Kernel commit 492855939bdb added a limit to the number of outstanding DMA requests for a type1 vfio container. However, lazy unmapping in s390 can in fact cause quite a large number of outstanding DMA requests to build up prior to being purged, potentially the entire guest DMA space. This results in unexpected 'VFIO_MAP_DMA failed: No space left on device' conditions seen in QEMU.
This patchset adds support to qemu to retrieve the number of allowable DMA requests via the VFIO_IOMMU_GET_INFO ioctl. The patches are separated into vfio hits which add support for reading in VFIO_IOMMU_GET_INFO capability chains and getting the per-container dma_avail value, and s390 hits to track DMA usage on a per-container basis. Associated kernel patch: https://marc.info/?l=kvm&m=160012235303068&w=2 Changes from v2: - Renamed references of dma_limit to dma_avail based on kernel change as the ioctl now reports the amount currently available vs the supposed limit. - Because the ioctl now provides a 'living value' vs the limit, it doesn't seem appropriate to put it in the VFIOConatiner structure without doing the dma_avail tracking there. So patch 1 shrinks to vfio helper routines to find the capability and leaves vfio_connect_container alone. - Subsequently, in patch 2 s390-pci issues its own VFIO_IOMMU_GET_INFO to read the dma_avail value and use it. Matthew Rosato (3): vfio: Find DMA available capability s390x/pci: Honor DMA limits set by vfio vfio: Create shared routine for scanning info capabilities hw/s390x/s390-pci-bus.c | 99 ++++++++++++++++++++++++++++++++++++++++--- hw/s390x/s390-pci-bus.h | 9 ++++ hw/s390x/s390-pci-inst.c | 29 ++++++++++--- hw/s390x/s390-pci-inst.h | 3 ++ hw/vfio/common.c | 50 ++++++++++++++++++---- include/hw/vfio/vfio-common.h | 2 + 6 files changed, 173 insertions(+), 19 deletions(-) -- 1.8.3.1