Am 08.05.19 um 12:11 schrieb Daniel Vetter:
On Wed, May 08, 2019 at 11:15:33AM +0200, Daniel Vetter wrote:On Tue, May 07, 2019 at 10:13:30AM +0200, Christian König wrote:To allow a smooth transition from pinning buffer objects to dynamic invalidation we first start to cache the sg_table for an attachment. Signed-off-by: Christian König <christian.koenig@xxxxxxx> --- drivers/dma-buf/dma-buf.c | 24 ++++++++++++++++++++++++ include/linux/dma-buf.h | 14 ++++++++++++++ 2 files changed, 38 insertions(+) diff --git a/drivers/dma-buf/dma-buf.c b/drivers/dma-buf/dma-buf.c index 7c858020d14b..775e13f54083 100644 --- a/drivers/dma-buf/dma-buf.c +++ b/drivers/dma-buf/dma-buf.c @@ -573,6 +573,20 @@ struct dma_buf_attachment *dma_buf_attach(struct dma_buf *dmabuf, list_add(&attach->node, &dmabuf->attachments);mutex_unlock(&dmabuf->lock);+ + if (!dma_buf_is_dynamic(dmabuf)) { + struct sg_table *sgt; + + sgt = dmabuf->ops->map_dma_buf(attach, DMA_BIDIRECTIONAL); + if (!sgt) + sgt = ERR_PTR(-ENOMEM); + if (IS_ERR(sgt)) { + dma_buf_detach(dmabuf, attach); + return ERR_CAST(sgt); + } + attach->sgt = sgt; + } + return attach;err_attach:@@ -595,6 +609,10 @@ void dma_buf_detach(struct dma_buf *dmabuf, struct dma_buf_attachment *attach) if (WARN_ON(!dmabuf || !attach)) return;+ if (attach->sgt)+ dmabuf->ops->unmap_dma_buf(attach, attach->sgt, + DMA_BIDIRECTIONAL); + mutex_lock(&dmabuf->lock); list_del(&attach->node); if (dmabuf->ops->detach) @@ -630,6 +648,9 @@ struct sg_table *dma_buf_map_attachment(struct dma_buf_attachment *attach, if (WARN_ON(!attach || !attach->dmabuf)) return ERR_PTR(-EINVAL);+ if (attach->sgt)+ return attach->sgt; + sg_table = attach->dmabuf->ops->map_dma_buf(attach, direction); if (!sg_table) sg_table = ERR_PTR(-ENOMEM); @@ -657,6 +678,9 @@ void dma_buf_unmap_attachment(struct dma_buf_attachment *attach, if (WARN_ON(!attach || !attach->dmabuf || !sg_table)) return;+ if (attach->sgt == sg_table)+ return; + attach->dmabuf->ops->unmap_dma_buf(attach, sg_table, direction); } diff --git a/include/linux/dma-buf.h b/include/linux/dma-buf.h index 58725f890b5b..52031fdc75bb 100644 --- a/include/linux/dma-buf.h +++ b/include/linux/dma-buf.h @@ -322,6 +322,7 @@ struct dma_buf_attachment { struct dma_buf *dmabuf; struct device *dev; struct list_head node; + struct sg_table *sgt; void *priv; };@@ -373,6 +374,19 @@ static inline void get_dma_buf(struct dma_buf *dmabuf)get_file(dmabuf->file); }+/**+ * dma_buf_is_dynamic - check if a DMA-buf uses dynamic mappings. + * @dmabuf: the DMA-buf to check + * + * Returns true if a DMA-buf exporter wants to create dynamic sg table mappings + * for each attachment. False if only a single static sg table should be used. + */ +static inline bool dma_buf_is_dynamic(struct dma_buf *dmabuf) +{ + /* Always use a static mapping for now */ + return false;Hm I still expect that later on we'll want this to be decided by the attachment: It's only dynamic if both the exporter and the importer support dynamic dma-buf management, otherwise we need to pin. But anyway, I feel like we need to go over the entire thing anyway once more when p2p has landed, on this: Reviewed-by: Daniel Vetter <daniel.vetter@xxxxxxxx>Correction that I only just realized, but need to retract that r-b on this and the drm sgt cache removal patch: You now hardcode the direction to DMA_BIDIRECTIONAL, drm_prime did only keep the cache for a given direction. Now x86 drivers always set DMA_BIDIRECTIONAL, but arm soc drivers (and also v4l videobuf layer) try to guess whether it should be DMA_TO_DEVICE or DMA_FROM_DEVICE. I have no idea what the implications are, also all the cache coherency on dma-bufs is kinda ill-defined. And we can't throw the sgt cache away at map time if it doesn't fit like drm_prime does, because that reintroduces the reservation object lock, defeating the entire purpose of this. Also we can't just assume drm_prime works for everyone, since the special cases all roll their own dma-buf import. I have also not checked what exactly exporters do. No idea what to do here now. /me cries
Well it is kind of abusing to use the DMA direction here in the first place and I actually considered to completely remove it.
The key problem is that the DMA direction defines things from the view point of the CPU, e.g. DMA_TO_DEVICE means the data has been accessed by the CPU and we are now pushing it to a device.
But DMA-buf is essentially about device to device transfers, so DMA_TO_DEVICE as well as DMA_FROM_DEVICE is completely ambiguous because you don't know the point of view to start with.
Additional to that the caching implementation in drm_prime.c didn't handled it full either:
- /* - * two mappings with different directions for the same attachment are - * not allowed - */ - if (WARN_ON(prime_attach->dir != DMA_NONE)) - return ERR_PTR(-EBUSY);
Do the ARM guys and V4L guys actually do anything with the direction except for passing it on dma_map_sg_attrs() ?
Regards, Christian.
-Daniel+} + struct dma_buf_attachment *dma_buf_attach(struct dma_buf *dmabuf, struct device *dev); void dma_buf_detach(struct dma_buf *dmabuf, -- 2.17.1 _______________________________________________ dri-devel mailing list dri-devel@xxxxxxxxxxxxxxxxxxxxx https://lists.freedesktop.org/mailman/listinfo/dri-devel-- Daniel Vetter Software Engineer, Intel Corporation http://blog.ffwll.ch
_______________________________________________ dri-devel mailing list dri-devel@xxxxxxxxxxxxxxxxxxxxx https://lists.freedesktop.org/mailman/listinfo/dri-devel
