On Tue, 2024-08-20 at 18:54 -0700, Haitao Huang wrote:
> With EPC cgroups, the global reclamation function,
> sgx_reclaim_pages_global(), can no longer apply to the global LRU as
> pages are now in per-cgroup LRUs.
>
> Create a wrapper, sgx_cgroup_reclaim_global() to invoke
> sgx_cgroup_reclaim_pages() passing in the root cgroup.
>
[...]
> Call this wrapper
> from sgx_reclaim_pages_global() when cgroup is enabled.
>
This is not done in this patch.
> The wrapper will
> scan and attempt to reclaim SGX_NR_TO_SCAN pages just like the current
> global reclaim.
>
> Note this simple implementation doesn't _exactly_ mimic the current
> global EPC reclaim (which always tries to do the actual reclaim in batch
> of SGX_NR_TO_SCAN pages): in rare cases when LRUs have less than
> SGX_NR_TO_SCAN reclaimable pages, the actual reclaim of EPC pages will
> be split into smaller batches _across_ multiple LRUs with each being
> smaller than SGX_NR_TO_SCAN pages.
>
> A more precise way to mimic the current global EPC reclaim would be to
> have a new function to only "scan" (or "isolate") SGX_NR_TO_SCAN pages
> _across_ the given EPC cgroup _AND_ its descendants, and then do the
> actual reclaim in one batch. But this is unnecessarily complicated at
> this stage to address such rare cases.
>
> Signed-off-by: Haitao Huang <haitao.huang@xxxxxxxxxxxxxxx>
> ---
> arch/x86/kernel/cpu/sgx/epc_cgroup.c | 12 ++++++++++++
> arch/x86/kernel/cpu/sgx/epc_cgroup.h | 3 +++
> arch/x86/kernel/cpu/sgx/main.c | 7 +++++++
> 3 files changed, 22 insertions(+)
>
> diff --git a/arch/x86/kernel/cpu/sgx/epc_cgroup.c b/arch/x86/kernel/cpu/sgx/epc_cgroup.c
> index ae65cac858f8..23a61689e0d9 100644
> --- a/arch/x86/kernel/cpu/sgx/epc_cgroup.c
> +++ b/arch/x86/kernel/cpu/sgx/epc_cgroup.c
> @@ -240,6 +240,18 @@ static bool sgx_cgroup_should_reclaim(struct sgx_cgroup *sgx_cg)
> return (cur >= max);
> }
>
> +/**
> + * sgx_cgroup_reclaim_pages_global() - Perform one round of global reclamation.
> + *
> + * @charge_mm: The mm to be charged for the backing store of reclaimed pages.
> + *
> + * Try to scan and attempt reclamation from root cgroup for %SGX_NR_TO_SCAN pages.
> + */
> +void sgx_cgroup_reclaim_pages_global(struct mm_struct *charge_mm)
> +{
> + sgx_cgroup_reclaim_pages(&sgx_cg_root, charge_mm, SGX_NR_TO_SCAN);
> +}
> +
>
[...]
>
> static void sgx_reclaim_pages_global(struct mm_struct *charge_mm)
> {
> + /*
> + * Now all EPC pages are still tracked in the @sgx_global_lru.
> + * Still reclaim from it.
> + *
> + * When EPC pages are tracked in the actual per-cgroup LRUs,
> + * sgx_cgroup_reclaim_pages_global() will be called.
> + */
> sgx_reclaim_pages(&sgx_global_lru, charge_mm);
> }
>
I didn't realize the only functional change of this patch is to add a new
helper sgx_cgroup_reclaim_pages_global() (hmm... it's not even a functional
change because the helper isn't called).
It might make sense to have this as a separate patch with the comment of
sgx_can_reclaim_global() being moved here, from the perspective that this
patch only adds the building block to do global reclaim from per-cgroup LRUs
but doesn't actually turn that on.
But given this patch only adds a helper it might also make sense to just merge
this with the patch "x86/sgx: Turn on per-cgroup EPC reclamation".
I have no hard opinion on this. Maybe we can keep the current way unless
other people comment in.
Btw, I don't quite follow why you placed this patch here. It should be placed
right before the patch "x86/sgx: Turn on per-cgroup EPC reclamation" (if we
want to make this as a separate patch).
Or, the patch "x86/sgx: implement direct reclamation for cgroups" should be
moved to an earlier place after patch "x86/sgx: Implement async reclamation
for cgroup".
