The way cookie_init_hw_msi_region() allocates the iommu_dma_msi_page structures doesn't match the way iommu_put_dma_cookie() frees them.
The former performs a single allocation of all the required structures, while the latter tries to free them one at a time. It doesn't quite work for the main use case (the GICv3 ITS where the range is 64kB) when the base granule size is 4kB.
This leads to a nice slab corruption on teardown, which is easily observable by simply creating a VF on a SRIOV-capable device, and tearing it down immediately (no need to even make use of it). Fortunately, this only affects systems where the ITS isn't translated by the SMMU, which are both rare and non-standard.
Fix it by allocating iommu_dma_msi_page structures one at a time.
Fixes: 7c1b058c8b5a3 ("iommu/dma: Handle IOMMU API reserved regions") Signed-off-by: Marc Zyngier maz@kernel.org Reviewed-by: Eric Auger eric.auger@redhat.com Cc: Robin Murphy robin.murphy@arm.com Cc: Joerg Roedel jroedel@suse.de Cc: Will Deacon will@kernel.org Cc: stable@vger.kernel.org --- * From v1: - Got rid of the superfluous error handling (Robin) - Clarified that it only affects a very small set of systems - Added Eric's RB (which I assume still stands)
drivers/iommu/dma-iommu.c | 16 ++++++++-------- 1 file changed, 8 insertions(+), 8 deletions(-)
diff --git a/drivers/iommu/dma-iommu.c b/drivers/iommu/dma-iommu.c index a2e96a5fd9a7..ba128d1cdaee 100644 --- a/drivers/iommu/dma-iommu.c +++ b/drivers/iommu/dma-iommu.c @@ -177,15 +177,15 @@ static int cookie_init_hw_msi_region(struct iommu_dma_cookie *cookie, start -= iova_offset(iovad, start); num_pages = iova_align(iovad, end - start) >> iova_shift(iovad);
- msi_page = kcalloc(num_pages, sizeof(*msi_page), GFP_KERNEL); - if (!msi_page) - return -ENOMEM; - for (i = 0; i < num_pages; i++) { - msi_page[i].phys = start; - msi_page[i].iova = start; - INIT_LIST_HEAD(&msi_page[i].list); - list_add(&msi_page[i].list, &cookie->msi_page_list); + msi_page = kmalloc(sizeof(*msi_page), GFP_KERNEL); + if (!msi_page) + return -ENOMEM; + + msi_page->phys = start; + msi_page->iova = start; + INIT_LIST_HEAD(&msi_page->list); + list_add(&msi_page->list, &cookie->msi_page_list); start += iovad->granule; }
On 04/03/2020 11:11 am, Marc Zyngier wrote:
The way cookie_init_hw_msi_region() allocates the iommu_dma_msi_page structures doesn't match the way iommu_put_dma_cookie() frees them.
The former performs a single allocation of all the required structures, while the latter tries to free them one at a time. It doesn't quite work for the main use case (the GICv3 ITS where the range is 64kB) when the base granule size is 4kB.
This leads to a nice slab corruption on teardown, which is easily observable by simply creating a VF on a SRIOV-capable device, and tearing it down immediately (no need to even make use of it). Fortunately, this only affects systems where the ITS isn't translated by the SMMU, which are both rare and non-standard.
Fix it by allocating iommu_dma_msi_page structures one at a time.
Reviewed-by: Robin Murphy robin.murphy@arm.com
Fixes: 7c1b058c8b5a3 ("iommu/dma: Handle IOMMU API reserved regions") Signed-off-by: Marc Zyngier maz@kernel.org Reviewed-by: Eric Auger eric.auger@redhat.com Cc: Robin Murphy robin.murphy@arm.com Cc: Joerg Roedel jroedel@suse.de Cc: Will Deacon will@kernel.org Cc: stable@vger.kernel.org
From v1:
- Got rid of the superfluous error handling (Robin)
- Clarified that it only affects a very small set of systems
- Added Eric's RB (which I assume still stands)
drivers/iommu/dma-iommu.c | 16 ++++++++-------- 1 file changed, 8 insertions(+), 8 deletions(-)
diff --git a/drivers/iommu/dma-iommu.c b/drivers/iommu/dma-iommu.c index a2e96a5fd9a7..ba128d1cdaee 100644 --- a/drivers/iommu/dma-iommu.c +++ b/drivers/iommu/dma-iommu.c @@ -177,15 +177,15 @@ static int cookie_init_hw_msi_region(struct iommu_dma_cookie *cookie, start -= iova_offset(iovad, start); num_pages = iova_align(iovad, end - start) >> iova_shift(iovad);
- msi_page = kcalloc(num_pages, sizeof(*msi_page), GFP_KERNEL);
- if (!msi_page)
return -ENOMEM;
- for (i = 0; i < num_pages; i++) {
msi_page[i].phys = start;
msi_page[i].iova = start;
INIT_LIST_HEAD(&msi_page[i].list);
list_add(&msi_page[i].list, &cookie->msi_page_list);
msi_page = kmalloc(sizeof(*msi_page), GFP_KERNEL);
if (!msi_page)
return -ENOMEM;
msi_page->phys = start;
msi_page->iova = start;
INIT_LIST_HEAD(&msi_page->list);
start += iovad->granule; }list_add(&msi_page->list, &cookie->msi_page_list);
On Wed, Mar 04, 2020 at 11:11:17AM +0000, Marc Zyngier wrote:
The way cookie_init_hw_msi_region() allocates the iommu_dma_msi_page structures doesn't match the way iommu_put_dma_cookie() frees them.
The former performs a single allocation of all the required structures, while the latter tries to free them one at a time. It doesn't quite work for the main use case (the GICv3 ITS where the range is 64kB) when the base granule size is 4kB.
This leads to a nice slab corruption on teardown, which is easily observable by simply creating a VF on a SRIOV-capable device, and tearing it down immediately (no need to even make use of it). Fortunately, this only affects systems where the ITS isn't translated by the SMMU, which are both rare and non-standard.
Fix it by allocating iommu_dma_msi_page structures one at a time.
Fixes: 7c1b058c8b5a3 ("iommu/dma: Handle IOMMU API reserved regions") Signed-off-by: Marc Zyngier maz@kernel.org Reviewed-by: Eric Auger eric.auger@redhat.com Cc: Robin Murphy robin.murphy@arm.com Cc: Joerg Roedel jroedel@suse.de Cc: Will Deacon will@kernel.org Cc: stable@vger.kernel.org
Applied for v5.6, thanks.
linux-stable-mirror@lists.linaro.org