On Thu, Jan 7, 2021 at 4:48 PM Ashish Kalra <ashish.kalra@xxxxxxx> wrote: > > > On Thu, Jan 07, 2021 at 01:34:14AM +0000, Ashish Kalra wrote: > > > Hello Steve, > > > > > > My thoughts here ... > > > > > > On Wed, Jan 06, 2021 at 05:01:33PM -0800, Steve Rutherford wrote: > > > > Avoiding an rbtree for such a small (but unstable) list seems correct. > > > > > > > > > > I agree. > > > > > > > For the unencrypted region list strategy, the only questions that I > > > > have are fairly secondary. > > > > - How should the kernel upper bound the size of the list in the face > > > > of malicious guests, but still support large guests? (Something > > > > similar to the size provided in the bitmap API would work). > > > > - What serialization format should be used for the ioctl API? > > > > (Usermode could send down a pointer to a user region and a size. The > > > > kernel could then populate that with an array of structs containing > > > > bases and limits for unencrypted regions.) > > > > - How will the kernel tag a guest as having exceeded its maximum list > > > > size, in order to indicate that the list is now incomplete? (Track a > > > > poison bit, and send it up when getting the serialized list of > > > > regions). > > > > > > > > In my view, there are two main competitors to this strategy: > > > > - (Existing) Bitmap API > > > > - A guest memory donation based model > > > > > > > > The existing bitmap API avoids any issues with growing too large, > > > > since it's size is predictable. > > > > > > > > To elaborate on the memory donation based model, the guest could put > > > > an encryption status data structure into unencrypted guest memory, and > > > > then use a hypercall to inform the host where the base of that > > > > structure is located. The main advantage of this is that it side steps > > > > any issues around malicious guests causing large allocations. > > > > > > > > The unencrypted region list seems very practical. It's biggest > > > > advantage over the bitmap is how cheap it will be to pass the > > > > structure up from the kernel. A memory donation based model could > > > > achieve similar performance, but with some additional complexity. > > > > > > > > Does anyone view the memory donation model as worth the complexity? > > > > Does anyone think the simplicity of the bitmap is a better tradeoff > > > > compared to an unencrypted region list? > > > > > > One advantage in sticking with the bitmap is that it maps very nicely to > > > the dirty bitmap page tracking logic in KVM/Qemu. The way Brijesh > > > designed and implemented it is very similar to dirty page bitmap tracking > > > and syncing between KVM and Qemu. The same logic is re-used for the page > > > encryption bitmap which means quite mininal changes and code resuse in > > > Qemu. > > > > > > Any changes to the backing data structure, will require additional > > > mapping logic to be added to Qemu. > > > > > > This is one advantage in keeping the bitmap logic. > > > > > So if nobody is in favor of keeping the (current) bitmap logic, we will > move to the unencrypted region list approach. Sounds good to me. > > Thanks, > Ashish
