On Fri, 2024-04-19 at 13:55 -0500, Haitao Huang wrote:
On Thu, 18 Apr 2024 20:32:14 -0500, Huang, Kai <kai.huang@xxxxxxxxx>
wrote:
>
>
> On 16/04/2024 3:20 pm, Haitao Huang wrote:
> > From: Kristen Carlson Accardi <kristen@xxxxxxxxxxxxxxx>
> > In cases EPC pages need be allocated during a page fault and the
cgroup
> > usage is near its limit, an asynchronous reclamation needs be
triggered
> > to avoid blocking the page fault handling.
> > Create a workqueue, corresponding work item and function
definitions
> > for EPC cgroup to support the asynchronous reclamation.
> > In case the workqueue allocation is failed during init, disable
cgroup.
>
> It's fine and reasonable to disable (SGX EPC) cgroup. The problem is
> "exactly what does this mean" isn't quite clear.
>
First, this is really some corner case most people don't care: during
init, kernel can't even allocate a workqueue object. So I don't think we
should write extra code to implement some sophisticated solution. Any
solution we come up with may just not work as the way user want or solve
the real issue due to the fact such allocation failure even happens at
init time.
I think for such boot time failure we can either choose directly
BUG_ON(),
or we try to handle it _nicely_, but not half-way. My experience is
adding BUG_ON() should be avoided in general, but it might be acceptable
during kernel boot. I will leave it to others.
[...]
>
> ..., IIUC you choose a (third) solution that is even one more step
back:
>
> It just makes try_charge() always succeed, but EPC pages are still
> managed in the "per-cgroup" list.
>
> But this solution, AFAICT, doesn't work. The reason is when you fail
to
> allocate EPC page you will do the global reclaim, but now the global
> list is empty.
>
> Am I missing anything?
But when cgroups enabled in config, global reclamation starts from root
and reclaim from the whole hierarchy if user may still be able to
create.
Just that we don't have async/sync per-cgroup reclaim triggered.
OK. I missed this as it is in a later patch.
>
> So my thinking is, we have two options:
>
> 1) Modify the MISC cgroup core code to allow the kernel to disable one
> particular resource. It shouldn't be hard, e.g., we can add a
> 'disabled' flag to the 'struct misc_res'.
>
> Hmm.. wait, after checking, the MISC cgroup won't show any control
files
> if the "capacity" of the resource is 0:
>
> "
> * Miscellaneous resources capacity for the entire machine. 0
capacity
> * means resource is not initialized or not present in the host.
> "
>
> So I really suppose we should go with this route, i.e., by just
setting
> the EPC capacity to 0?
>
> Note misc_cg_try_charge() will fail if capacity is 0, but we can make
it
> return success by explicitly check whether SGX cgroup is disabled by
> using a helper, e.g., sgx_cgroup_disabled().
>
> And you always return the root SGX cgroup in sgx_get_current_cg() when
> sgx_cgroup_disabled() is true.
>
> And in sgx_reclaim_pages_global(), you do something like:
>
> static void sgx_reclaim_pages_global(..)
> {
> #ifdef CONFIG_CGROUP_MISC
> if (sgx_cgroup_disabled())
> sgx_reclaim_pages(&sgx_root_cg.lru);
> else
> sgx_cgroup_reclaim_pages(misc_cg_root());
> #else
> sgx_reclaim_pages(&sgx_global_list);
> #endif
> }
>
> I am perhaps missing some other spots too but you got the idea.
>
> At last, after typing those, I believe we should have a separate patch
> to handle disable SGX cgroup at initialization time. And you can even
> put this patch _somewhere_ after the patch
>
> "x86/sgx: Implement basic EPC misc cgroup functionality"
>
> and before this patch.
>
> It makes sense to have such patch anyway, because with it we can
easily
> to add a kernel command line 'sgx_cgroup=disabled" if the user wants
it
> disabled (when someone has such requirement in the future).
>
I think we can add support for "sgx_cgroup=disabled" in future if indeed
needed. But just for init failure, no?
It's not about the commandline, which we can add in the future when
needed. It's about we need to have a way to handle SGX cgroup being
disabled at boot time nicely, because we already have a case where we
need
to do so.
Your approach looks half-way to me, and is not future extendible. If we
choose to do it, do it right -- that is, we need a way to disable it
completely in both kernel and userspace so that userspace won't be able
to
see it.
