Re: [Qemu-devel] [PATCH qemu v14 17/18] vfio/spapr: Use VFIO_SPAPR_TCE_v2_IOMMU

[Alexey Kardashevskiy <aik@ozlabs.ru>]

On 03/23/2016 12:08 PM, David Gibson wrote:
> On Tue, Mar 22, 2016 at 04:54:07PM +1100, Alexey Kardashevskiy wrote:
>> On 03/22/2016 04:14 PM, David Gibson wrote:
>>> On Mon, Mar 21, 2016 at 06:47:05PM +1100, Alexey Kardashevskiy wrote:
>>>> New VFIO_SPAPR_TCE_v2_IOMMU type supports dynamic DMA window management.
>>>> This adds ability to VFIO common code to dynamically allocate/remove
>>>> DMA windows in the host kernel when new VFIO container is added/removed.
>>>>
>>>> This adds VFIO_IOMMU_SPAPR_TCE_CREATE ioctl to vfio_listener_region_add
>>>> and adds just created IOMMU into the host IOMMU list; the opposite
>>>> action is taken in vfio_listener_region_del.
>>>>
>>>> When creating a new window, this uses euristic to decide on the TCE table
>>>> levels number.
>>>>
>>>> This should cause no guest visible change in behavior.
>>>>
>>>> Signed-off-by: Alexey Kardashevskiy <aik@ozlabs.ru>
>>>> ---
>>>> Changes:
>>>> v14:
>>>> * new to the series
>>>>
>>>> ---
>>>> TODO:
>>>> * export levels to PHB
>>>> ---
>>>>   hw/vfio/common.c | 108 ++++++++++++++++++++++++++++++++++++++++++++++++++++---
>>>>   trace-events     |   2 ++
>>>>   2 files changed, 105 insertions(+), 5 deletions(-)
>>>>
>>>> diff --git a/hw/vfio/common.c b/hw/vfio/common.c
>>>> index 4e873b7..421d6eb 100644
>>>> --- a/hw/vfio/common.c
>>>> +++ b/hw/vfio/common.c
>>>> @@ -279,6 +279,14 @@ static int vfio_host_iommu_add(VFIOContainer *container,
>>>>       return 0;
>>>>   }
>>>>
>>>> +static void vfio_host_iommu_del(VFIOContainer *container, hwaddr min_iova)
>>>> +{
>>>> +    VFIOHostIOMMU *hiommu = vfio_host_iommu_lookup(container, min_iova, 0x1000);
>>>
>>> The hard-coded 0x1000 looks dubious..
>>
>> Well, that's the minimal page size...
>
> Really?  Some BookE CPUs support 1KiB page size..


Hm. For IOMMU? Ok. s/0x1000/1/ should do then :)


>
>>>> +    g_assert(hiommu);
>>>> +    QLIST_REMOVE(hiommu, hiommu_next);
>>>> +}
>>>> +
>>>>   static bool vfio_listener_skipped_section(MemoryRegionSection *section)
>>>>   {
>>>>       return (!memory_region_is_ram(section->mr) &&
>>>> @@ -392,6 +400,61 @@ static void vfio_listener_region_add(MemoryListener *listener,
>>>>       }
>>>>       end = int128_get64(llend);
>>>>
>>>> +    if (container->iommu_type == VFIO_SPAPR_TCE_v2_IOMMU) {
>>>
>>> I think this would be clearer split out into a helper function,
>>> vfio_create_host_window() or something.
>>
>>
>> It is rather vfio_spapr_create_host_window() and we were avoiding
>> xxx_spapr_xxx so far. I'd cut-n-paste the SPAPR PCI AS listener to a
>> separate file but this usually triggers more discussion and never ends well.
>>
>>
>>
>>>> +        unsigned entries, pages;
>>>> +        struct vfio_iommu_spapr_tce_create create = { .argsz = sizeof(create) };
>>>> +
>>>> +        g_assert(section->mr->iommu_ops);
>>>> +        g_assert(memory_region_is_iommu(section->mr));
>>>
>>> I don't think you need these asserts.  AFAICT the same logic should
>>> work if a RAM MR was added directly to PCI address space - this would
>>> create the new host window, then the existing code for adding a RAM MR
>>> would map that block of RAM statically into the new window.
>>
>> In what configuration/machine can we do that on SPAPR?
>
> spapr guests won't ever do that.  But you can run an x86 guest on a
> powernv host and this situation could come up.


