On 03/17/2016 05:10 PM, David Gibson wrote:
On Thu, Mar 17, 2016 at 04:04:29PM +1100, Alexey Kardashevskiy wrote:On 03/15/2016 04:42 PM, David Gibson wrote:On Tue, Mar 15, 2016 at 01:53:48PM +1100, Alexey Kardashevskiy wrote:On 03/03/2016 05:30 PM, David Gibson wrote:On Tue, Mar 01, 2016 at 08:10:36PM +1100, Alexey Kardashevskiy wrote:This makes use of the new "memory registering" feature. The idea is to provide the userspace ability to notify the host kernel about pages which are going to be used for DMA. Having this information, the host kernel can pin them all once per user process, do locked pages accounting (once) and not spent time on doing that in real time with possible failures which cannot be handled nicely in some cases.This adds a prereg memory listener which listens on address_space_memory and notifies a VFIO container about memory which needs to be pinned/unpinned. VFIO MMIO regions (i.e. "skip dump" regions) are skipped. As there is no per-IOMMU-type release() callback anymore, this stores the IOMMU type in the container so vfio_listener_release() can device if it needs to unregister @prereg_listener. The feature is only enabled for SPAPR IOMMU v2. The host kernel changes are required. Since v2 does not need/support VFIO_IOMMU_ENABLE, this does not call it when v2 is detected and enabled. This does not change the guest visible interface. Signed-off-by: Alexey Kardashevskiy <a...@ozlabs.ru> --- hw/vfio/Makefile.objs | 1 + hw/vfio/common.c | 39 +++++++++--- hw/vfio/prereg.c | 138 ++++++++++++++++++++++++++++++++++++++++++ include/hw/vfio/vfio-common.h | 4 ++ trace-events | 2 + 5 files changed, 175 insertions(+), 9 deletions(-) create mode 100644 hw/vfio/prereg.c diff --git a/hw/vfio/Makefile.objs b/hw/vfio/Makefile.objs index ceddbb8..5800e0e 100644 --- a/hw/vfio/Makefile.objs +++ b/hw/vfio/Makefile.objs @@ -4,4 +4,5 @@ obj-$(CONFIG_PCI) += pci.o pci-quirks.o obj-$(CONFIG_SOFTMMU) += platform.o obj-$(CONFIG_SOFTMMU) += calxeda-xgmac.o obj-$(CONFIG_SOFTMMU) += amd-xgbe.o +obj-$(CONFIG_SOFTMMU) += prereg.o endif diff --git a/hw/vfio/common.c b/hw/vfio/common.c index 3aaa6b5..f2a03e0 100644 --- a/hw/vfio/common.c +++ b/hw/vfio/common.c @@ -531,6 +531,9 @@ static const MemoryListener vfio_iommu_listener = { static void vfio_listener_release(VFIOContainer *container) { memory_listener_unregister(&container->iommu_listener.listener); + if (container->iommu_type == VFIO_SPAPR_TCE_v2_IOMMU) { + memory_listener_unregister(&container->prereg_listener.listener); + } } int vfio_mmap_region(Object *obj, VFIORegion *region, @@ -722,8 +725,8 @@ static int vfio_connect_container(VFIOGroup *group, AddressSpace *as) goto free_container_exit; } - ret = ioctl(fd, VFIO_SET_IOMMU, - v2 ? VFIO_TYPE1v2_IOMMU : VFIO_TYPE1_IOMMU); + container->iommu_type = v2 ? VFIO_TYPE1v2_IOMMU : VFIO_TYPE1_IOMMU; + ret = ioctl(fd, VFIO_SET_IOMMU, container->iommu_type); if (ret) { error_report("vfio: failed to set iommu for container: %m"); ret = -errno; @@ -748,8 +751,10 @@ static int vfio_connect_container(VFIOGroup *group, AddressSpace *as) if ((ret == 0) && (info.flags & VFIO_IOMMU_INFO_PGSIZES)) { container->iova_pgsizes = info.iova_pgsizes; } - } else if (ioctl(fd, VFIO_CHECK_EXTENSION, VFIO_SPAPR_TCE_IOMMU)) { + } else if (ioctl(fd, VFIO_CHECK_EXTENSION, VFIO_SPAPR_TCE_IOMMU) || + ioctl(fd, VFIO_CHECK_EXTENSION, VFIO_SPAPR_TCE_v2_IOMMU)) { struct vfio_iommu_spapr_tce_info info; + bool v2 = !!ioctl(fd, VFIO_CHECK_EXTENSION, VFIO_SPAPR_TCE_v2_IOMMU); ret = ioctl(group->fd, VFIO_GROUP_SET_CONTAINER, &fd); if (ret) { @@ -757,7 +762,9 @@ static int vfio_connect_container(VFIOGroup *group, AddressSpace *as) ret = -errno; goto free_container_exit; } - ret = ioctl(fd, VFIO_SET_IOMMU, VFIO_SPAPR_TCE_IOMMU); + container->iommu_type = + v2 ? VFIO_SPAPR_TCE_v2_IOMMU : VFIO_SPAPR_TCE_IOMMU; + ret = ioctl(fd, VFIO_SET_IOMMU, container->iommu_type);It'd be nice to consolidate the setting of container->iommu_type and then the SET_IOMMU ioctl() rather than having more or less duplicated logic for Type1 and SPAPR, but it's not a big deal.May be but I cannot think of any nice way of doing this though.if (ret) { error_report("vfio: failed to set iommu for container: %m"); ret = -errno; @@ -769,11 +776,25 @@ static int vfio_connect_container(VFIOGroup *group, AddressSpace *as) * when container fd is closed so we do not call it explicitly * in this file. */ - ret = ioctl(fd, VFIO_IOMMU_ENABLE); - if (ret) { - error_report("vfio: failed to enable container: %m"); - ret = -errno; - goto free_container_exit; + if (!v2) { + ret = ioctl(fd, VFIO_IOMMU_ENABLE); + if (ret) { + error_report("vfio: failed to enable container: %m"); + ret = -errno; + goto free_container_exit; + } + } else { + container->prereg_listener.container = container; + container->prereg_listener.listener = vfio_prereg_listener; + + memory_listener_register(&container->prereg_listener.listener, + &address_space_memory);This assumes that the target address space of the (guest) IOMMU is address_space_memory. Which is fine - vfio already assumes that - but it reminds me that it'd be nice to have an explicit check for that (I guess it would have to go in vfio_iommu_map_notify()). So that if someone constructs a machine where that's not the case, it'll at least be obvious why VFIO isn't working.Ok, I'll add a small patch for this in the next respin.Ok.+ if (container->error) { + error_report("vfio: RAM memory listener initialization failed for container"); + memory_listener_unregister( + &container->prereg_listener.listener); + goto free_container_exit; + } }Looks like you don't have an error path which will handle the case where the prereg listener is registered, but registering the normal PCI AS listener fails - I believe you will fail to unregister the prereg listener in that case.In this case, the control goes to listener_release_exit: which calls vfio_listener_release() which unregisters both listeners (it is a few chunks above).Ah.. yes. In which case this could also jump to listener_release_exit and avoid the explicit unreg(), yes?Sorry, I do not follow you here. It does jump to listener_release_exit already.I mean you can use the listener_release_exit label on failure of the prereg listener as well as failure of the regular listener.
When vfio_prereg_listener fails, vfio_memory_listener is not registered - it just looks cleaner to jump further to free_container_exit than to listener_release_exit, does not it?
/* diff --git a/hw/vfio/prereg.c b/hw/vfio/prereg.c new file mode 100644 index 0000000..66cd696 --- /dev/null +++ b/hw/vfio/prereg.c @@ -0,0 +1,138 @@ +/* + * DMA memory preregistration + * + * Authors: + * Alexey Kardashevskiy <a...@ozlabs.ru> + * + * This work is licensed under the terms of the GNU GPL, version 2. See + * the COPYING file in the top-level directory. + */ + +#include "qemu/osdep.h" +#include <sys/ioctl.h> +#include <linux/vfio.h> + +#include "hw/vfio/vfio-common.h" +#include "hw/vfio/vfio.h" +#include "qemu/error-report.h" +#include "trace.h" + +static bool vfio_prereg_listener_skipped_section(MemoryRegionSection *section) +{ + return (!memory_region_is_ram(section->mr) && + !memory_region_is_iommu(section->mr)) || + memory_region_is_skip_dump(section->mr); +} + +static void vfio_prereg_listener_region_add(MemoryListener *listener, + MemoryRegionSection *section) +{ + VFIOMemoryListener *vlistener = container_of(listener, VFIOMemoryListener, + listener); + VFIOContainer *container = vlistener->container; + hwaddr iova; + Int128 llend; + int ret; + hwaddr page_mask = qemu_real_host_page_mask; + struct vfio_iommu_spapr_register_memory reg = { + .argsz = sizeof(reg), + .flags = 0, + }; + + if (vfio_prereg_listener_skipped_section(section)) { + trace_vfio_listener_region_add_skip( + section->offset_within_address_space, + section->offset_within_address_space + + int128_get64(int128_sub(section->size, int128_one()))); + return; + }You should probably explicitly check for IOMMU regions and abort if you find one. An IOMMU AS appearing within address_space_memory would be bad news.Oh, vfio_prereg_listener_skipped_section() allows memory_region_is_iommu(), I'll remove it.Well, that's part. But IOMMU regions appearing here shouldn't just be ignored - they should be treated as a fatal error.Is this always an error when they appear in &address_space_memory? Because if it is so, we should do this check somewhere in memory_region_transaction_commit() (memory_region_add_subregion() cannot check this, there is no AS).Hmm.. not necessarily, although it would certainly be strange. I can imagine some sort of off-board memory device which incorporates an "IO"MMU which will remap requests both from the cpu and from DMA devices. But, that wouldn't work with VFIO, since we no longer know where the host memory is which is backing address_space_memory to preregister it. Well.. it might be possible by looking at the *target* AS of an iommu registered in address_space_memory, setting up a listener on that, and keeping on going until you reach a real memory region. The case is so unlikely that it's not worth actually implementing the code for. But I think it is worth a sanity check.
Ok then. g_assert() or hwerror()? -- Alexey
