diff mbox series

[RFC,4/5] RDMA/mlx5: Add fallback for P2P DMA errors

Message ID 20241201103659.420677-5-ymaman@nvidia.com (mailing list archive)
State New
Headers show
Series GPU Direct RDMA (P2P DMA) for Device Private Pages | expand

Commit Message

Yonatan Maman Dec. 1, 2024, 10:36 a.m. UTC 
From: Yonatan Maman <Ymaman@Nvidia.com>

Handle P2P DMA mapping errors when the transaction requires traversing
an inaccessible host bridge that is not in the allowlist:

- In `populate_mtt`, if a P2P mapping fails, the `HMM_PFN_ALLOW_P2P` flag
  is cleared only for the PFNs that returned a mapping error.

- In `pagefault_real_mr`, if a P2P mapping error occurs, the mapping is
  retried with the `HMM_PFN_ALLOW_P2P` flag only for the PFNs that didn't
  fail, ensuring a fallback to standard DMA(host memory) for the rest,
if possible.

Signed-off-by: Yonatan Maman <Ymaman@Nvidia.com>
Signed-off-by: Gal Shalom <GalShalom@Nvidia.com>
---
 drivers/infiniband/hw/mlx5/odp.c | 24 +++++++++++++++++++++---
 1 file changed, 21 insertions(+), 3 deletions(-)
diff mbox series

Patch

diff --git a/drivers/infiniband/hw/mlx5/odp.c b/drivers/infiniband/hw/mlx5/odp.c
index fbb2a5670c32..f7a1291ec7d1 100644
--- a/drivers/infiniband/hw/mlx5/odp.c
+++ b/drivers/infiniband/hw/mlx5/odp.c
@@ -169,6 +169,7 @@  static int populate_mtt(__be64 *pas, size_t start, size_t nentries,
 	struct pci_p2pdma_map_state p2pdma_state = {};
 	struct ib_device *dev = odp->umem.ibdev;
 	size_t i;
+	int ret = 0;
 
 	if (flags & MLX5_IB_UPD_XLT_ZAP)
 		return 0;
@@ -184,8 +185,11 @@  static int populate_mtt(__be64 *pas, size_t start, size_t nentries,
 
 		dma_addr = hmm_dma_map_pfn(dev->dma_device, &odp->map,
 					   start + i, &p2pdma_state);
-		if (ib_dma_mapping_error(dev, dma_addr))
-			return -EFAULT;
+		if (ib_dma_mapping_error(dev, dma_addr)) {
+			odp->map.pfn_list[start + i] &= ~(HMM_PFN_ALLOW_P2P);
+			ret = -EFAULT;
+			continue;
+		}
 
 		dma_addr |= MLX5_IB_MTT_READ;
 		if ((pfn & HMM_PFN_WRITE) && !downgrade)
@@ -194,7 +198,7 @@  static int populate_mtt(__be64 *pas, size_t start, size_t nentries,
 		pas[i] = cpu_to_be64(dma_addr);
 		odp->npages++;
 	}
-	return 0;
+	return ret;
 }
 
 int mlx5_odp_populate_xlt(void *xlt, size_t idx, size_t nentries,
@@ -696,6 +700,10 @@  static int pagefault_real_mr(struct mlx5_ib_mr *mr, struct ib_umem_odp *odp,
 	if (odp->umem.writable && !downgrade)
 		access_mask |= HMM_PFN_WRITE;
 
+	/*
+	 * try fault with HMM_PFN_ALLOW_P2P flag
+	 */
+	access_mask |= HMM_PFN_ALLOW_P2P;
 	np = ib_umem_odp_map_dma_and_lock(odp, user_va, bcnt, access_mask, fault);
 	if (np < 0)
 		return np;
@@ -705,6 +713,16 @@  static int pagefault_real_mr(struct mlx5_ib_mr *mr, struct ib_umem_odp *odp,
 	 * ib_umem_odp_map_dma_and_lock already checks this.
 	 */
 	ret = mlx5r_umr_update_xlt(mr, start_idx, np, page_shift, xlt_flags);
+	if (ret == -EFAULT) {
+		/*
+		 * Indicate P2P Mapping Error, retry with no HMM_PFN_ALLOW_P2P
+		 */
+		access_mask &= ~HMM_PFN_ALLOW_P2P;
+		np = ib_umem_odp_map_dma_and_lock(odp, user_va, bcnt, access_mask, fault);
+		if (np < 0)
+			return np;
+		ret = mlx5r_umr_update_xlt(mr, start_idx, np, page_shift, xlt_flags);
+	}
 	mutex_unlock(&odp->umem_mutex);
 
 	if (ret < 0) {