I am pretty sure VFIO won't work in this case anyway.

> In any case there's no point asserting if the code is correct anyway.

Assert here says (at least) "not tested" or "not expected to happen".


>
>>>> +        trace_vfio_listener_region_add_iommu(iova, end - 1);
>>>> +        /*
>>>> +         * FIXME: For VFIO iommu types which have KVM acceleration to
>>>> +         * avoid bouncing all map/unmaps through qemu this way, this
>>>> +         * would be the right place to wire that up (tell the KVM
>>>> +         * device emulation the VFIO iommu handles to use).
>>>> +         */
>>>> +        create.window_size = memory_region_size(section->mr);
>>>> +        create.page_shift =
>>>> +                ctz64(section->mr->iommu_ops->get_page_sizes(section->mr));
>>>
>>> Ah.. except that I guess you'd need to fall back to host page size
>>> here to handle a RAM MR.
>>
>> Can you give an example of such RAM MR being added to PCI AS on
>> SPAPR?
>
> On spapr, no.  But you can run other machine types as guests (at least
> with TCG) on a host with the spapr IOMMU.
>
>>>> +        /*
>>>> +         * SPAPR host supports multilevel TCE tables, there is some
>>>> +         * euristic to decide how many levels we want for our table:
>>>> +         * 0..64 = 1; 65..4096 = 2; 4097..262144 = 3; 262145.. = 4
>>>> +         */
>>>> +        entries = create.window_size >> create.page_shift;
>>>> +        pages = (entries * sizeof(uint64_t)) / getpagesize();
>>>> +        create.levels = ctz64(pow2ceil(pages) - 1) / 6 + 1;
>>>> +
>>>> +        ret = ioctl(container->fd, VFIO_IOMMU_SPAPR_TCE_CREATE, &create);
>>>> +        if (ret) {
>>>> +            error_report("Failed to create a window, ret = %d (%m)", ret);
>>>> +            goto fail;
>>>> +        }
>>>> +
>>>> +        if (create.start_addr != section->offset_within_address_space ||
>>>> +            vfio_host_iommu_lookup(container, create.start_addr,
>>>> +                                   create.start_addr + create.window_size - 1)) {
>>>
>>> Under what circumstances can this trigger?  Is the kernel ioctl
>>> allowed to return a different window start address than the one
>>> requested?
>>
>> You already asked this some time ago :) The userspace cannot request
>> address, the host kernel returns one.
>
> Ok.  For generality it would be nice if you could succeed here as long
> as the new host window covers the requested guest window, even if it
> doesn't match exactly.  And for that matter to not request the new
> window if the host already has a window covering the guest region.


That would be dead code - when would it possibly work? I mean I could 
instrument an artificial test but the actual user which might appear later 
will likely be soooo different so this won't help anyway.


>>> The second check looks very strange - if it returns true doesn't that
>>> mean you *do* have host window which can accomodate this guest region,
>>> which is what you want?
>>
>> This should not happen, this is what this check is for. Can make it assert()
>> or something like this.
>
> Oh.. I see.  Because you've done the ioctl, but not recorded the new
> host window in the list yet.
>
> No, I think the correct approach is to look for an existing host
> window containing the requested guest window *before* you try to
> create a new host window.  If one is already there, you can just carry
> on.

Right, I'll change this.


