On 2021-10-12 11:34 p.m., Bart Van Assche wrote:

On 10/12/21 17:30, Logan Gunthorpe wrote:

quoted

Best I can see from the code is that someone is passing an sg_cnt of
zero. Previously that would have returned -ENOMEM, but now it might be
ignored, in which case it would hit that WARNING and return -EIO.

That is not what is happening. The debug patch shown below learned me
the following:
* The sg_cnt argument of rdma_rw_ctx_init() is not zero.
* After the rdma_rw_map_sgtable() call, sgt.nents is zero.

The debug patch that I used is as follows:

Ah, hmm. Perhaps it's this... The virt path in
ib_dma_map_sgtable_attrs() doesn't set the sgt.nents...

Maybe try this something like the patch below.

Thanks,

Logan

--

diff --git a/include/rdma/ib_verbs.h b/include/rdma/ib_verbs.h
index 4b50d9a3018a..4ba642fc8a19 100644
--- a/include/rdma/ib_verbs.h
+++ b/include/rdma/ib_verbs.h

@@ -4097,8 +4097,13 @@ static inline int ib_dma_map_sgtable_attrs(struct ib_dev>
                                           enum dma_data_direction direction,
                                           unsigned long dma_attrs)
 {
+       int nents;
+
        if (ib_uses_virt_dma(dev)) {
-               ib_dma_virt_map_sg(dev, sgt->sgl, sgt->orig_nents);
+               nents = ib_dma_virt_map_sg(dev, sgt->sgl, sgt->orig_nents);
+               if (!nents)
+                       return -EIO;
+               sgt->nents = nents;
                return 0;
        }
        return dma_map_sgtable(dev->dma_device, sgt, direction, dma_attrs);