On Tue, Apr 28, 2020 at 05:36:03PM +0000, Ruhl, Michael J wrote:
-----Original Message----- From: Daniel Vetter daniel@ffwll.ch Sent: Tuesday, April 28, 2020 11:02 AM To: Ruhl, Michael J michael.j.ruhl@intel.com Cc: dri-devel@lists.freedesktop.org; daniel@ffwll.ch; Xiong, Jianxin jianxin.xiong@intel.com Subject: Re: [PATCH 3/5] drm/i915/dmabuf: Add LMEM knowledge to dmabuf map handler
On Wed, Apr 22, 2020 at 05:25:17PM -0400, Michael J. Ruhl wrote:
LMEM backed buffer objects do not have struct page information. Instead the dma_address of the struct sg is used to store the LMEM address information (relative to the device, this is not the CPU physical address).
The dmabuf map handler requires pages to do a dma_map_xx.
Add new mapping/unmapping functions, based on the LMEM usage of the dma_address to allow LMEM backed buffer objects to be mapped.
Before mapping check the peer2peer distance to verify that P2P dma can occur.
So this is supposed to check the importer's allow_peer2peer flag, and that one is supposed to require the implementation of ->move_notify. Which requires a pile of locking changes to align with dma_resv.
Yeah, I was avoiding that step for the moment. I am not completely comfortable with the how and why of how the move_notify is supposed to work, so I left the current mechanism "as is", and am planning on adding the move_notify part as a next step.
By not doing all that you avoid the lockdep splats, but you're also breaking the peer2peer dma-buf contract big time :-)
OK. I have some concerns because of the differences between the AMD and i915 implementations. It is not clear to me how compatible they currently are, and if "matched" the implementation how that would affect the i915 driver.
That's kinda the problem. They're not compatible :-/
I'm also pretty sure that we'll discover a bunch of places where the current debug checks and lockdep annotations we have are insufficient, and we'll need to add more to lock down this cross driver interface. -Daniel
I think this needs more work, or I need better glasses in case I'm not spotting where this is all done.
Nope, your glasses are good.
I will take a close look at how to incorporate the peer2peer stuff.
Mike
-Daniel
Signed-off-by: Michael J. Ruhl michael.j.ruhl@intel.com
drivers/gpu/drm/i915/gem/i915_gem_dmabuf.c | 82
++++++++++++++++++++--
1 file changed, 76 insertions(+), 6 deletions(-)
diff --git a/drivers/gpu/drm/i915/gem/i915_gem_dmabuf.c
b/drivers/gpu/drm/i915/gem/i915_gem_dmabuf.c
index 7ea4abb6a896..402c989cc23d 100644 --- a/drivers/gpu/drm/i915/gem/i915_gem_dmabuf.c +++ b/drivers/gpu/drm/i915/gem/i915_gem_dmabuf.c @@ -7,6 +7,7 @@ #include <linux/dma-buf.h> #include <linux/highmem.h> #include <linux/dma-resv.h> +#include <linux/pci-p2pdma.h> #include <linux/scatterlist.h>
#include "i915_drv.h" @@ -18,6 +19,67 @@ static struct drm_i915_gem_object
*dma_buf_to_obj(struct dma_buf *buf)
return to_intel_bo(buf->priv); }
+static void dmabuf_unmap_addr(struct device *dev, struct scatterlist *sgl,
int nents, enum dma_data_direction dir)+{
- struct scatterlist *sg;
- int i;
- for_each_sg(sgl, sg, nents, i)
dma_unmap_resource(dev, sg_dma_address(sg),sg_dma_len(sg),
dir, 0);+}
+/**
- dmabuf_map_addr - Update LMEM address to a physical address and
map the
- resource.
- @dev: valid device
- @obj: valid i915 GEM object
- @sg: scatter list to appy mapping to
- @nents: number of entries in the scatter list
- @dir: DMA direction
- The dma_address of the scatter list is the LMEM "address". From this
the
- actual physical address can be determined.
- Return of 0 means error.
- */
+static int dmabuf_map_addr(struct device *dev, struct
drm_i915_gem_object *obj,
struct scatterlist *sgl, int nents,enum dma_data_direction dir)+{
- struct scatterlist *sg;
- phys_addr_t addr;
- int distance;
- int i;
- distance = pci_p2pdma_distance_many(obj->base.dev->pdev, &dev,
1,
true);- if (distance < 0) {
pr_info("%s: from: %s to: %s distance: %d\n", __func__,pci_name(obj->base.dev->pdev), dev_name(dev),distance);return 0;- }
- for_each_sg(sgl, sg, nents, i) {
addr = sg_dma_address(sg) + obj->mm.region->io_start;sg->dma_address = dma_map_resource(dev, addr, sg-length, dir,
0);if (dma_mapping_error(dev, sg->dma_address))goto unmap;sg->dma_length = sg->length;- }
- return nents;
+unmap:
- dmabuf_unmap_addr(dev, sgl, i, dir);
- return 0;
+}
static struct sg_table *i915_gem_map_dma_buf(struct
dma_buf_attachment *attach,
enum dma_data_direction dir){ @@ -44,12 +106,17 @@ static struct sg_table
*i915_gem_map_dma_buf(struct dma_buf_attachment *attach,
dst = sgt->sgl; for_each_sg(obj->mm.pages->sgl, src, obj->mm.pages->nents, i) { sg_set_page(dst, sg_page(src), src->length, 0);
sg_dma_address(dst) = sg_dma_address(src);
- if (!dma_map_sg_attrs(attach->dev,
sgt->sgl, sgt->nents, dir,DMA_ATTR_SKIP_CPU_SYNC)) {
- if (i915_gem_object_has_struct_page(obj))
ret = dma_map_sg_attrs(attach->dev, sgt->sgl, sgt->nents,dir,
DMA_ATTR_SKIP_CPU_SYNC);- else
ret = dmabuf_map_addr(attach->dev, obj, sgt->sgl, sgt-nents,
dir);- if (!ret) { ret = -ENOMEM; goto err_free_sg; }
@@ -72,9 +139,12 @@ static void i915_gem_unmap_dma_buf(struct
dma_buf_attachment *attach,
{ struct drm_i915_gem_object *obj = dma_buf_to_obj(attach- dmabuf);
- dma_unmap_sg_attrs(attach->dev,
sgt->sgl, sgt->nents, dir,DMA_ATTR_SKIP_CPU_SYNC);
- if (i915_gem_object_has_struct_page(obj))
dma_unmap_sg_attrs(attach->dev, sgt->sgl, sgt->nents, dir,DMA_ATTR_SKIP_CPU_SYNC);- else
dmabuf_unmap_addr(attach->dev, sgt->sgl, sgt->nents, dir);- sg_free_table(sgt); kfree(sgt);
-- 2.21.0
-- Daniel Vetter Software Engineer, Intel Corporation http://blog.ffwll.ch