| Message ID | 20250408124042.2695955-3-amachhiw@linux.ibm.com (mailing list archive) |
|---|---|
| State | New |
| Headers | show |
| Series | vfio/spapr: Fix L2 crash with PCI device passthrough | expand |
On 4/8/25 14:40, Amit Machhiwal wrote: > An L2 KVM guest fails to boot inside a pSeries LPAR when booted with a > memory more than 128 GB and PCI device passthrough. The L2 guest also > crashes when it is booted with a memory greater than 128 GB and a PCI > device is hotplugged later. > > The issue arises from a conditional check for `levels > 1` in > `spapr_tce_create_table()` within L1 KVM. This check is meant to prevent > multi-level TCEs, which are not supported by the PowerVM hypervisor. As > a result, when QEMU makes a `VFIO_IOMMU_SPAPR_TCE_CREATE` ioctl call > with `levels > 1`, it triggers the conditional check and returns > `EINVAL`, causing the guest to crash with the following errors: > > 2025-03-04T06:36:36.133117Z qemu-system-ppc64: Failed to create a window, ret = -1 (Invalid argument) > 2025-03-04T06:36:36.133176Z qemu-system-ppc64: Failed to create SPAPR window: Invalid argument > qemu: hardware error: vfio: DMA mapping failed, unable to continue > > Fix this by checking the supported DDW "levels" returned by the > VFIO_IOMMU_SPAPR_TCE_GET_INFO ioctl before attempting the TCE create > ioctl in KVM. > > The patch has been tested on KVM guests with memory configurations of up > to 390GB, and 450GB on PowerVM and bare-metal environments respectively. > > Signed-off-by: Amit Machhiwal <amachhiw@linux.ibm.com> Reviewed-by: Cédric Le Goater <clg@redhat.com> Thanks, C. > --- > hw/vfio/spapr.c | 36 +++++++++++++++++++++++++++--------- > 1 file changed, 27 insertions(+), 9 deletions(-) > > diff --git a/hw/vfio/spapr.c b/hw/vfio/spapr.c > index dd9207679dbe..32611096fa29 100644 > --- a/hw/vfio/spapr.c > +++ b/hw/vfio/spapr.c > @@ -26,6 +26,7 @@ typedef struct VFIOSpaprContainer { > VFIOContainer container; > MemoryListener prereg_listener; > QLIST_HEAD(, VFIOHostDMAWindow) hostwin_list; > + unsigned int levels; > } VFIOSpaprContainer; > > OBJECT_DECLARE_SIMPLE_TYPE(VFIOSpaprContainer, VFIO_IOMMU_SPAPR); > @@ -236,9 +237,11 @@ static bool vfio_spapr_create_window(VFIOContainer *container, > { > int ret = 0; > VFIOContainerBase *bcontainer = &container->bcontainer; > + VFIOSpaprContainer *scontainer = container_of(container, VFIOSpaprContainer, > + container); > IOMMUMemoryRegion *iommu_mr = IOMMU_MEMORY_REGION(section->mr); > uint64_t pagesize = memory_region_iommu_get_min_page_size(iommu_mr), pgmask; > - unsigned entries, bits_total, bits_per_level, max_levels; > + unsigned entries, bits_total, bits_per_level, max_levels, ddw_levels; > struct vfio_iommu_spapr_tce_create create = { .argsz = sizeof(create) }; > long rampagesize = qemu_minrampagesize(); > > @@ -291,16 +294,29 @@ static bool vfio_spapr_create_window(VFIOContainer *container, > */ > bits_per_level = ctz64(qemu_real_host_page_size()) + 8; > create.levels = bits_total / bits_per_level; > - if (bits_total % bits_per_level) { > - ++create.levels; > - } > - max_levels = (64 - create.page_shift) / ctz64(qemu_real_host_page_size()); > - for ( ; create.levels <= max_levels; ++create.levels) { > - ret = ioctl(container->fd, VFIO_IOMMU_SPAPR_TCE_CREATE, &create); > - if (!ret) { > - break; > + > + ddw_levels = scontainer->levels; > + if (ddw_levels > 1) { > + if (bits_total % bits_per_level) { > + ++create.levels; > } > + max_levels = (64 - create.page_shift) / ctz64(qemu_real_host_page_size()); > + for ( ; create.levels <= max_levels; ++create.levels) { > + ret = ioctl(container->fd, VFIO_IOMMU_SPAPR_TCE_CREATE, &create); > + if (!ret) { > + break; > + } > + } > + } else { /* ddw_levels == 1 */ > + if (create.levels > ddw_levels) { > + error_setg_errno(errp, EINVAL, "Host doesn't support multi-level TCE tables" > + ". Use larger IO page size. Supported mask is 0x%lx", > + bcontainer->pgsizes); > + return false; > + } > + ret = ioctl(container->fd, VFIO_IOMMU_SPAPR_TCE_CREATE, &create); > } > + > if (ret) { > error_setg_errno(errp, errno, "Failed to create a window, ret = %d", ret); > return false; > @@ -501,6 +517,8 @@ static bool vfio_spapr_container_setup(VFIOContainerBase *bcontainer, > goto listener_unregister_exit; > } > > + scontainer->levels = info.ddw.levels; > + > if (v2) { > bcontainer->pgsizes = info.ddw.pgsizes; > /*
