diff mbox

[intel-sgx-kernel-dev,RFC,12/12] docs, cgroup, intel_sgx: add SGX EPC cgroup controller documentation

Message ID 1497461858-20309-13-git-send-email-sean.j.christopherson@intel.com (mailing list archive)
State New, archived
Headers show

Commit Message

Sean Christopherson June 14, 2017, 5:37 p.m. UTC
Signed-off-by: Sean Christopherson <sean.j.christopherson@intel.com>
---
 Documentation/cgroup-v2.txt | 205 +++++++++++++++++++++++++++++++++++++++++++-
 1 file changed, 201 insertions(+), 4 deletions(-)
diff mbox

Patch

diff --git a/Documentation/cgroup-v2.txt b/Documentation/cgroup-v2.txt
index dc5e2dc..2c85cd3 100644
--- a/Documentation/cgroup-v2.txt
+++ b/Documentation/cgroup-v2.txt
@@ -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