On Wed Aug 21, 2024 at 4:53 AM EEST, Haitao Huang wrote: > Currently in the EPC page allocation, the kernel simply fails the > allocation when the current EPC cgroup fails to charge due to its usage > reaching limit. This is not ideal. When that happens, a better way is > to reclaim EPC page(s) from the current EPC cgroup to reduce its usage > so the new allocation can succeed. > > Currently, all EPC pages are tracked in a single global LRU, and the > "global EPC reclamation" supports the following 3 cases: > > 1) On-demand asynchronous reclamation: For allocation requests that can > not wait for reclamation but can be retried, an asynchronous > reclamation is triggered, in which the global reclaimer, ksgxd, keeps > reclaiming EPC pages until the free page count is above a minimal > threshold. > > 2) On-demand synchronous reclamation: For allocations that can wait for > reclamation, the EPC page allocator, sgx_alloc_epc_page() reclaims > EPC page(s) immediately until at least one free page is available for > allocation. > > 3) Preemptive reclamation: For some allocation requests, e.g., > allocation for reloading a reclaimed page to change its permissions > or page type, the kernel invokes sgx_reclaim_direct() to preemptively > reclaim EPC page(s) as a best effort to minimize on-demand > reclamation for subsequent allocations. > > Similarly, a "per-cgroup reclamation" is needed to support the above 3 > cases as well: > > 1) For on-demand asynchronous reclamation, a per-cgroup reclamation > needs to be invoked to maintain a minimal difference between the > usage and the limit for each cgroup, analogous to the minimal free > page threshold maintained by the global reclaimer. > > 2) For on-demand synchronous reclamation, sgx_cgroup_try_charge() needs > to invoke the per-cgroup reclamation until the cgroup usage become > at least one page lower than its limit. > > 3) For preemptive reclamation, sgx_reclaim_direct() needs to invoke the > per-cgroup reclamation to minimize per-cgroup on-demand reclamation > for subsequent allocations. > > To support the per-cgroup reclamation, introduce a "per-cgroup LRU" to > track all EPC pages belong to the owner cgroup to utilize the existing > sgx_reclaim_pages(). > > Currently, the global reclamation treats all EPC pages equally as it > scans all EPC pages in FIFO order in the global LRU. The "per-cgroup > reclamation" needs to somehow achieve the same fairness of all EPC pages > that are tracked in the multiple LRUs of the given cgroup and all the > descendants to reflect the nature of the cgroup. > > The idea is to achieve such fairness by scanning "all EPC cgroups" of > the subtree (the given cgroup and all the descendants) equally in turns, > and in the scan to each cgroup, apply the existing sgx_reclaim_pages() > to its LRU. This basic flow is encapsulated in a new function, > sgx_cgroup_reclaim_pages(). > > Export sgx_reclaim_pages() for use in sgx_cgroup_reclaim_pages(). And > modify sgx_reclaim_pages() to return the number of pages scanned so > sgx_cgroup_reclaim_pages() can track scanning progress and determine > whether enough scanning is done or to continue the scanning for next > descendant. > > Whenever reclaiming in a subtree of a given root is needed, start the > scanning from the next descendant where scanning was stopped at last > time. To keep track of the next descendant cgroup to scan, add a new > field, next_cg, in the sgx_cgroup struct. Create an iterator function, > sgx_cgroup_next_get(), atomically returns a valid reference of the > descendant for next round of scanning and advances @next_cg to next > valid descendant in a preorder walk. This iterator function is used in > sgx_cgroup_reclaim_pages() to iterate descendants for scanning. > Separately also advances @next_cg to next valid descendant when the > cgroup referenced by @next_cg is to be freed. > > Add support for on-demand synchronous reclamation in > sgx_cgroup_try_charge(), applying sgx_cgroup_reclaim_pages() iteratively > until cgroup usage is lower than its limit. > > Later patches will reuse sgx_cgroup_reclaim_pages() to add support for > asynchronous and preemptive reclamation. > > Note all reclaimable EPC pages are still tracked in the global LRU thus > no per-cgroup reclamation is actually active at the moment: -ENOMEM is > returned by __sgx_cgroup_try_charge() when LRUs are empty. Per-cgroup > tracking and reclamation will be turned on in the end after all > necessary infrastructure is in place. > > Co-developed-by: Sean Christopherson <sean.j.christopherson@xxxxxxxxx> > Signed-off-by: Sean Christopherson <sean.j.christopherson@xxxxxxxxx> > Co-developed-by: Kristen Carlson Accardi <kristen@xxxxxxxxxxxxxxx> > Signed-off-by: Kristen Carlson Accardi <kristen@xxxxxxxxxxxxxxx> > Signed-off-by: Haitao Huang <haitao.huang@xxxxxxxxxxxxxxx> Reviewed-by: Jarkko Sakkinen <jarkko@xxxxxxxxxx> BR, Jarkko
