Skip to contentSkip to conversation

drm/panfrost: Add support for 2MB page entries

PATCH In review Rob Herring 4 participants 4 messages Apr 11, 2019 – Apr 12, 2019

Subsystems (3)
lore
Share & tools
lore

Rob Herring

author maintainer
#0 lore raw

PATCH 1/1 drm/panfrost: Add support for 2MB page entries

Add support for 2MB sized pages. This will improve our map and unmap times and save a bit of memory by avoiding 3rd level page tables for contiguous allocations.

As we use shmem for buffers and huge page allocations for shmem are off by default, there isn't an improvement out of the box and userspace must enable THP for shmem.

It's not clear if the h/w can support 1GB page sizes which standard ARM long format descriptors support. In any case, it is unlikely we'll see any contiguous 1GB allocations on current h/w.

Cc Tomeu Vizoso <[email protected]>
Cc David Airlie <[email protected]>
Cc Daniel Vetter <[email protected]>
---
 drivers/gpu/drm/panfrost/panfrost_mmu.c | 32 ++++++++++++++++++-------
 1 file changed, 24 insertions(+), 8 deletions(-)
drivers/gpu/drm/panfrost/panfrost_mmu.c +24 -8
index 502af37d4b62..762b1bd2a8c2 100644
@@ -137,6 +137,14 @@ static void mmu_disable(struct panfrost_device *pfdev, u32 as_nr)
 	write_cmd(pfdev, as_nr, AS_COMMAND_UPDATE);
 }
 
+static size_t get_pgsize(u64 addr, size_t size)
+{
+	if (addr & (SZ_2M - 1) || size < SZ_2M)
+		return SZ_4K;
+
+	return SZ_2M;
+}
+
 int panfrost_mmu_map(struct panfrost_gem_object *bo)
 {
 	struct drm_gem_object *obj = &bo->base.base;
@@ -165,10 +173,12 @@ int panfrost_mmu_map(struct panfrost_gem_object *bo)
 		dev_dbg(pfdev->dev, "map: iova=%llx, paddr=%lx, len=%zx", iova, paddr, len);
 
 		while (len) {
-			ops->map(ops, iova, paddr, SZ_4K, IOMMU_WRITE | IOMMU_READ);
-			iova += SZ_4K;
-			paddr += SZ_4K;
-			len -= SZ_4K;
+			size_t pgsize = get_pgsize(iova | paddr, len);
+
+			ops->map(ops, iova, paddr, pgsize, IOMMU_WRITE | IOMMU_READ);
+			iova += pgsize;
+			paddr += pgsize;
+			len -= pgsize;
 		}
 	}
 
@@ -202,9 +212,15 @@ void panfrost_mmu_unmap(struct panfrost_gem_object *bo)
 	mutex_lock(&pfdev->mmu->lock);
 
 	while (unmapped_len < len) {
-		ops->unmap(ops, iova, SZ_4K);
-		iova += SZ_4K;
-		unmapped_len += SZ_4K;
+		size_t unmapped_page;
+		size_t pgsize = get_pgsize(iova, len - unmapped_len);
+
+		unmapped_page = ops->unmap(ops, iova, pgsize);
+		if (!unmapped_page)
+			break;
+
+		iova += unmapped_page;
+		unmapped_len += unmapped_page;
 	}
 
 	mmu_hw_do_operation(pfdev, 0, bo->node.start << PAGE_SHIFT,
@@ -342,7 +358,7 @@ int panfrost_mmu_init(struct panfrost_device *pfdev)
 	mmu_write(pfdev, MMU_INT_MASK, ~0);
 
 	pfdev->mmu->pgtbl_cfg = (struct io_pgtable_cfg) {
-		.pgsize_bitmap	= SZ_4K, // | SZ_2M | SZ_1G),
+		.pgsize_bitmap	= SZ_4K | SZ_2M,
 		.ias		= FIELD_GET(0xff, pfdev->features.mmu_features),
 		.oas		= FIELD_GET(0xff00, pfdev->features.mmu_features),
 		.tlb		= &mmu_tlb_ops,
Signature · 6 lines
2.19.1

_______________________________________________
dri-devel mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/dri-devel
maintainer Reviewed-by on v1 1/1
#3 lore raw
In reply to Rob Herring
Quoted 8 lines · Rob Herring

Add support for 2MB sized pages. This will improve our map and unmap times and save a bit of memory by avoiding 3rd level page tables for contiguous allocations.

As we use shmem for buffers and huge page allocations for shmem are off by default, there isn't an improvement out of the box and userspace must enable THP for shmem.

Quoted 8 lines

It's not clear if the h/w can support 1GB page sizes which standard ARM long format descriptors support. In any case, it is unlikely we'll see any contiguous 1GB allocations on current h/w.

Cc Tomeu Vizoso <[email protected]>
Cc David Airlie <[email protected]>
Cc Daniel Vetter <[email protected]>

LGTM, and I see a nice speed up too!

_______________________________________________
dri-devel mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/dri-devel