iommufd: Fix a race with concurrent allocation and unmap

commit 8602018 upstream.

iopt_unmap_iova_range() releases the lock on iova_rwsem inside the loop
body when getting to the more expensive unmap operations. This is fine on
its own, except the loop condition is based on the first area that matches
the unmap address range. If a concurrent call to map picks an area that
was unmapped in previous iterations, the loop mistakenly tries to unmap
it.

This is reproducible by having one userspace thread map buffers and pass
them to another thread that unmaps them. The problem manifests as EBUSY
errors with single page mappings.

Fix this by advancing the start pointer after unmapping an area. This
ensures each iteration only examines the IOVA range that remains mapped,
which is guaranteed not to have overlaps.

Cc: stable@vger.kernel.org
Fixes: 51fe614 ("iommufd: Data structure to provide IOVA to PFN mapping")
Link: https://patch.msgid.link/r/CAAJpGJSR4r_ds1JOjmkqHtsBPyxu8GntoeW08Sk5RNQPmgi+tg@mail.gmail.com
Signed-off-by: Sina Hassani <sina@openai.com>
Reviewed-by: Kevin Tian <kevin.tian@intel.com>
Signed-off-by: Jason Gunthorpe <jgg@nvidia.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>

Author: sina-openai

drivers/iommu/iommufd/io_pagetable.c

@@ -724,6 +724,16 @@ static int iopt_unmap_iova_range(struct io_pagetable *iopt, unsigned long start,
 		unmapped_bytes += area_last - area_first + 1;
 
 		down_write(&iopt->iova_rwsem);
+
+		/*
+		 * After releasing the iova_rwsem concurrent allocation could
+		 * place new areas at IOVAs we have already unmapped. Keep
+		 * moving the start of the search forward to ignore the area
+		 * already unmapped.
+		 */
+		if (area_last >= last)
+			break;
+		start = area_last + 1;
 	}
 
 out_unlock_iova: