On 01/27/2016 06:43 PM, Alex Williamson wrote:
On Wed, 2016-01-27 at 10:28 +0100, Pierre Morel wrote:On 01/26/2016 06:00 PM, Alex Williamson wrote:Alex, I think we misunderstood us.On Tue, 2016-01-26 at 15:51 +0100, Pierre Morel wrote:On 01/22/2016 11:19 PM, Alex Williamson wrote:On Fri, 2016-01-22 at 15:14 -0700, Alex Williamson wrote:On Thu, 2016-01-21 at 14:15 +0100, Pierre Morel wrote:On 01/20/2016 04:46 PM, Alex Williamson wrote:On Wed, 2016-01-20 at 16:14 +0100, Pierre Morel wrote:On 01/12/2016 07:16 PM, Alex Williamson wrote:On Tue, 2016-01-12 at 16:11 +0100, Pierre Morel wrote:In vfio_listener_region_add(), we try to validate that the region is not zero sized and hasn't overflowed the addresses space.But the calculation uses the size of the region instead ofusing the region's limit (size - 1).This leads to Int128 overflow when the region hasbeen initialized to UINT64_MAX because in this case memory_region_init() transform the size from UINT64_MAX to int128_2_64().Let's really use the limit by sustracting one to the sizeand take care to use the limit for functions using limit and size to call functions which need size.Signed-off-by: Pierre Morel <pmo...@linux.vnet.ibm.com>---Changes from v2:- all, just ignore v2, sorry about this, this is build after v1Changes from v1:- adjust the tests by knowing we already substracted one to end.hw/vfio/common.c | 14 +++++++-------1 files changed, 7 insertions(+), 7 deletions(-)diff --git a/hw/vfio/common.c b/hw/vfio/common.cindex 6797208..a5f6643 100644 --- a/hw/vfio/common.c +++ b/hw/vfio/common.c @@ -348,12 +348,12 @@ static void vfio_listener_region_add(MemoryListener *listener, if (int128_ge(int128_make64(iova), llend)) { return; } - end = int128_get64(llend); + end = int128_get64(int128_sub(llend, int128_one()));- if ((iova < container->min_iova) || ((end - 1) > container-max_iova)) {+ if ((iova < container->min_iova) || (end > container-max_iova)) {error_report("vfio: IOMMU container %p can't map guest IOVA region" " 0x%"HWADDR_PRIx"..0x%"HWADDR_PRIx, - container, iova, end - 1); + container, iova, end); ret = -EFAULT; goto fail; } @@ -363,7 +363,7 @@ static void vfio_listener_region_add(MemoryListener *listener, if (memory_region_is_iommu(section->mr)) { VFIOGuestIOMMU *giommu;- trace_vfio_listener_region_add_iommu(iova, end - 1);+ trace_vfio_listener_region_add_iommu(iova, end); /* * FIXME: We should do some checking to see if the * capabilities of the host VFIO IOMMU are adequate to model @@ -394,13 +394,13 @@ static void vfio_listener_region_add(MemoryListener *listener, section->offset_within_region + (iova - section->offset_within_address_space);- trace_vfio_listener_region_add_ram(iova, end - 1, vaddr);+ trace_vfio_listener_region_add_ram(iova, end, vaddr);- ret = vfio_dma_map(container, iova, end - iova, vaddr,section-readonly);+ ret = vfio_dma_map(container, iova, end - iova + 1, vaddr, section->readonly); if (ret) { error_report("vfio_dma_map(%p, 0x%"HWADDR_PRIx", " "0x%"HWADDR_PRIx", %p) = %d (%m)", - container, iova, end - iova, vaddr, ret); + container, iova, end - iova + 1, vaddr, ret); goto fail; }Hmm, did we just push the overflow from one place to another? If we're mapping a full region of size int128_2_64() starting at iova zero, then this becomes (0xffff_ffff_ffff_ffff - 0 + 1) = 0. So I think we need to calculate size with 128bit arithmetic too and let it assert if we overflow, ie:diff --git a/hw/vfio/common.c b/hw/vfio/common.cindex a5f6643..13ad90b 100644 --- a/hw/vfio/common.c +++ b/hw/vfio/common.c @@ -321,7 +321,7 @@ static void vfio_listener_region_add(MemoryListener *listener, MemoryRegionSection *section) { VFIOContainer *container = container_of(listener, VFIOContainer, listener); - hwaddr iova, end; + hwaddr iova, end, size; Int128 llend; void *vaddr; int ret; @@ -348,7 +348,9 @@ static void vfio_listener_region_add(MemoryListener *listener, if (int128_ge(int128_make64(iova), llend)) { return; } + end = int128_get64(int128_sub(llend, int128_one())); + size = int128_get64(int128_sub(llend, int128_make64(iova)));here again, if iova is null, since llend is section->size (2^64) ...if ((iova < container->min_iova) || (end > container-max_iova)) {error_report("vfio: IOMMU container %p can't map guest IOVA region" @@ -396,11 +398,11 @@ static void vfio_listener_region_add(MemoryListener *listener,trace_vfio_listener_region_add_ram(iova, end, vaddr); - ret = vfio_dma_map(container, iova, end - iova + 1, vaddr,section->readonly); + ret = vfio_dma_map(container, iova, size, vaddr, section-readonly);if (ret) { error_report("vfio_dma_map(%p, 0x%"HWADDR_PRIx", " "0x%"HWADDR_PRIx", %p) = %d (%m)", - container, iova, end - iova + 1, vaddr, ret); + container, iova, size, vaddr, ret); goto fail; }Does that still solve your scenario? Perhaps vfio-iommu-type1should have used first/last rather than start/size for mapping since we seem to have an off-by-one for mapping a full 64bit space. Seems like we could do it with two calls to vfio_dma_map if