On Wed, Jun 23, 2021 at 09:10:05AM -0300, Jason Gunthorpe wrote:

On Wed, Jun 23, 2021 at 08:23:17AM +0300, Leon Romanovsky wrote:

quoted

On Tue, Jun 22, 2021 at 10:18:16AM -0300, Jason Gunthorpe wrote:

quoted

On Tue, Jun 22, 2021 at 02:39:42PM +0300, Leon Romanovsky wrote:

quoted

diff --git a/drivers/infiniband/core/umem.c b/drivers/infiniband/core/umem.c
index 0eb40025075f..a76ef6a6bac5 100644
+++ b/drivers/infiniband/core/umem.c

@@ -51,11 +51,11 @@ static void __ib_umem_release(struct ib_device *dev, struct ib_umem *umem, int d
 	struct scatterlist *sg;
 	unsigned int i;
 
-	if (umem->nmap > 0)
-		ib_dma_unmap_sg(dev, umem->sg_head.sgl, umem->sg_nents,
-				DMA_BIDIRECTIONAL);
+	if (dirty)
+		ib_dma_unmap_sgtable_attrs(dev, &umem->sg_head,
+					   DMA_BIDIRECTIONAL, 0);
 
-	for_each_sg(umem->sg_head.sgl, sg, umem->sg_nents, i)
+	for_each_sgtable_dma_sg(&umem->sg_head, sg, i)
 		unpin_user_page_range_dirty_lock(sg_page(sg),
 			DIV_ROUND_UP(sg->length, PAGE_SIZE), make_dirty);

This isn't right, can't mix sg_page with a _dma_ API

Jason, why is that?

We use same pages that were passed to __sg_alloc_table_from_pages() in __ib_umem_get().

A sgl has two lists inside it a 'dma' list and a 'page' list, they are
not the same length and not interchangable.

If you use for_each_sgtable_dma_sg() then you iterate over the 'dma'
list and have to use 'dma' accessors

If you use for_each_sgtable_sg() then you iterate over the 'page' list
and have to use 'page' acessors

Mixing dma iteration with page accessors or vice-versa, like above, is
always a bug.

You can see it alos because the old code used umem->sg_nents which is
the CPU list length while this new code is using the dma list length.

Ohh, I see difference between types now, thank you for the explanation,
will consult with Maor once he returns next week to the office.

Thanks

Jason