1SD_EVENT_ADD_MEMORY_PRESSUsRdE_(e3v)ent_add_memory_preSsDs_uErVeENT_ADD_MEMORY_PRESSURE(3)
2
3
4
6 sd_event_add_memory_pressure, sd_event_source_set_memory_pressure_type,
7 sd_event_source_set_memory_pressure_period, sd_event_trim_memory - Add
8 and configure an event source run as result of memory pressure
9
11 #include <systemd/sd-event.h>
12
13 typedef struct sd_event_source sd_event_source;
14
15 int sd_event_add_memory_pressure(sd_event *event,
16 sd_event_source **ret_source,
17 sd_event_handler_t handler,
18 void *userdata);
19
20 int sd_event_source_set_memory_pressure_type(sd_event_source *source,
21 const char *type);
22
23 int sd_event_source_set_memory_pressure_period(sd_event_source *source,
24 uint64_t threshold_usec,
25 uint64_t window_usec);
26
27 int sd_event_trim_memory(void);
28
30 sd_event_add_memory_pressure() adds a new event source that is
31 triggered whenever memory pressure is seen. This functionality is built
32 around the Linux kernel's Pressure Stall Information (PSI)[1] logic.
33
34 Expects an event loop object as first parameter, and returns the
35 allocated event source object in the second parameter, on success. The
36 handler parameter is a function to call when memory pressure is seen,
37 or NULL. The handler function will be passed the userdata pointer,
38 which may be chosen freely by the caller. The handler may return
39 negative to signal an error (see below), other return values are
40 ignored. If handler is NULL, a default handler that compacts allocation
41 caches maintained by libsystemd as well as glibc (via malloc_trim(3))
42 will be used.
43
44 To destroy an event source object use sd_event_source_unref(3), but
45 note that the event source is only removed from the event loop when all
46 references to the event source are dropped. To make sure an event
47 source does not fire anymore, even if it is still referenced, disable
48 the event source using sd_event_source_set_enabled(3) with
49 SD_EVENT_OFF.
50
51 If the second parameter of sd_event_add_memory_pressure() is NULL no
52 reference to the event source object is returned. In this case the
53 event source is considered "floating", and will be destroyed implicitly
54 when the event loop itself is destroyed.
55
56 The event source will fire according to the following logic:
57
58 1. If the $MEMORY_PRESSURE_WATCH/$MEMORY_PRESSURE_WRITE environment
59 variables are set at the time the event source is established, it
60 will watch the file, FIFO or AF_UNIX socket specified via
61 $MEMORY_PRESSURE_WATCH (which must contain an absolute path name)
62 for POLLPRI (in case it is a regular file) or POLLIN events
63 (otherwise). After opening the inode, it will write the (decoded)
64 Base64 data provided via $MEMORY_PRESSURE_WRITE into it before it
65 starts polling on it (the variable may be unset in which case this
66 is skipped). Typically, if used, $MEMORY_PRESSURE_WATCH will
67 contain a path such as /proc/pressure/memory or a path to a
68 specific memory.pressure file in the control group file system
69 (cgroupfs).
70
71 2. If these environment variables are not set, the local PSI interface
72 file memory.pressure of the control group the invoking process is
73 running in is used.
74
75 3. If that file does not exist, the system-wide PSI interface file
76 /proc/pressure/memory is watched instead.
77
78 Or in other words: preferably any explicit configuration passed in by
79 an invoking service manager (or similar) is used as notification
80 source, before falling back to local notifications of the service, and
81 finally to global notifications of the system.
82
83 Well-behaving services and applications are recommended to react to
84 memory pressure events by executing one or more of the following
85 operations, in order to ensure optimal behaviour even on loaded and
86 resource-constrained systems:
87
88 • Release allocation caches such as malloc_trim() or similar, both
89 implemented in the libraries consumed by the program and in private
90 allocation caches of the program itself.
91
92 • Release any other form of in-memory caches that can easily be
93 recovered if needed (e.g. browser caches).
94
95 • Terminate idle worker threads or processes, or similar.
96
97 • Even exit entirely from the program if it is idle and can be
98 automatically started when needed (for example via socket or bus
99 activation).