we really wanted to. Thanks,AlexYou are right, every try to solve this will push the overflow somewhere else.There is just no way to express 2^64 with 64 bits, we have theint128() solution, but if we solve it here, we fall in the linux ioctl call anyway.Intuitively, making two calls do not seem right to me. But, what do you think of something like: - creating a new VFIO extention - and in ioctl(), since we have a flag entry in thevfio_iommu_type1_dma_map, may be adding a new flag meaning "map all virtual memory" ? or meaning "use first/last" ? I think this would break existing code unless we add a new VFIO extension.Backup, is there ever a case where we actually need to map the entire 64bit address space? This is fairly well impossible on x86. I'm pointing out an issue, but I don't know that we need to solve it with more than an assert since it's never likely to happen. Thanks,AlexIf I understood right, IOVA is the IO virtual address, it is then possible to map the virtual address page 0xffff_ffff_ffff_f000 to something reasonable inside the real memory.It is.Eventual we do not need to map the last virtual page but I think that in a general case the all virtual memory, as viewed by the device through the IOMMU should be mapped to avoid any uninitialized virtual memory access.When using vfio, a device only has access to the IOVA space which has been explicitly mapped. This would be a security issue otherwise since kernel vfio can't rely on userspace to wipe the device IOVA space.yes.Alex, I am not sure of that, when preparing DMA from the device, theIt is the same reason that make us map the all virtual memory for the CPU MMU.We don't really do that either, CPU mapping works based on page tables and non-existent entries simply don't exist. We don't fully populate the page tables in advance, this would be a horrible waste of memory.guest will provide the destination address and these destination addresses will be translated by the IOMMU when the device start the DMA.The guest can make any decision by preparing the DMA and if I have wellunderstood, this is transparent to QEMU. What is not transparent is the IOMMU translation.Then, when the device starts the DMA the destination address can beanything inside the virtual memory and the IOMMU will translate this. To be able to translate, a table entry for this virtual address must exist in the IOMMU page table.If you have several level of page table you may only fill the first levelfor all entries, and may be, have only one first level entry initialized and the belonging second level entries filled. Which greatly reduces the size of the tables.But if you do not fill one of the first level entry the behavior of theIOMMU and then of the DMA is done according to what ever has been left in this entry.It seems like you're arguing that the guest is going to have a 2^64 bit address space for DMA targets. Sure, the guest driver can program the device to DMA anywhere in that address space, but what should the IOMMU actually consider an valid DMA? It has to be things that have been mapped, like guest physical RAM or peer-to-peer DMA targets. How would the IOMMU handle a stray DMA for anything else? By default the DMA target would not be mapped and the DMA would generate a fault. How that fault is handled is specific to the architecture and platform. If you want to provide the device with a full 2^64 bit address space where nothing will fault, you're going to need to do it via a lot of mappings pointing to the same host physical page. I'm really not sure where we're going with this though.As I do not know if I do not understand you or if you do not understand me, I try to explain me better. You may find the description obvious, the all is ok.In the architectures I know, DMA and IOMMU are two separate things.IOMMU must initialize table entries for the all space DMA is able to access. It must do it, no exception allowed. But some, and may be quite all the entries in the table are invalid entries, existing entries with the invalid bit (or no valid bit) set.Most, if not all IOMMU have several level of tables.Suppose a two level table mapping a 2^32 memory space. You have first level with 4k entries with an indirection to the second page level. Second page level mapping 256 entries of 4k pages.In this case, to have a valid page table accessing 1 byte somewhere inmemory you have: 4k - 1 entries of first level with invalid bit set 1 first level entry which is valid and point to a second level tablein second level table you have 255 entries with invalid bit setone entry valid pointing to a page where the byte is.This just to illustrate what I want to say when I say all entries areinitialized. If one let garbage inside one entry it happens random things... we do not want that.Are we OK with this?Yes, basic page tables.My fault, I use the word map indifferently to refer to a valid mappingMay be I missed something, or may be I worry too much, but I see this as a restriction on the