From: Mark Kettenis Subject: Re: SEV-SNP: bus_dma(9): Rescind and claim DMA bounce buffer To: hshoexer Cc: tech@openbsd.org Date: Sat, 26 Sep 2026 21:17:51 +0200 > Date: Thu, 24 Sep 2026 20:43:15 +0200 > From: hshoexer > > On Thu, Sep 24, 2026 at 07:33:39PM +0200, Mark Kettenis wrote: > > > Date: Thu, 24 Sep 2026 18:14:41 +0200 > > > From: hshoexer > > > > > > Hi, > > > > > > when creating DMA maps, rescind DMA bounce buffer pages to the > > > hypervisor. When destroying DMA maps, claim these pages back from > > > the hypervisor. > > > > > > ok? > > > > Sorry, too much jargon I'm not familliar with. Can you explain what > > does actually does? > > sure: > > With SEV and SEV-ES: For pages to be shared with the hypervisor > the guest maps them without the crypt bit (PMAP_NOCRYPT). That's > what we do when setting up pages to be used as bounce buffers. > However, the hypervisor can still access "private" pages, e.g. pages > that are mapped encrypted by setting the crypt bit in the PTE. This > corrupts the cipher text, pages are not integrity protected. SEV > and SEV-ES only provide "privacy". > > SEV-SNP provides integrity protection. All physcal pages are managed > with the so called Reverse Mapping Table (RMP). The MMU not only > uses the information from the page table but also from the RMP. > > Pages have to be either assigned to the Hypervisor/Host OS or the > guest (VM). Note: There are more possible page "owners", I'm > simplifying a bit. > > When the hypervisor touches a guest owned page, #GP is raised. When > a page is owned by the hypervisor, the guest is allowed to touch > this page as long as it is mapped unencrypted. This is what we > want to do for bounce buffering. > > When the hypervisor forcefully withdraws a page from the guest, the > page is "invalidated". When the guest touches such a page, a #VC > trap is raised in the guest. This allows the guest to detect this > "corruption". That's the integrity protection, SEV-SNP provides. > > On guest launch the guest firmware -- SEV-SNP enable OVMF for > qemu/KVM -- "claims" all pages assigned to the VM with these steps: > > 1) Remove the mapping for unencrypted access, i.e. map it encrypted > 2) Request the hypervisor to set "assigned" to "guest owned", using > the GHCB protocol function "page stage change" > 3) "validate" the page using the pvalidate instruction. > > Afte that these pages are marked in the RMP to be owned by the > guest, the hypervisor can not access these pages anymore without > raising #GP. > > To share pages with the hypervisor the -- for bounce buffering -- > the guest "rescinds" pages back to the hypervisor with these three > steps: > > 1) Set "validate" in the RMP to 0 using pvalidate instruction. > 2) Request the hypervisor to set "assigned" in the RMP to "hypervisor > owned", using the GHCB protocol "page state change" > 3) Establish a mapping with the C-bit cleared; with this mapping > the page can be accessed unencrypted. > > So, when setting up bounce buffers on dmamap creation, the guest > "rescinds" pages. When destroying the dmamap the guest "claims" > the pages again. Claim and resincd is implemented in diff #1. > > Naming is hard. Thanks. I think sun4v used "grant" and "revoke" in the somewhat similar case of sharing pages with other domains. I think there is a problem with the diff though in that it breaks builds without amdsev enabled... > > > Take care, > > > HJ. > > > --- > > > sys/arch/amd64/amd64/bus_dma.c | 5 +++++ > > > 1 file changed, 5 insertions(+) > > > > > > diff --git a/sys/arch/amd64/amd64/bus_dma.c b/sys/arch/amd64/amd64/bus_dma.c > > > index f4efff95a18..2e69319739c 100644 > > > --- a/sys/arch/amd64/amd64/bus_dma.c > > > +++ b/sys/arch/amd64/amd64/bus_dma.c > > > @@ -100,6 +100,7 @@ > > > #endif > > > > > > #include > > > +#include > > > > > > #include > > > > > > @@ -255,6 +256,8 @@ _bus_dmamap_create(bus_dma_tag_t t, bus_size_t size, int nsegments, > > > return (ENOMEM); > > > } > > > > > > + snp_rescind_pages(&mlist); > > > + > > > sva = va; > > > ssize = sz; > > > pgnext = TAILQ_FIRST(&mlist); > > > @@ -269,6 +272,7 @@ _bus_dmamap_create(bus_dma_tag_t t, bus_size_t size, int nsegments, > > > map->_dm_npages = 0; > > > km_free((void *)sva, ssize, &kv_any, &kp_none); > > > free(map, M_DEVBUF, mapsize); > > > + snp_claim_pages(&mlist); > > > uvm_pglistfree(&mlist); > > > return (ENOMEM); > > > } > > > @@ -307,6 +311,7 @@ _bus_dmamap_destroy(bus_dma_tag_t t, bus_dmamap_t map) > > > for (pg = map->_dm_pages; map->_dm_npages--; pg++) { > > > TAILQ_INSERT_TAIL(&mlist, *pg, pageq); > > > } > > > + snp_claim_pages(&mlist); > > > uvm_pglistfree(&mlist); > > > } > > > > > > -- > > > 2.47.3 > > > > > > > >