On Fri, Dec 02, 2022, Dave Hansen wrote: > On 12/2/22 13:40, Kristen Carlson Accardi wrote: > > On Fri, 2022-12-02 at 13:35 -0800, Dave Hansen wrote: > >> On 12/2/22 10:36, Kristen Carlson Accardi wrote: > >>> When allocating new Version Array (VA) pages, pass the struct > >>> sgx_encl > >>> of the enclave that is allocating the page. sgx_alloc_epc_page() > >>> will > >>> store this value in the encl_owner field of the struct > >>> sgx_epc_page. In > >>> a later patch, VA pages will be placed in an unreclaimable queue, > >>> and then when the cgroup max limit is reached and there are no more > >>> reclaimable pages and the enclave must be oom killed, all the > >>> VA pages associated with that enclave can be uncharged and freed. > >> What does this have to do with the 'encl' that is being passed, > >> though? > >> > >> In other words, why is this new sgx_epc_page-to-encl mapping needed > >> for > >> VA pages now, but it wasn't before? > > When we OOM kill an enclave, we want to get rid of all the associated > > VA pages too. Prior to this patch, there wasn't a way to easily get the > > VA pages associated with an enclave. > > Given an enclave, we have encl->va_pages to look up all the VA pages. > Also, this patch's code allows you to go from a va page to an enclave. Yep. > That seems like it's going the other direction from what an OOM-kill > would need to do. Providing a backpointer from a VA page to its enclave allows OOM-killing the enclave if its cgroup is over the limit but there are no reclaimable pages for said cgroup (for SGX's definition of "reclaimable"). I.e. if all of an enclave's "regular" pages have been swapped out, the only thing left resident in the EPC will be the enclave's VA pages, which are not reclaimable in the kernel's current SGX implementation.