supported hardware if we compare host and guest hardware support compatibility.I don't see the issue, there's arguably a bug in the API that doesn't allow us to map the full 64bit IOVA space of a device in a single mapping, but we can do it in two. Besides, there's really no case where a device needs a fully populated IOTLB unless you're actually giving the device access to 16 EMB of memory.s/EMB/EB/ Or I suppose technically EiByes, I agree with this, we do not need to access so much memory.We can live with it, because in fact you are right and today I am not aware of a hardware wanting to access this page but a hardware designers knowing having a IOMMU may want to access exactly this kind of strange virtual page for special features and this would work on the host but not inside of the guest.The API issue is not that we can't map 0xffff_ffff_ffff_f000, it's that we can't map 0x0 through 0xffff_ffff_ffff_ffff in a single mapping because we pass the size instead of the end address (where size here would be 2^64). We can map 0x0 through 0xffff_ffff_ffff_efff, followed by 0xffff_ffff_ffff_f000 through 0xffff_ffff_ffff_ffff, but again, why would you ever need to do this? Thanks,AlexThe thing is that It could be useful to say we map all the virtual memory.Why? And again, you can do it, just not in a single mapping, which really seems like a theoretical problem since you're mapping the IOVA space of a device, which lives within and consumes a small portion of this address space, so you're pretty much always looking at doing multiple mappings.or invalid mapping. May be the source of the misunderstood.as seen earlier, it depends on the IOMMU architecture, some tablesHaving a size of 2^64 was a possibility. On the other hands, with the actual implementation the "memory_region_iommu_replay" would take on long long long time.In fact, depending on the IOMMU capabilities and usage we do not need tocall the "memory_region_iommu_replay" at that time or even not at all.The mapping itself would take a long time, even with GiB pages we're talking about populating 2^34 page table entries.can be ignored if the entry of the parent table is invalid. And anyway, the host must do it.Another source of misunderstanding can be that I may not understandOf course many of those entries would need to point to the same physical page to cover empty space since otherwise you'd need to run this on a processor that supports more than 2^64 bits of address space, but you probably don't want to waste lots of memory covering empty space, so that means a smaller page size, which means orders of magnitude more mappings and more space wasted in the page tables... All of this should have some useful value, which is escaping me. Thanks,Alexthe QEMU code as good as I think.And may be the last (I hope so) source of misunderstanding is that I didnot explain well the goal of this which is the optimization of IOMMU for some architectures:Depending on the IOMMU architecture you may not be forced to map the memorywith a valid entry from the beginning. In fact you have the following possibility: - map all memory with invalid entry, may be the host do it per default (he better do) - only update entries that the guest updated.To do this you must intercept guest updates.There are several possibilities to do the interception depending on the host architecture, the IOMMU and the IOMMU-TLB architecture.In this case, you would map all the memory the DMA may theoretically accesswith invalid entries, you have no need of replaying the IOMMU entries on start, and you will only updates entries on each DMA_MAP/UNMAP calls.The migration case would need some kind of IOMMU replay but it isanother problem.You've still lost me
sorry, when I do not understand I ask until I understand. and since I did not understand it follows stupid questions. Qemu code is sometime hard to understand for me and there is no room for approximations.
on the actual problem you're trying to solve. You want to invalidate all of the IOTLB entries for a device, but that's exactly the starting state of an IOMMU domain and you already have the ability to get back to that state by unmapping the entire address space, with the caveat that it takes two calls to flush the full 2^64 bit address space.
Thanks, I think that it is the point I did not full understand. I think I understand better, thanks to your patience.
So don't you just need to fix the overflow issue that you've identified and add a workaround for the API issue?
OK, I try to do this.
I don't understand why you'd ever need to _map_, ie. create valid IOTLB entries,
me neither:-) . I misunderstood, I do not need it.
for the entire address space, but if you don't even want to track the highest DMA address in use to make unmapping more efficient, the tools are still there for you. Thanks, Alex
thanks you very much for your patience and explanations, I revisit the way to fix the overflow issue and contact you again when I find something proper. Pierre
