@@ -51,8 +51,12 @@ CONTENTS
5-4-1. PID Interface Files
5-5. RDMA
5-5-1. RDMA Interface Files
- 5-6. Misc
- 5-6-1. perf_event
+ 5-6. SGX EPC
+ 5-6-1. SGX EPC Interface Files
+ 5-6-2. Usage Guidelines
+ 5-6-3. Migration
+ 5-7. Misc
+ 5-7-1. perf_event
6. Namespace
6-1. Basics
6-2. The Root and Views
@@ -1216,9 +1220,202 @@ of RDMA resources.
ocrdma1 hca_handle=1 hca_object=23
-5-6. Misc
+5-6. SGX EPC
-5-6-1. perf_event
+The "sgx_epc" controller regulates distribution of SGX EPC memory,
+which is a subset of system RAM that is used to provide SGX-enabled
+applications with protected memory, and is otherwise inaccessible,
+i.e. shows up as reserved in /proc/iomem and cannot be read/written
+outside of an SGX enclave.
+
+EPC memory is independent from normal system memory; it is managed
+by the SGX subsystem and is not accounted by the memory controller.
+This approach is necessary as EPC memory must be reserved at boot,
+i.e. memory cannot be converted between EPC and normal memory while
+the system is running. Note that this is true only for EPC memory
+itself; normal memory allocations related to SGX and EPC memory,
+e.g. the shmem backing file for EPC, are accounted, limited and
+protected by the memory controller.
+
+Much like normal system memory, EPC memory can be overcommitted via
+virtual memory techniques and pages can be swapped out of the EPC
+to their backing store (normal system memory allocated via shmem).
+The SGX EPC subsystem is analogous to the memory subsytem, and the
+SGX EPC controller is in turn analogous to the memory controller;
+it implements limit and protection models for EPC memory.
+
+See Documentation/x86/intel_sgx.rst for more info on SGX and EPC.
+
+5-6-1. SGX EPC Interface Files
+
+All SGX EPC memory amounts are in bytes unless explicitly stated
+otherwise. If a value which is not PAGE_SIZE aligned is written,
+the actual value used by the controller will be rounded down to
+the closest PAGE_SIZE multiple.
+
+ sgx_epc.current
+
+ A read-only single value file which exists on all cgroups.
+
+ The total amount of EPC memory currently being used by the
+ cgroup and its descendants.
+
+ sgx_epc.low
+
+ A read-write single value file which exists on non-root
+ cgroups. The default is "0".
+
+ Best-effort protection of EPC usage. If the EPC usage of a
+ cgroup is below its limits, and all its ancestors are below
+ their low limits, then the cgroup's EPC won't be reclaimed
+ unless EPC cannot be reclaimed from unprotected cgroups,
+ e.g. all sibling cgroups are also below their low limit.
+
+ Setting low to a value more than the amount of EPC available
+ is discouraged. The low limit is effectively ignored if the
+ cgroup's high or max limit is less than its low limit.
+
+ sgx_epc.high
+
+ A read-write single value file which exists on non-root
+ cgroups. The default is "max".
+
+ EPC usage best-effort limit. This is the main mechanism to
+ control EPC usage of a cgroup. If a cgroup's usage goes
+ over the high boundary, EPC pages will be reclaimed from
+ the cgroup until it is back under the high limit.
+
+ Going over the high limit does not prevent allocation of
+ additional EPC pages, e.g. EPC usage will often spike above
+ the high limit during enclave creation, when a large number
+ of EPC pages are EADDed in a short period.
+
+ sgx_epc.max
+
+ A read-write single value file which exists on non-root
+ cgroups. The default is "max".
+
+ EPC usage hard limit. If a cgroup's EPC usage reaches this
+ limit, EPC allocations, e.g. for page fault handling, will
+ be blocked until EPC can be reclaimed from the cgroup. If
+ EPC cannot be reclaimed in a timely manner, reclaim will be
+ forced, e.g. by ignoring LRU.
+
+ The max limit is intended to be a last line of defense; it
+ should rarely come into play on a properly configured and
+ monitored system.
+
+ sgx_epc.stats
+
+ A read-write flat-keyed file which exists on all cgroups.