>
>>>> +            struct vfio_iommu_spapr_tce_remove remove = {
>>>> +                .argsz = sizeof(remove),
>>>> +                .start_addr = create.start_addr
>>>> +            };
>>>> +            error_report("Host doesn't support DMA window at %"HWADDR_PRIx", must be %"PRIx64,
>>>> +                         section->offset_within_address_space,
>>>> +                         create.start_addr);
>>>> +            ioctl(container->fd, VFIO_IOMMU_SPAPR_TCE_REMOVE, &remove);
>>>> +            ret = -EINVAL;
>>>> +            goto fail;
>>>> +        }
>>>> +        trace_vfio_spapr_create_window(create.page_shift,
>>>> +                                       create.window_size,
>>>> +                                       create.start_addr);
>>>> +
>>>> +        vfio_host_iommu_add(container, create.start_addr,
>>>> +                            create.start_addr + create.window_size - 1,
>>>> +                            1ULL << create.page_shift);
>>>> +    }
>>>> +
>>>>       if (!vfio_host_iommu_lookup(container, iova, end - 1)) {
>>>>           error_report("vfio: IOMMU container %p can't map guest IOVA region"
>>>>                        " 0x%"HWADDR_PRIx"..0x%"HWADDR_PRIx,
>>>> @@ -525,6 +588,22 @@ static void vfio_listener_region_del(MemoryListener *listener,
>>>>                        container, iova, end - iova, ret);
>>>>       }
>>>>
>>>> +    if (container->iommu_type == VFIO_SPAPR_TCE_v2_IOMMU) {
>>>> +        struct vfio_iommu_spapr_tce_remove remove = {
>>>> +            .argsz = sizeof(remove),
>>>> +            .start_addr = section->offset_within_address_space,
>>>> +        };
>>>> +        ret = ioctl(container->fd, VFIO_IOMMU_SPAPR_TCE_REMOVE, &remove);
>>>> +        if (ret) {
>>>> +            error_report("Failed to remove window at %"PRIx64,
>>>> +                         remove.start_addr);
>>>> +        }
>>>> +
>>>> +        vfio_host_iommu_del(container, section->offset_within_address_space);
>>>> +
>>>> +        trace_vfio_spapr_remove_window(remove.start_addr);
>>>> +    }
>>>> +
>>>>       if (iommu && iommu->iommu_ops && iommu->iommu_ops->vfio_stop) {
>>>>           iommu->iommu_ops->vfio_stop(section->mr);
>>>>       }
>>>> @@ -928,11 +1007,30 @@ static int vfio_connect_container(VFIOGroup *group, AddressSpace *as)
>>>>               goto listener_release_exit;
>>>>           }
>>>>
>>>> -        /* The default table uses 4K pages */
>>>> -        vfio_host_iommu_add(container, info.dma32_window_start,
>>>> -                            info.dma32_window_start +
>>>> -                            info.dma32_window_size - 1,
>>>> -                            0x1000);
>>>> +        if (v2) {
>>>> +            /*
>>>> +             * There is a default window in just created container.
>>>> +             * To make region_add/del simpler, we better remove this
>>>> +             * window now and let those iommu_listener callbacks
>>>> +             * create/remove them when needed.
>>>> +             */
>>>> +            struct vfio_iommu_spapr_tce_remove remove = {
>>>> +                .argsz = sizeof(remove),
>>>> +                .start_addr = info.dma32_window_start,
>>>> +            };
>>>> +            ret = ioctl(fd, VFIO_IOMMU_SPAPR_TCE_REMOVE, &remove);
>>>> +            if (ret) {
>>>> +                error_report("vfio: VFIO_IOMMU_SPAPR_TCE_REMOVE failed: %m");
>>>> +                ret = -errno;
>>>> +                goto free_container_exit;
>>>> +            }
>>>> +        } else {
>>>> +            /* The default table uses 4K pages */
>>>> +            vfio_host_iommu_add(container, info.dma32_window_start,
>>>> +                                info.dma32_window_start +
>>>> +                                info.dma32_window_size - 1,
>>>> +                                0x1000);
>>>> +        }
>>>>       } else {
>>>>           error_report("vfio: No available IOMMU models");
>>>>           ret = -EINVAL;
>>>> diff --git a/trace-events b/trace-events
>>>> index cc619e1..f2b75a3 100644
>>>> --- a/trace-events
>>>> +++ b/trace-events
>>>> @@ -1736,6 +1736,8 @@ vfio_region_finalize(const char *name, int index) "Device %s, region %d"
>>>>   vfio_region_mmaps_set_enabled(const char *name, bool enabled) "Region %s mmaps enabled: %d"
>>>>   vfio_ram_register(uint64_t va, uint64_t size, int ret) "va=%"PRIx64" size=%"PRIx64" ret=%d"
>>>>   vfio_ram_unregister(uint64_t va, uint64_t size, int ret) "va=%"PRIx64" size=%"PRIx64" ret=%d"
>>>> +vfio_spapr_create_window(int ps, uint64_t ws, uint64_t off) "pageshift=0x%x winsize=0x%"PRIx64" offset=0x%"PRIx64
>>>> +vfio_spapr_remove_window(uint64_t off) "offset=%"PRIx64
>>>>
>>>>   # hw/vfio/platform.c
>>>>   vfio_platform_base_device_init(char *name, int groupid) "%s belongs to group #%d"
>>>
>>
>>
>


-- 
Alexey