Re: [PATCH v2 1/9] KVM: x86: Add AMD SEV specific Hypercall3

[Ashish Kalra <ashish.kalra@amd.com>]

On Fri, Dec 18, 2020 at 07:56:41PM +0000, Dr. David Alan Gilbert wrote:
> * Kalra, Ashish (Ashish.Kalra@amd.com) wrote:
> > Hello Dave,
> > 
> > On Dec 18, 2020, at 1:40 PM, Dr. David Alan Gilbert <dgilbert@redhat.com> wrote:
> > 
> > * Ashish Kalra (ashish.kalra@amd.com) wrote:
> > On Fri, Dec 11, 2020 at 10:55:42PM +0000, Ashish Kalra wrote:
> > Hello All,
> > 
> > On Tue, Dec 08, 2020 at 10:29:05AM -0600, Brijesh Singh wrote:
> > 
> > On 12/7/20 9:09 PM, Steve Rutherford wrote:
> > On Mon, Dec 7, 2020 at 12:42 PM Sean Christopherson <seanjc@google.com> wrote:
> > On Sun, Dec 06, 2020, Paolo Bonzini wrote:
> > On 03/12/20 01:34, Sean Christopherson wrote:
> > On Tue, Dec 01, 2020, Ashish Kalra wrote:
> > From: Brijesh Singh <brijesh.singh@amd.com>
> > 
> > KVM hypercall framework relies on alternative framework to patch the
> > VMCALL -> VMMCALL on AMD platform. If a hypercall is made before
> > apply_alternative() is called then it defaults to VMCALL. The approach
> > works fine on non SEV guest. A VMCALL would causes #UD, and hypervisor
> > will be able to decode the instruction and do the right things. But
> > when SEV is active, guest memory is encrypted with guest key and
> > hypervisor will not be able to decode the instruction bytes.
> > 
> > Add SEV specific hypercall3, it unconditionally uses VMMCALL. The hypercall
> > will be used by the SEV guest to notify encrypted pages to the hypervisor.
> > What if we invert KVM_HYPERCALL and X86_FEATURE_VMMCALL to default to VMMCALL
> > and opt into VMCALL?  It's a synthetic feature flag either way, and I don't
> > think there are any existing KVM hypercalls that happen before alternatives are
> > patched, i.e. it'll be a nop for sane kernel builds.
> > 
> > I'm also skeptical that a KVM specific hypercall is the right approach for the
> > encryption behavior, but I'll take that up in the patches later in the series.
> > Do you think that it's the guest that should "donate" memory for the bitmap
> > instead?
> > No.  Two things I'd like to explore:
> > 
> >  1. Making the hypercall to announce/request private vs. shared common across
> >     hypervisors (KVM, Hyper-V, VMware, etc...) and technologies (SEV-* and TDX).
> >     I'm concerned that we'll end up with multiple hypercalls that do more or
> >     less the same thing, e.g. KVM+SEV, Hyper-V+SEV, TDX, etc...  Maybe it's a
> >     pipe dream, but I'd like to at least explore options before shoving in KVM-
> >     only hypercalls.
> > 
> > 
> >  2. Tracking shared memory via a list of ranges instead of a using bitmap to
> >     track all of guest memory.  For most use cases, the vast majority of guest
> >     memory will be private, most ranges will be 2mb+, and conversions between
> >     private and shared will be uncommon events, i.e. the overhead to walk and
> >     split/merge list entries is hopefully not a big concern.  I suspect a list
> >     would consume far less memory, hopefully without impacting performance.
> > For a fancier data structure, I'd suggest an interval tree. Linux
> > already has an rbtree-based interval tree implementation, which would
> > likely work, and would probably assuage any performance concerns.
> > 
> > Something like this would not be worth doing unless most of the shared
> > pages were physically contiguous. A sample Ubuntu 20.04 VM on GCP had
> > 60ish discontiguous shared regions. This is by no means a thorough
> > search, but it's suggestive. If this is typical, then the bitmap would
> > be far less efficient than most any interval-based data structure.
> > 
> > You'd have to allow userspace to upper bound the number of intervals
> > (similar to the maximum bitmap size), to prevent host OOMs due to
> > malicious guests. There's something nice about the guest donating
> > memory for this, since that would eliminate the OOM risk.
> > 
> > 
> > Tracking the list of ranges may not be bad idea, especially if we use
> > the some kind of rbtree-based data structure to update the ranges. It
> > will certainly be better than bitmap which grows based on the guest
> > memory size and as you guys see in the practice most of the pages will
> > be guest private. I am not sure if guest donating a memory will cover
> > all the cases, e.g what if we do a memory hotplug (increase the guest
> > ram from 2GB to 64GB), will donated memory range will be enough to store
> > the metadata.
> > 
> > .
> > 
> > With reference to internal discussions regarding the above, i am going
> > to look into specific items as listed below :
> > 
> > 1). "hypercall" related :
> > a). Explore the SEV-SNP page change request structure (included in GHCB),
> > see if there is something common there than can be re-used for SEV/SEV-ES
> > page encryption status hypercalls.
> > b). Explore if there is any common hypercall framework i can use in
> > Linux/KVM.
> > 
> > 2). related to the "backing" data structure - explore using a range-based
> > list or something like rbtree-based interval tree data structure
> > (as mentioned by Steve above) to replace the current bitmap based
> > implementation.
> > 
> > 
> > 
> > I do agree that a range-based list or an interval tree data structure is a
> > really good "logical" fit for the guest page encryption status tracking.
> > 
> > We can only keep track of the guest unencrypted shared pages in the
> > range(s) list (which will keep the data structure quite compact) and all
> > the guest private/encrypted memory does not really need any tracking in
> > the list, anything not in the list will be encrypted/private.
> > 
> > Also looking at a more "practical" use case, here is the current log of
> > page encryption status hypercalls when booting a linux guest :
> > 
> > ...
> > 
> > <snip>
> > 
> > [   56.146336] page_enc_status_hc invoked, gpa = 1f018000, npages  = 1, enc = 1
> > [   56.146351] page_enc_status_hc invoked, gpa = 1f00e000, npages  = 1, enc = 0
> > [   56.147261] page_enc_status_hc invoked, gpa = 1f00e000, npages  = 1, enc = 0
> > [   56.147271] page_enc_status_hc invoked, gpa = 1f018000, npages  = 1, enc = 0
> > ....
> > 
> > [   56.180730] page_enc_status_hc invoked, gpa = 1f008000, npages  = 1, enc = 0
> > [   56.180741] page_enc_status_hc invoked, gpa = 1f006000, npages  = 1, enc = 0
> > [   56.180768] page_enc_status_hc invoked, gpa = 1f008000, npages  = 1, enc = 1
> > [   56.180782] page_enc_status_hc invoked, gpa = 1f006000, npages  = 1, enc = 1
> > 
> > ....
> > [   56.197110] page_enc_status_hc invoked, gpa = 1f007000, npages  = 1, enc = 0
> > [   56.197120] page_enc_status_hc invoked, gpa = 1f005000, npages  = 1, enc = 0
> > [   56.197136] page_enc_status_hc invoked, gpa = 1f007000, npages  = 1, enc = 1
> > [   56.197148] page_enc_status_hc invoked, gpa = 1f005000, npages  = 1, enc = 1
> > ....
> > 
> > [   56.222679] page_enc_status_hc invoked, gpa = 1e83b000, npages  = 1, enc = 0
> > [   56.222691] page_enc_status_hc invoked, gpa = 1e839000, npages  = 1, enc = 0
> > [   56.222707] page_enc_status_hc invoked, gpa = 1e83b000, npages  = 1, enc = 1
> > [   56.222720] page_enc_status_hc invoked, gpa = 1e839000, npages  = 1, enc = 1
> > ....
> > 
> > [   56.313747] page_enc_status_hc invoked, gpa = 1e5eb000, npages  = 1, enc = 0
> > [   56.313771] page_enc_status_hc invoked, gpa = 1e5e9000, npages  = 1, enc = 0
> > [   56.313789] page_enc_status_hc invoked, gpa = 1e5eb000, npages  = 1, enc = 1
> > [   56.313803] page_enc_status_hc invoked, gpa = 1e5e9000, npages  = 1, enc = 1
> > ....
> > [   56.459276] page_enc_status_hc invoked, gpa = 1d767000, npages  = 100, enc = 0
> > [   56.459428] page_enc_status_hc invoked, gpa = 1e501000, npages  = 1, enc = 1
> > [   56.460037] page_enc_status_hc invoked, gpa = 1d767000, npages  = 100, enc = 1
> > [   56.460216] page_enc_status_hc invoked, gpa = 1e501000, npages  = 1, enc = 0
> > [   56.460299] page_enc_status_hc invoked, gpa = 1d767000, npages  = 100, enc = 0
> > [   56.460448] page_enc_status_hc invoked, gpa = 1e501000, npages  = 1, enc = 1
> > ....
> > 
> > As can be observed here, all guest MMIO ranges are initially setup as
> > shared, and those are all contigious guest page ranges.
> > 
> > After that the encryption status hypercalls are invoked when DMA gets
> > triggered during disk i/o while booting the guest ... here again the
> > guest page ranges are contigious, though mostly single page is touched
> > and a lot of page re-use is observed.
> > 
> > So a range-based list/structure will be a "good" fit for such usage
> > scenarios.
> > 
> > It seems surprisingly common to flick the same pages back and forth between
> > encrypted and clear for quite a while;  why is this?
> > 
> > 
> > dma_alloc_coherent()'s will allocate pages and then call
> > set_decrypted() on them and then at dma_free_coherent(), set_encrypted()
> > is called on the pages to be freed. So these observations in the logs
> > where a lot of single 4K pages are seeing C-bit transitions and
> > corresponding hypercalls are the ones associated with
> > dma_alloc_coherent().
> 
> It makes me wonder if it might be worth teaching it to hold onto those
> DMA pages somewhere until it needs them for something else and avoid the
> extra hypercalls; just something to think about.
> 
> Dave