100
101 Any of the suggested operations should help easing memory pressure
102 situations and allowing the system to make progress by reclaiming the
103 memory for other purposes.
104
105 This event source typically fires on memory pressure stalls, i.e. when
106 operational latency above a configured threshold already has been seen.
107 This should be taken into consideration when discussing whether later
108 latency to re-aquire any released resources is acceptable: it's usually
109 more important to think of the latencies that already happened than
110 those coming up in future.
111
112 The sd_event_source_set_memory_pressure_type() and
113 sd_event_source_set_memory_pressure_period() functions can be used to
114 fine-tune the PSI parameters for pressure notifications. The former
115 takes either "some", "full" as second parameter, the latter takes
116 threshold and period times in microseconds as parameters. For details
117 about these three parameters see the PSI documentation. Note that these
118 two calls must be invoked immediately after allocating the event
119 source, as they must be configured before polling begins. Also note
120 that these calls will fail if memory pressure parameterization has been
121 passed in via the $MEMORY_PRESSURE_WATCH/$MEMORY_PRESSURE_WRITE
122 environment variables (or in other words: configuration supplied by a
123 service manager wins over internal settings).
124
125 The sd_event_trim_memory() function releases various internal
126 allocation caches maintained by libsystemd and then invokes glibc's
127 malloc_trim(3). This makes the operation executed when the handler
128 function parameter of sd_event_add_memory_pressure is passed as NULL
129 directly accessible for invocation at any time (see above). This
130 function will log a structured log message at LOG_DEBUG level (with
131 message ID f9b0be465ad540d0850ad32172d57c21) about the memory pressure
132 operation.
133
134 For further details see Memory Pressure Handling in systemd[2].
135
137 On success, these functions return 0 or a positive integer. On failure,
138 they return a negative errno-style error code.
139
140 Errors
141 Returned errors may indicate the following problems:
142
143 -ENOMEM
144 Not enough memory to allocate an object.
145
146 -EINVAL
147 An invalid argument has been passed.
148
149 -EHOSTDOWN
150 The $MEMORY_PRESSURE_WATCH variable has been set to the literal
151 string /dev/null, in order to explicitly disable memory pressure
152 handling.
153
154 -EBADMSG
155 The $MEMORY_PRESSURE_WATCH variable has been set to an invalid
156 string, for example a relative rather than an absolute path.
157
158 -ENOTTY
159 The $MEMORY_PRESSURE_WATCH variable points to a regular file
160 outside of the procfs or cgroupfs file systems.
161
162 -EOPNOTSUPP
163 No configuration via $MEMORY_PRESSURE_WATCH has been specified and
164 the local kernel does not support the PSI interface.
165
166 -EBUSY
167 This is returned by sd_event_source_set_memory_pressure_type() and
168 sd_event_source_set_memory_pressure_period() if invoked on event
169 sources at a time later than immediately after allocating them.
170
171 -ESTALE
172 The event loop is already terminated.
173
174 -ECHILD
175 The event loop has been created in a different process, library or
176 module instance.
177
178 -EDOM
179 The passed event source is not a signal event source.
180
182 Functions described here are available as a shared library, which can
183 be compiled against and linked to with the libsystemd pkg-config(1)
184 file.
185
186 The code described here uses getenv(3), which is declared to be not
187 multi-thread-safe. This means that the code calling the functions
188 described here must not call setenv(3) from a parallel thread. It is
189 recommended to only do calls to setenv() from an early phase of the
190 program when no other threads have been started.
191
193 systemd(1), sd-event(3), sd_event_new(3), sd_event_add_io(3),
194 sd_event_add_time(3), sd_event_add_child(3), sd_event_add_inotify(3),
195 sd_event_add_defer(3), sd_event_source_set_enabled(3),
196 sd_event_source_set_description(3), sd_event_source_set_userdata(3),
197 sd_event_source_set_floating(3)
198
200 1. Pressure Stall Information (PSI)
201 https://docs.kernel.org/accounting/psi.html
202
203 2. Memory Pressure Handling in systemd
204 https://systemd.io/MEMORY_PRESSURE
205
206
207
208systemd 254 SD_EVENT_ADD_MEMORY_PRESSURE(3)