When the system is under memory pressure (i.e. some component of the OS requires memory allocation but there is only very little or none available), it can attempt various things to make more memory available again (“reclaim”):
The kernel can flush out memory pages backed by files on disk, under the knowledge that it can reread them from disk when needed again. Candidate pages are the many memory mapped executable files and shared libraries on disk, among others.
The kernel can flush out memory packages not backed by files on disk
(“anonymous” memory, i.e. memory allocated via malloc() and similar calls,
or tmpfs file system contents) if there’s swap to write it to.
Userspace can proactively release memory it allocated but doesn’t immediately require back to the kernel. This includes allocation caches, and other forms of caches that are not required for normal operation to continue.
The latter is what we want to focus on in this document: how to ensure userspace process can detect mounting memory pressure early and release memory back to the kernel as it happens, relieving the memory pressure before it becomes too critical.
The effects of memory pressure during runtime generally are growing latencies during operation: when a program requires memory but the system is busy writing out memory to (relatively slow) disks in order make some available, this generally surfaces in scheduling latencies, and applications and services will slow down until memory pressure is relieved. Hence, to ensure stable service latencies it is essential to release unneeded memory back to the kernel early on.
On Linux the Pressure Stall Information
(PSI) Linux kernel interface is
the primary way to determine the system or a part of it is under memory
pressure. PSI makes available to userspace a poll()-able file descriptor that
gets notifications whenever memory pressure latencies for the system or a
control group grow beyond some level.
systemd itself makes use of PSI, and helps applications to do so too.
Specifically:
Most of systemd’s long running components watch for PSI memory pressure events, and release allocation caches and other resources once seen.
systemd’s service manager provides a protocol for asking services to monitor PSI events and configure the appropriate pressure thresholds.
systemd’s sd-event event loop API provides a high-level call
sd_event_add_memory_pressure() enabling programs using it to efficiently
hook into the PSI memory pressure protocol provided by the service manager,
with very few lines of code.
If memory pressure handling for a specific service is enabled via
MemoryPressureWatch= the memory pressure service protocol is used to tell the
service code about this. Specifically two environment variables are set by the
service manager, and typically consumed by the service:
The $MEMORY_PRESSURE_WATCH environment variable will contain an absolute
path in the file system to the file to watch for memory pressure events. This
will usually point to a PSI file such as the memory.pressure file of the
service’s cgroup. In order to make debugging easier, and allow later
extension it is recommended for applications to also allow this path to refer
to an AF_UNIX stream socket in the file system or a FIFO inode in the file
system. Regardless which of the three types of inodes this absolute path
refers to, all three are poll()-able for memory pressure events. The
variable can also be set to the literal string /dev/null. If so the service
code should take this as indication that memory pressure monitoring is not
desired and should be turned off.
The $MEMORY_PRESSURE_WRITE environment variable is optional. If set by the
service manager it contains Base64 encoded data (that may contain arbitrary
binary values, including NUL bytes) that should be written into the path
provided via $MEMORY_PRESSURE_WATCH right after opening it. Typically, if
talking directly to a PSI kernel file this will contain information about the
threshold settings configurable in the service manager.
When a service initializes it hence should look for
$MEMORY_PRESSURE_WATCH. If set, it should try to open the specified path. If
it detects the path to refer to a regular file it should assume it refers to a
PSI kernel file. If so, it should write the data from $MEMORY_PRESSURE_WRITE
into the file descriptor (after Base64-decoding it, and only if the variable is
set) and then watch for POLLPRI events on it. If it detects the paths refers
to a FIFO inode, it should open it, write the $MEMORY_PRESSURE_WRITE data
into it (as above) and then watch for POLLIN events on it. Whenever POLLIN
is seen it should read and discard any data queued in the FIFO. If the path
refers to an AF_UNIX socket in the file system, the application should
connect() a stream socket to it, write $MEMORY_PRESSURE_WRITE into it (as
above) and watch for POLLIN, discarding any data it might receive.
To summarize:
If $MEMORY_PRESSURE_WATCH points to a regular file: open and watch for
POLLPRI, never read from the file descriptor.
If $MEMORY_PRESSURE_WATCH points to a FIFO: open and watch for POLLIN,
read/discard any incoming data.
If $MEMORY_PRESSURE_WATCH points to an AF_UNIX socket: connect and watch
for POLLIN, read/discard any incoming data.