Following up on this discussion and looking at the hypercall logs and DMA usage scenarios on the SEV, I have the following additional observations and comments :

It is mostly the Guest MMIO regions setup as un-encrypted by uefi/edk2 initially, which will be the "static" nodes in the backing data structure for page encryption status. 
These will be like 15-20 nodes/entries.

Drivers doing DMA allocations using GFP_ATOMIC will be fetching DMA buffers from the pre-allocated unencrypted atomic pool, hence it will be a "static" node added at kernel startup.

As we see with the logs, almost all runtime C-bit transitions and corresponding hypercalls will be from DMA I/O and dma_alloc_coherent/dma_free_coherent calls, these will be 
using 4K/single pages and mostly fragmented ranges, so if we use a "rbtree" based interval tree then there will be a lot of tree insertions and deletions 
(dma_alloc_coherent followed with a dma_free_coherent), so this will lead to a lot of expensive tree rotations and re-balancing, compared to much less complex 
and faster linked list node insertions and deletions (if we use a list based structure to represent these interval ranges).

Also as the static nodes in the structure will be quite limited (all the above DMA I/O added ranges will simply be inserted and removed), so a linked list lookup 
won't be too expensive compared to a tree lookup. In other words, this be a fixed size list.

Looking at the above, I am now more inclined to use a list based structure to represent the page encryption status.

Looking fwd. to any comments/feedback/thoughts on the above.

Thanks,
Ashish