On Thu, May 19, 2022 at 2:05 PM Kristen Carlson Accardi
<[email protected]> wrote:
>
> When the system runs out of enclave memory, SGX can reclaim EPC pages
> by swapping to normal RAM. These backing pages are allocated via a
> per-enclave shared memory area. Since SGX allows unlimited over
> commit on EPC memory, the reclaimer thread can allocate a large
> number of backing RAM pages in response to EPC memory pressure.
>
> When the shared memory backing RAM allocation occurs during
> the reclaimer thread context, the shared memory is charged to
> the root memory control group, and the shmem usage of the enclave
> is not properly accounted for, making cgroups ineffective at
> limiting the amount of RAM an enclave can consume.
>
> For example, when using a cgroup to launch a set of test
> enclaves, the kernel does not properly account for 50% - 75% of
> shmem page allocations on average. In the worst case, when
> nearly all allocations occur during the reclaimer thread, the
> kernel accounts less than a percent of the amount of shmem used
> by the enclave's cgroup to the correct cgroup.
>
> SGX stores a list of mm_structs that are associated with
> an enclave. Pick one of them during reclaim and charge that
> mm's memcg with the shmem allocation. The one that gets picked
> is arbitrary, but this list almost always only has one mm. The
> cases where there is more than one mm with different memcg's
> are not worth considering.
>
> Create a new function - sgx_encl_alloc_backing(). This function
> is used whenever a new backing storage page needs to be
> allocated. Previously the same function was used for page
> allocation as well as retrieving a previously allocated page.
> Prior to backing page allocation, if there is a mm_struct associated
> with the enclave that is requesting the allocation, it is set
> as the active memory control group.
>
> Signed-off-by: Kristen Carlson Accardi <[email protected]>
For the memcg part:
Reviewed-by: Shakeel Butt <[email protected]>
On Thu, 2022-05-19 at 14:22 -0700, Shakeel Butt wrote:
> On Thu, May 19, 2022 at 2:05 PM Kristen Carlson Accardi
> <[email protected]> wrote:
> > When the system runs out of enclave memory, SGX can reclaim EPC
> > pages
> > by swapping to normal RAM. These backing pages are allocated via a
> > per-enclave shared memory area. Since SGX allows unlimited over
> > commit on EPC memory, the reclaimer thread can allocate a large
> > number of backing RAM pages in response to EPC memory pressure.
> >
> > When the shared memory backing RAM allocation occurs during
> > the reclaimer thread context, the shared memory is charged to
> > the root memory control group, and the shmem usage of the enclave
> > is not properly accounted for, making cgroups ineffective at
> > limiting the amount of RAM an enclave can consume.
> >
> > For example, when using a cgroup to launch a set of test
> > enclaves, the kernel does not properly account for 50% - 75% of
> > shmem page allocations on average. In the worst case, when
> > nearly all allocations occur during the reclaimer thread, the
> > kernel accounts less than a percent of the amount of shmem used
> > by the enclave's cgroup to the correct cgroup.
> >
> > SGX stores a list of mm_structs that are associated with
> > an enclave. Pick one of them during reclaim and charge that
> > mm's memcg with the shmem allocation. The one that gets picked
> > is arbitrary, but this list almost always only has one mm. The
> > cases where there is more than one mm with different memcg's
> > are not worth considering.
> >
> > Create a new function - sgx_encl_alloc_backing(). This function
> > is used whenever a new backing storage page needs to be
> > allocated. Previously the same function was used for page
> > allocation as well as retrieving a previously allocated page.
> > Prior to backing page allocation, if there is a mm_struct
> > associated
> > with the enclave that is requesting the allocation, it is set
> > as the active memory control group.
> >
> > Signed-off-by: Kristen Carlson Accardi <[email protected]>
>
> For the memcg part:
>
> Reviewed-by: Shakeel Butt <[email protected]>
Thanks!