[RFC PATCH 3/3] vfio/type1: Introduce pfn_list mutex

[Alex Williamson <alex.williamson@xxxxxxxxxx>]

We can promote external page {un}pinning to a reader lock, allowing
concurrency since these don't change the vfio_iommu state.  We do need
to protect the vpfn list per vfio_dma in place of that serialization
though.

Signed-off-by: Alex Williamson <alex.williamson@xxxxxxxxxx>
---
 drivers/vfio/vfio_iommu_type1.c |   24 +++++++++++++++++++-----
 1 file changed, 19 insertions(+), 5 deletions(-)

diff --git a/drivers/vfio/vfio_iommu_type1.c b/drivers/vfio/vfio_iommu_type1.c
index e78067cc74b3..ea63306c16f7 100644
--- a/drivers/vfio/vfio_iommu_type1.c
+++ b/drivers/vfio/vfio_iommu_type1.c
@@ -90,6 +90,7 @@ struct vfio_dma {
 	bool			iommu_mapped;
 	bool			lock_cap;	/* capable(CAP_IPC_LOCK) */
 	struct task_struct	*task;
+	struct mutex		pfn_list_lock;
 	struct rb_root		pfn_list;	/* Ex-user pinned pfn list */
 };
 
@@ -539,7 +540,7 @@ static int vfio_iommu_type1_pin_pages(void *iommu_data,
 	if (!iommu->v2)
 		return -EACCES;
 
-	down_write(&iommu->lock);
+	down_read(&iommu->lock);
 
 	/* Fail if notifier list is empty */
 	if (!iommu->notifier.head) {
@@ -570,8 +571,11 @@ static int vfio_iommu_type1_pin_pages(void *iommu_data,
 			goto pin_unwind;
 		}
 
+		mutex_lock(&dma->pfn_list_lock);
+
 		vpfn = vfio_iova_get_vfio_pfn(dma, iova);
 		if (vpfn) {
+			mutex_unlock(&dma->pfn_list_lock);
 			phys_pfn[i] = vpfn->pfn;
 			continue;
 		}
@@ -579,14 +583,19 @@ static int vfio_iommu_type1_pin_pages(void *iommu_data,
 		remote_vaddr = dma->vaddr + iova - dma->iova;
 		ret = vfio_pin_page_external(dma, remote_vaddr, &phys_pfn[i],
 					     do_accounting);
-		if (ret)
+		if (ret) {
+			mutex_unlock(&dma->pfn_list_lock);
 			goto pin_unwind;
+		}
 
 		ret = vfio_add_to_pfn_list(dma, iova, phys_pfn[i]);
 		if (ret) {
 			vfio_unpin_page_external(dma, iova, do_accounting);
+			mutex_unlock(&dma->pfn_list_lock);
 			goto pin_unwind;
 		}
+
+		mutex_unlock(&dma->pfn_list_lock);
 	}
 
 	ret = i;
@@ -599,11 +608,13 @@ static int vfio_iommu_type1_pin_pages(void *iommu_data,
 
 		iova = user_pfn[j] << PAGE_SHIFT;
 		dma = vfio_find_dma(iommu, iova, PAGE_SIZE);
+		mutex_lock(&dma->pfn_list_lock);
 		vfio_unpin_page_external(dma, iova, do_accounting);
+		mutex_unlock(&dma->pfn_list_lock);
 		phys_pfn[j] = 0;
 	}
 pin_done:
-	up_write(&iommu->lock);
+	up_read(&iommu->lock);
 	return ret;
 }
 
@@ -622,7 +633,7 @@ static int vfio_iommu_type1_unpin_pages(void *iommu_data,
 	if (!iommu->v2)
 		return -EACCES;
 
-	down_write(&iommu->lock);
+	down_read(&iommu->lock);
 
 	do_accounting = !IS_IOMMU_CAP_DOMAIN_IN_CONTAINER(iommu);
 	for (i = 0; i < npage; i++) {
@@ -633,11 +644,13 @@ static int vfio_iommu_type1_unpin_pages(void *iommu_data,
 		dma = vfio_find_dma(iommu, iova, PAGE_SIZE);
 		if (!dma)
 			goto unpin_exit;
+		mutex_lock(&dma->pfn_list_lock);
 		vfio_unpin_page_external(dma, iova, do_accounting);
+		mutex_unlock(&dma->pfn_list_lock);
 	}
 
 unpin_exit:
-	up_write(&iommu->lock);
+	up_read(&iommu->lock);
 	return i > npage ? npage : (i > 0 ? i : -EINVAL);
 }
 
@@ -1109,6 +1122,7 @@ static int vfio_dma_do_map(struct vfio_iommu *iommu,
 	dma->iova = iova;
 	dma->vaddr = vaddr;
 	dma->prot = prot;
+	mutex_init(&dma->pfn_list_lock);
 
 	/*
 	 * We need to be able to both add to a task's locked memory and test