If $MEMORY_PRESSURE_WATCH contains the literal string /dev/null, turn off
memory pressure handling.
(And in each case, immediately after opening/connecting to the path, write the
decoded $MEMORY_PRESSURE_WRITE data into it.)
Whenever a POLLPRI/POLLIN event is seen the service is under memory
pressure. It should use this as hint to release suitable redundant resources,
for example:
glibc’s memory allocation cache, via
malloc_trim(). Similar,
allocation caches implemented in the service itself.
Any other local caches, such DNS caches, or web caches (in particular if service is a web browser).
Terminate any idle worker threads or processes.
Run a garbage collection (GC) cycle, if the runtime environment supports it.
Terminate the process if idle, and can be automatically started when needed next.
Which actions precisely to take depends on the service in question. Note that the notifications are delivered when memory allocation latency already degraded beyond some point. Hence when discussing which resources to keep and which to discard, keep in mind it’s typically acceptable that latencies incurred recovering discarded resources at a later point are acceptable, given that latencies already are affected negatively.
In case the path supplied via $MEMORY_PRESSURE_WATCH points to a PSI kernel
API file, or to an AF_UNIX opening it multiple times is safe and reliable,
and should deliver notifications to each of the opened file descriptors. This
is specifically useful for services that consist of multiple processes, and
where each of them shall be able to release resources on memory pressure.
The POLLPRI/POLLIN conditions will be triggered every time memory pressure
is detected, but not continuously. It is thus safe to keep poll()-ing on the
same file descriptor continuously, and executing resource release operations
whenever the file descriptor triggers without having to expect overloading the
process.
(Currently, the protocol defined here only allows configuration of a single
“degree” of memory pressure, there’s no distinction made on how strong the
pressure is. In future, if it becomes apparent that there’s clear need to
extend this we might eventually add different degrees, most likely by adding
additional environment variables such as $MEMORY_PRESSURE_WRITE_LOW and
$MEMORY_PRESSURE_WRITE_HIGH or similar, which may contain different settings
for lower or higher memory pressure thresholds.)
The service manager provides two per-service settings that control the memory pressure handling:
The
MemoryPressureWatch=
setting controls whether to enable the memory pressure protocol for the
service in question.
The MemoryPressureThresholdSec= setting allows to configure the threshold
when to signal memory pressure to the services. It takes a time value
(usually in the millisecond range) that defines a threshold per 1s time
window: if memory allocation latencies grow beyond this threshold
notifications are generated towards the service, requesting it to release
resources.
The /etc/systemd/system.conf file provides two settings that may be used to
select the default values for the above settings. If the threshold isn’t
configured via the per-service nor system-wide option, it defaults to 100ms.
When memory pressure monitoring is enabled for a service via
MemoryPressureWatch= this primarily does three things:
It enables cgroup memory accounting for the service (this is a requirement for per-cgroup PSI)
It sets the aforementioned two environment variables for processes invoked for the service, based on the control group of the service and provided settings.
The memory.pressure PSI control group file associated with the service’s
cgroup is delegated to the service (i.e. permissions are relaxed so that
unprivileged service payload code can open the file for writing).
sd-eventThe
sd-event
event loop library provides two API calls that encapsulate the
functionality described above:
The
sd_event_add_memory_pressure()
call implements the service-side of the memory pressure protocol and
integrates it with an sd-event event loop. It reads the two environment
variables, connects/opens the specified file, writes the specified data to it,
then watches it for events.
The sd_event_trim_memory() call may be called to trim the calling
processes’ memory. It’s a wrapper around glibc’s malloc_trim(), but first
releases allocation caches maintained by libsystemd internally. This function
serves as the default when a NULL callback is supplied to
sd_event_add_memory_pressure().
When implementing a service using sd-event, for automatic memory pressure
handling, it’s typically sufficient to add a line such as:
(void) sd_event_add_memory_pressure(event, NULL, NULL, NULL);
– right after allocating the event loop object event.
Other programming environments might have native APIs to watch memory pressure/low memory events. Most notable is probably GLib’s GMemoryMonitor. It currently uses the per-system Linux PSI interface as the backend, but operates differently than the above: memory pressure events are picked up by a system service, which then propagates this through D-Bus to the applications. This is typically less than ideal, since this means each notification event has to traverse three processes before being handled. This traversal creates additional latencies at a time where the system is already experiencing adverse latencies. Moreover, it focusses on system-wide PSI events, even though service-local ones are generally the better approach.