mlx4_core: Fix possible chunk sg list overflow in mlx4_alloc_icm()
authorSebastien Dugue <sebastien.dugue@bull.net>
Thu, 20 May 2010 22:58:22 +0000 (15:58 -0700)
committerRoland Dreier <rolandd@cisco.com>
Thu, 20 May 2010 22:58:22 +0000 (15:58 -0700)
If the number of sg entries in the ICM chunk reaches MLX4_ICM_CHUNK_LEN,
we must set chunk to NULL even for coherent mappings so that the next
time through the loop will allocate another chunk.  Otherwise we'll
overflow the sg list the next time through the loop.  This will lead to
memory corruption if this case is hit.

mthca does not have this bug.

Signed-off-by: Sebastien Dugue <sebastien.dugue@bull.net>
Cc: <stable@kernel.org>
Signed-off-by: Roland Dreier <rolandd@cisco.com>
drivers/net/mlx4/icm.c

index 57288ca1395f1e6b0974655396afb5341566b37a..ef62f1749b8a47e1d13aa9ed117b78d9a5364605 100644 (file)
@@ -175,9 +175,10 @@ struct mlx4_icm *mlx4_alloc_icm(struct mlx4_dev *dev, int npages,
 
                                if (chunk->nsg <= 0)
                                        goto fail;
+                       }
 
+                       if (chunk->npages == MLX4_ICM_CHUNK_LEN)
                                chunk = NULL;
-                       }
 
                        npages -= 1 << cur_order;
                } else {