| Message ID | 1460342864-27891-1-git-send-email-shawnguo@kernel.org (mailing list archive) |
|---|---|
| State | New, archived |
| Headers | show |
Hi Shawn, On 11/04/16 03:47, Shawn Guo wrote: > The macro sg_dma_len(sg) is commonly used to retrieve length of sg, > which could be 'dma_length' or 'length' field, depending on whether > NEED_SG_DMA_LENGTH is enabled or not. On the other hand, many driver > code use helper function sg_set_page() to set an sg entry pointing at > a page, with offset and length set up in one call. But sg_set_page() > does not consider NEED_SG_DMA_LENGTH case and only set up 'length' > field. This causes problem on platforms like ARM64, where > NEED_SG_DMA_LENGTH is enabled by default, i.e. sg_set_page() sets up > 'length' while sg_dma_len(sg) returns 'dma_length' field. > > The patch changes sg_set_page() to use sg_dma_len() for sg length setup > as well, so that NEED_SG_DMA_LENGTH case can be handled. > > Signed-off-by: Shawn Guo <shawnguo@kernel.org> > --- > include/linux/scatterlist.h | 2 +- > 1 file changed, 1 insertion(+), 1 deletion(-) > > diff --git a/include/linux/scatterlist.h b/include/linux/scatterlist.h > index 556ec1ea2574..b0e32ea594c3 100644 > --- a/include/linux/scatterlist.h > +++ b/include/linux/scatterlist.h > @@ -114,7 +114,7 @@ static inline void sg_set_page(struct scatterlist *sg, struct page *page, > { > sg_assign_page(sg, page); > sg->offset = offset; > - sg->length = len; > + sg_dma_len(sg) = len; This looks wrong. If a driver is building a scatterlist, then it needs to fill in the page, offset and length fields to describe the physical layout - leaving sg->length uninitialised would be a recipe for disaster - then pass it to the DMA API. Only the DMA API implementation should be setting dma_addr and dma_len, and they may not correspond to the physical layout at all (e.g. an IOMMU could concatenate the entire list into a single much longer segment at whatever arbitrary DMA address it chooses). Is there some particular driver hiccup behind this? Robin. > } > > static inline struct page *sg_page(struct scatterlist *sg) >
On Mon, Apr 11, 2016 at 11:31:04AM +0100, Robin Murphy wrote: > Hi Shawn, > > On 11/04/16 03:47, Shawn Guo wrote: > >The macro sg_dma_len(sg) is commonly used to retrieve length of sg, > >which could be 'dma_length' or 'length' field, depending on whether > >NEED_SG_DMA_LENGTH is enabled or not. On the other hand, many driver > >code use helper function sg_set_page() to set an sg entry pointing at > >a page, with offset and length set up in one call. But sg_set_page() > >does not consider NEED_SG_DMA_LENGTH case and only set up 'length' > >field. This causes problem on platforms like ARM64, where > >NEED_SG_DMA_LENGTH is enabled by default, i.e. sg_set_page() sets up > >'length' while sg_dma_len(sg) returns 'dma_length' field. > > > >The patch changes sg_set_page() to use sg_dma_len() for sg length setup > >as well, so that NEED_SG_DMA_LENGTH case can be handled. > > > >Signed-off-by: Shawn Guo <shawnguo@kernel.org> > >--- > > include/linux/scatterlist.h | 2 +- > > 1 file changed, 1 insertion(+), 1 deletion(-) > > > >diff --git a/include/linux/scatterlist.h b/include/linux/scatterlist.h > >index 556ec1ea2574..b0e32ea594c3 100644 > >--- a/include/linux/scatterlist.h > >+++ b/include/linux/scatterlist.h > >@@ -114,7 +114,7 @@ static inline void sg_set_page(struct scatterlist *sg, struct page *page, > > { > > sg_assign_page(sg, page); > > sg->offset = offset; > >- sg->length = len; > >+ sg_dma_len(sg) = len; > > This looks wrong. If a driver is building a scatterlist, then it > needs to fill in the page, offset and length fields to describe the > physical layout - leaving sg->length uninitialised would be a recipe > for disaster - then pass it to the DMA API. Only the DMA API > implementation should be setting dma_addr and dma_len, and they may > not correspond to the physical layout at all (e.g. an IOMMU could > concatenate the entire list into a single much longer segment at > whatever arbitrary DMA address it chooses). Thanks for the hint, Robin. Yes, you're right, the proposed change is just wrong. I just found the correct fix to my problem, i.e. commit 70bc916b2c80 (staging: android: ion: Set the length of the DMA sg entries in buffer). It should be applied for stable kernel, IMO. Shawn
diff --git a/include/linux/scatterlist.h b/include/linux/scatterlist.h index 556ec1ea2574..b0e32ea594c3 100644 --- a/include/linux/scatterlist.h +++ b/include/linux/scatterlist.h @@ -114,7 +114,7 @@ static inline void sg_set_page(struct scatterlist *sg, struct page *page, { sg_assign_page(sg, page); sg->offset = offset; - sg->length = len; + sg_dma_len(sg) = len; } static inline struct page *sg_page(struct scatterlist *sg)
The macro sg_dma_len(sg) is commonly used to retrieve length of sg, which could be 'dma_length' or 'length' field, depending on whether NEED_SG_DMA_LENGTH is enabled or not. On the other hand, many driver code use helper function sg_set_page() to set an sg entry pointing at a page, with offset and length set up in one call. But sg_set_page() does not consider NEED_SG_DMA_LENGTH case and only set up 'length' field. This causes problem on platforms like ARM64, where NEED_SG_DMA_LENGTH is enabled by default, i.e. sg_set_page() sets up 'length' while sg_dma_len(sg) returns 'dma_length' field. The patch changes sg_set_page() to use sg_dma_len() for sg length setup as well, so that NEED_SG_DMA_LENGTH case can be handled. Signed-off-by: Shawn Guo <shawnguo@kernel.org> --- include/linux/scatterlist.h | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-)