|
From: | Alexander Graf |
Subject: | Re: [Qemu-devel] [RFC PATCH 09/10] spapr_pci_vfio: Enable DDW |
Date: | Fri, 15 Aug 2014 09:37:27 +0200 |
User-agent: | Mozilla/5.0 (Macintosh; Intel Mac OS X 10.9; rv:31.0) Gecko/20100101 Thunderbird/31.0 |
On 15.08.14 05:16, Alexey Kardashevskiy wrote:
On 08/14/2014 11:38 PM, Alexander Graf wrote:On 13.08.14 02:18, Alexey Kardashevskiy wrote:On 08/13/2014 01:28 AM, Alexander Graf wrote:On 12.08.14 17:10, Alexey Kardashevskiy wrote:On 08/12/2014 07:37 PM, Alexander Graf wrote:On 12.08.14 02:03, Alexey Kardashevskiy wrote:On 08/12/2014 03:30 AM, Alexander Graf wrote:On 11.08.14 17:01, Alexey Kardashevskiy wrote:On 08/11/2014 10:02 PM, Alexander Graf wrote:On 31.07.14 11:34, Alexey Kardashevskiy wrote:This implements DDW for VFIO. Host kernel support is required for this. Signed-off-by: Alexey Kardashevskiy <address@hidden> --- hw/ppc/spapr_pci_vfio.c | 75 +++++++++++++++++++++++++++++++++++++++++++++++++ 1 file changed, 75 insertions(+) diff --git a/hw/ppc/spapr_pci_vfio.c b/hw/ppc/spapr_pci_vfio.c index d3bddf2..dc443e2 100644 --- a/hw/ppc/spapr_pci_vfio.c +++ b/hw/ppc/spapr_pci_vfio.c @@ -69,6 +69,77 @@ static void spapr_phb_vfio_finish_realize(sPAPRPHBState *sphb, Error **errp) /* Register default 32bit DMA window */ memory_region_add_subregion(&sphb->iommu_root, tcet->bus_offset, spapr_tce_get_iommu(tcet)); + + sphb->ddw_supported = !!(info.flags & VFIO_IOMMU_SPAPR_TCE_FLAG_DDW); +} + +static int spapr_pci_vfio_ddw_query(sPAPRPHBState *sphb, + uint32_t *windows_available, + uint32_t *page_size_mask) +{ + sPAPRPHBVFIOState *svphb = SPAPR_PCI_VFIO_HOST_BRIDGE(sphb); + struct vfio_iommu_spapr_tce_query query = { .argsz = sizeof(query) }; + int ret; + + ret = vfio_container_ioctl(&sphb->iommu_as, svphb->iommugroupid, + VFIO_IOMMU_SPAPR_TCE_QUERY, &query); + if (ret) { + return ret; + } + + *windows_available = query.windows_available; + *page_size_mask = query.page_size_mask; + + return ret; +} + +static int spapr_pci_vfio_ddw_create(sPAPRPHBState *sphb, uint32_t page_shift, + uint32_t window_shift, uint32_t liobn, + sPAPRTCETable **ptcet) +{ + sPAPRPHBVFIOState *svphb = SPAPR_PCI_VFIO_HOST_BRIDGE(sphb); + struct vfio_iommu_spapr_tce_create create = { + .argsz = sizeof(create), + .page_shift = page_shift, + .window_shift = window_shift, + .start_addr = 0 + }; + int ret; + + ret = vfio_container_ioctl(&sphb->iommu_as, svphb->iommugroupid, + VFIO_IOMMU_SPAPR_TCE_CREATE, &create); + if (ret) { + return ret; + } + + *ptcet = spapr_tce_new_table(DEVICE(sphb), liobn, create.start_addr, + page_shift, 1 << (window_shift - page_shift),I spot a 1 without ULL again - this time it might work out ok, but please just always use ULL when you pass around addresses.My bad. I keep forgetting this, I'll adjust my own checkpatch.py :)Please walk me though the abstraction levels on what each page size honoration means. If I use THP, what page size granularity can I use for TCE entries?[RFC PATCH 06/10] spapr_rtas: Add Dynamic DMA windows (DDW) RTAS calls support + const struct { int shift; uint32_t mask; } masks[] = { + { 12, DDW_PGSIZE_4K }, + { 16, DDW_PGSIZE_64K }, + { 24, DDW_PGSIZE_16M }, + { 25, DDW_PGSIZE_32M }, + { 26, DDW_PGSIZE_64M }, + { 27, DDW_PGSIZE_128M }, + { 28, DDW_PGSIZE_256M }, + { 34, DDW_PGSIZE_16G }, + }; Supported page sizes are returned by the host kernel via "query". For 16MB pages, page shift will return DDW_PGSIZE_4K|DDW_PGSIZE_64K|DDW_PGSIZE_16M. Or I did not understand the question...Why do we care about the sizes? Anything bigger than what we support should always work, no? What happens if the guest creates a 16MB map but my pages are 4kb mapped? Wouldn't the same logic be able to deal with 16G pages?It is DMA memory, if I split "virtual" 16M page to a bunch of real 4K pages, I have to make sure these 16M are continuous - there will be one TCE entry for it and no more translations besides IOMMU. What do I miss now?Who does the shadow translation where? Does it exist at all?IOMMU? I am not sure I am following you... This IOMMU will look as direct DMA for the guest but the real IOMMU table is sparse and it is populated via a bunch of H_PUT_TCE calls as the default small window. There is a direct mapping in the host called "bypass window" but it is not used here as sPAPR does not define that for paravirtualization.Ok, imagine I have 16MB of guest physical memory that is in reality backed by 256 64k pages on the host. The guest wants to create a 16M TCE entry for this (from its point of view contiguous) chunk of memory. Do we allow this?No, we do not. We tell the guest what it can use.Or do we force the guest to create 64k TCE entries?16MB TCE pages are only allowed if qemu is running with hugepages.That's unfortunate ;)This is my limitation, not SPAPR spec or anything like that.but as long as we have to pin TCEd memory anyway, I guess it doesn't hurt as badly.Yep.If we allow it, why would we ever put any restriction at the upper end of TCE entry sizes? If we already implement enough logic to map things lazily around, we could as well have the guest create a 256M TCE entry and just split it on the host view to 64k TCE entries.Oh, thiiiiiis is what you meant... Well, we could, just for now current linux guests support 4K/64K/16M only and they choose depending on what hypervisor supports - look at enable_ddw() in the guest. What you suggest seems to be an unnecessary code duplication for 16MB pages case. For bigger page sizes - for example, for 64GB guest, a TCE table with 16MB TCEs will be 32KB which is already awesome enough, no?In "normal" invironments guests won't be backed by 16M pages, but by 64k pages with the occasional THP huge page merge that you can't rely on. That's why I figured it'd be smart to support 16MB TCEs even when the underlying memory is only backed by 64k pages.Real TCE table will be using 64K pages anyway - whether we fake 16MB pages to the guest and split them silently in QEMU/host OR guest itself uses 64K pages, in any case we end up with a bigger 64K TCE table so I do not see what we win if we fake 16MB pages for the guest.
Well, what if you for example want to live migrate from a hugepage backed guest to a non-hugepage backed guest? In that case you have to migrate the TCEs as well, but the new location can't use the 16MB ones, right?
We also win in that we have fewer hypercalls to add TCEs, but I suppose that part is almost negligible.
Alex
My point it the code is already in the guest, it can do 64K and 16MB pages, if it was not there, yes, everything you suggest would make sense to implement :)
[Prev in Thread] | Current Thread | [Next in Thread] |