+ Reads from the file display the cgroup's statistics, while
+ writes reset the underlying counters (if applicable).
+
+ The entries are ordered to be human readable, and new entries
+ can show up in the middle. Don't rely on items remaining in a
+ fixed position; use the keys to look up specific values!
+
+ The following entries are defined.
+
+ pages
+
+ The total number of pages currently being used by the
+ cgroup and its descendants, i.e. sgx_epc.current / 4096.
+
+ direct
+
+ The number of pages currently being used by the cgroup
+ itself, excluding its descendants.
+
+ indirect
+
+ The number of pages currently being used by the cgroup's
+ descendants, excluding its own pages.
+
+ reclaimed
+
+ The number of pages that have been reclaimed from the
+ cgroup (since sgx_epc.stats was last reset).
+
+ reclamations
+
+ The number of times this cgroup's LRU lists have been
+ scanned for reclaim, i.e. the number of times the cgroup
+ has been selected for reclaim via any code path.
+
+ sgx_epc.events
+
+ A read-write flat-keyed file which exists on non-root cgroups.
+ Writes to the file reset the event counters to zero. A value
+ change in this file generates a file modified event.
+
+ The following entries are defined.
+
+ low
+
+ The number of times the cgroup has been reclaimed even
+ though its usage is under the low boundary, e.g. due to
+ all sibling cgroups also being low. This event usually
+ indicates that the low boundary is over-committed.
+
+ high
+
+ The number of times the cgroup has triggered a reclaim
+ due to its EPC usage exceeding its high EPC boundary.
+ This event is expected for cgroups whose EPC usage is
+ capped by its high limit rather than global pressure.
+
+ max
+
+ The number of times the cgroup has triggered a reclaim
+ due to its EPC usage approaching (or exceeding) its max
+ EPC boundary.
+
+5-6-2. Usage Guidelines
+
+"sgx_epc.high" and "sgx_epc.low" are the main mechanisms to control
+EPC usage; using "sgx_epc.max" as anything other than a safety net
+is inadvisable, SGX application performance will suffer greatly if
+a process encounters its max limit. Because a cgroup is allowed to
+breach its high limit, e.g. to fault in a page, performance is not
+artificially limited, whereas the max limit will effectively block
+a faulting application until the kernel can reclaim EPC memory from
+the cgroup.
+
+Exactly how "sgx_epc.high" is utilized will vary case by case, i.e.
+there is no one "correct" strategy. Deferring to global EPC memory
+pressure, e.g. by overcommitting on the high limit, may be the most
+effective approach for a particular situation, whereas a different
+scenario might warrant a more draconian usage of the high limit.
+Regardless of the strategy used, because breach of the high limit
+does not cause processes to block or be killed, a management agent
+has ample opportunity to monitor and react as needed, e.g. it can
+raise the offending cgroup's high limit or terminate the workload.
+
+Similarly, "sgx_epc.low" can play different roles depending on the
+situation, e.g. it can be set to a relatively high value to protect
+a mission critical workload, or it may be used to reserve a minimal
+amount of EPC memory simply to ensure forward progress. Employing
+"sgx_epc.low" in some capacity is generally recommended, especially
+when overcommitting "sgx_epc.high", as it is relatively common for
+a system to be under heavy EPC pressure; this holds true even on a
+carefully tuned system, as initializing an enclave requires all of
+the enclave's pages be brought into the EPC at some point prior to
+initialization, if only temporarily.
+
+
+5-6-3. Migration
+
+Once an EPC page is charged to a cgroup (during allocation), it
+remains charged to the original cgroup until the page is released
+or reclaimed. Migrating a process to a different cgroup doesn't
+move the EPC charges that it incurred while in the previous cgroup
+to its new cgroup.
+
+
+5-7. Misc
+
+5-7-1. perf_event
perf_event controller, if not mounted on a legacy hierarchy, is
automatically enabled on the v2 hierarchy so that perf events can
Signed-off-by: Sean Christopherson <sean.j.christopherson@intel.com> --- Documentation/cgroup-v2.txt | 205 +++++++++++++++++++++++++++++++++++++++++++- 1 file changed, 201 insertions(+), 4 deletions(-)