for the mapping */
+void msm_gem_vma_unpin(struct msm_gem_vma *vma)
+{
+	spin_lock(&vma->lock);
+	vma_unpin_locked(vma);
+	spin_unlock(&vma->lock);
+}
+
 /* Replace pin reference with fence: */
 void msm_gem_vma_unpin_fenced(struct msm_gem_vma *vma, struct msm_fence_context *fctx)
 {
+	spin_lock(&vma->lock);
 	vma->fctx[fctx->index] = fctx;
 	vma->fence[fctx->index] = fctx->last_fence;
 	vma->fence_mask |= BIT(fctx->index);
-	msm_gem_vma_unpin(vma);
+	vma_unpin_locked(vma);
+	spin_unlock(&vma->lock);
 }
 
 /* Map and pin vma: */
@@ -103,7 +121,9 @@  msm_gem_vma_map(struct msm_gem_vma *vma, int prot,
 		return -EINVAL;
 
 	/* Increase the usage counter */
+	spin_lock(&vma->lock);
 	vma->inuse++;
+	spin_unlock(&vma->lock);
 
 	if (vma->mapped)
 		return 0;
@@ -113,11 +133,22 @@  msm_gem_vma_map(struct msm_gem_vma *vma, int prot,
 	if (!aspace)
 		return 0;
 
+	/*
+	 * NOTE: iommu/io-pgtable can allocate pages, so we cannot hold
+	 * a lock across map/unmap which is also used in the job_run()
+	 * path, as this can cause deadlock in job_run() vs shrinker/
+	 * reclaim.
+	 *
+	 * Revisit this if we can come up with a scheme to pre-alloc pages
+	 * for the pgtable in map/unmap ops.
+	 */
 	ret = aspace->mmu->funcs->map(aspace->mmu, vma->iova, sgt, size, prot);
 
 	if (ret) {
 		vma->mapped = false;
+		spin_lock(&vma->lock);
 		vma->inuse--;
+		spin_unlock(&vma->lock);
 	}
 
 	return ret;
@@ -148,6 +179,7 @@  struct msm_gem_vma *msm_gem_vma_new(struct msm_gem_address_space *aspace)
 	if (!vma)
 		return NULL;
 
+	spin_lock_init(&vma->lock);
 	vma->aspace = aspace;
 
 	return vma;
diff --git a/drivers/gpu/drm/msm/msm_ringbuffer.c b/drivers/gpu/drm/msm/msm_ringbuffer.c
index 44a22b283730..31b4fbf96c36 100644
--- a/drivers/gpu/drm/msm/msm_ringbuffer.c
+++ b/drivers/gpu/drm/msm/msm_ringbuffer.c
@@ -23,8 +23,8 @@  static struct dma_fence *msm_job_run(struct drm_sched_job *job)
 	for (i = 0; i < submit->nr_bos; i++) {
 		struct drm_gem_object *obj = &submit->bos[i].obj->base;
 
-		msm_gem_lock(obj);
 		msm_gem_vma_unpin_fenced(submit->bos[i].vma, fctx);
+		msm_gem_lock(obj);
 		msm_gem_unpin_locked(obj);
 		msm_gem_unlock(obj);
 		submit->bos[i].flags &= ~(BO_VMA_PINNED | BO_OBJ_PINNED);