Jason Gunthorpe jgg@ziepe.ca writes:
On Wed, Sep 23, 2026 at 11:40:36AM +0100, Catalin Marinas wrote:
The simplest is probably to always zero in the backend and ignore __GFP_ZERO to the allocator. But it's probably only marginally smaller than passing a CC_SHARED_ZERO flag down. Get codex to try this as well and compare the diffstat.
For patch ordering I would convert to use the allocator first
The semantics of the new API should be clear
If you pass GFP_ZERO then the resulting allocated memory is zero
Otherwise the allocator does Whatever The Arch Needs to not leak private data out.
ok
Once places are converted to the allocator lets go see what is left and ask why it is left and what API it actually needs.
I'm also considering requiring the address passed to cc_make_shared() to be in the linear map. This is currently required by both TDX and CCA, while AMD SNP appears to support vmalloc addresses. The only user of that vmalloc support is Hyper-V VMBus GPADL setup (vmbus_establish_gpadl()). How should the generic CoCo shared-memory allocator handle this?
-aneesh