diff --git a/hw/vfio/spapr.c b/hw/vfio/spapr.c index dd9207679dbe..32611096fa29 100644 --- a/hw/vfio/spapr.c +++ b/hw/vfio/spapr.c @@ -26,6 +26,7 @@ typedef struct VFIOSpaprContainer { VFIOContainer container; MemoryListener prereg_listener; QLIST_HEAD(, VFIOHostDMAWindow) hostwin_list; + unsigned int levels; } VFIOSpaprContainer; OBJECT_DECLARE_SIMPLE_TYPE(VFIOSpaprContainer, VFIO_IOMMU_SPAPR); @@ -236,9 +237,11 @@ static bool vfio_spapr_create_window(VFIOContainer *container, { int ret = 0; VFIOContainerBase *bcontainer = &container->bcontainer; + VFIOSpaprContainer *scontainer = container_of(container, VFIOSpaprContainer, + container); IOMMUMemoryRegion *iommu_mr = IOMMU_MEMORY_REGION(section->mr); uint64_t pagesize = memory_region_iommu_get_min_page_size(iommu_mr), pgmask; - unsigned entries, bits_total, bits_per_level, max_levels; + unsigned entries, bits_total, bits_per_level, max_levels, ddw_levels; struct vfio_iommu_spapr_tce_create create = { .argsz = sizeof(create) }; long rampagesize = qemu_minrampagesize(); @@ -291,16 +294,29 @@ static bool vfio_spapr_create_window(VFIOContainer *container, */ bits_per_level = ctz64(qemu_real_host_page_size()) + 8; create.levels = bits_total / bits_per_level; - if (bits_total % bits_per_level) { - ++create.levels; - } - max_levels = (64 - create.page_shift) / ctz64(qemu_real_host_page_size()); - for ( ; create.levels <= max_levels; ++create.levels) { - ret = ioctl(container->fd, VFIO_IOMMU_SPAPR_TCE_CREATE, &create); - if (!ret) { - break; + + ddw_levels = scontainer->levels; + if (ddw_levels > 1) { + if (bits_total % bits_per_level) { + ++create.levels; } + max_levels = (64 - create.page_shift) / ctz64(qemu_real_host_page_size()); + for ( ; create.levels <= max_levels; ++create.levels) { + ret = ioctl(container->fd, VFIO_IOMMU_SPAPR_TCE_CREATE, &create); + if (!ret) { + break; + } + } + } else { /* ddw_levels == 1 */ + if (create.levels > ddw_levels) { + error_setg_errno(errp, EINVAL, "Host doesn't support multi-level TCE tables" + ". Use larger IO page size. Supported mask is 0x%lx", + bcontainer->pgsizes); + return false; + } + ret = ioctl(container->fd, VFIO_IOMMU_SPAPR_TCE_CREATE, &create); } + if (ret) { error_setg_errno(errp, errno, "Failed to create a window, ret = %d", ret); return false; @@ -501,6 +517,8 @@ static bool vfio_spapr_container_setup(VFIOContainerBase *bcontainer, goto listener_unregister_exit; } + scontainer->levels = info.ddw.levels; + if (v2) { bcontainer->pgsizes = info.ddw.pgsizes; /*
An L2 KVM guest fails to boot inside a pSeries LPAR when booted with a memory more than 128 GB and PCI device passthrough. The L2 guest also crashes when it is booted with a memory greater than 128 GB and a PCI device is hotplugged later. The issue arises from a conditional check for `levels > 1` in `spapr_tce_create_table()` within L1 KVM. This check is meant to prevent multi-level TCEs, which are not supported by the PowerVM hypervisor. As a result, when QEMU makes a `VFIO_IOMMU_SPAPR_TCE_CREATE` ioctl call with `levels > 1`, it triggers the conditional check and returns `EINVAL`, causing the guest to crash with the following errors: 2025-03-04T06:36:36.133117Z qemu-system-ppc64: Failed to create a window, ret = -1 (Invalid argument) 2025-03-04T06:36:36.133176Z qemu-system-ppc64: Failed to create SPAPR window: Invalid argument qemu: hardware error: vfio: DMA mapping failed, unable to continue Fix this by checking the supported DDW "levels" returned by the VFIO_IOMMU_SPAPR_TCE_GET_INFO ioctl before attempting the TCE create ioctl in KVM. The patch has been tested on KVM guests with memory configurations of up to 390GB, and 450GB on PowerVM and bare-metal environments respectively. Signed-off-by: Amit Machhiwal <amachhiw@linux.ibm.com> --- hw/vfio/spapr.c | 36 +++++++++++++++++++++++++++--------- 1 file changed, 27 insertions(+), 9 deletions(-)