1
1
mirror of https://github.com/systemd/systemd-stable.git synced 2024-12-23 17:34:00 +03:00

oom: make memory pressure duration configurable through oomd.conf

This commit is contained in:
Anita Zhang 2021-01-24 00:16:19 -08:00
parent 664e54b1bb
commit c20aa7b171
7 changed files with 31 additions and 8 deletions

View File

@ -65,13 +65,23 @@
will take action. A unit can override this value with <varname>ManagedOOMMemoryPressureLimitPercent=</varname>.
The memory pressure for this property represents the fraction of time in a 10 second window in which all tasks
in the cgroup were delayed. For each monitored cgroup, if the memory pressure on that cgroup exceeds the
limit set for more than 30 seconds, <command>systemd-oomd</command> will act on eligible descendant cgroups,
limit set for longer than the duration set by <varname>DefaultMemoryPressureDurationSec=</varname>,
<command>systemd-oomd</command> will act on eligible descendant cgroups,
starting from the ones with the most reclaim activity to the least reclaim activity. Which cgroups are
monitored and what action gets taken depends on what the unit has configured for
<varname>ManagedOOMMemoryPressure=</varname>. Takes a percentage value between 0% and 100%, inclusive.
Defaults to 60%.</para></listitem>
</varlistentry>
<varlistentry>
<term><varname>DefaultMemoryPressureDurationSec=</varname></term>
<listitem><para>Sets the amount of time a unit's cgroup needs to have exceeded memory pressure limits before
<command>systemd-oomd</command> will take action. Memory pressure limits are defined by
<varname>DefaultMemoryPressureLimitPercent=</varname> and <varname>ManagedOOMMemoryPressureLimitPercent=</varname>.
Defaults to 30 seconds when this property is unset or set to 0.</para></listitem>
</varlistentry>
</variablelist>
</refsect1>

View File

@ -306,7 +306,7 @@ static int monitor_cgroup_contexts_handler(sd_event_source *s, uint64_t usec, vo
m->post_action_delay_start = 0;
}
r = oomd_pressure_above(m->monitored_mem_pressure_cgroup_contexts, PRESSURE_DURATION_USEC, &targets);
r = oomd_pressure_above(m->monitored_mem_pressure_cgroup_contexts, m->default_mem_pressure_duration_usec, &targets);
if (r == -ENOMEM)
return log_error_errno(r, "Failed to check if memory pressure exceeded limits");
else if (r == 1) {
@ -325,7 +325,7 @@ static int monitor_cgroup_contexts_handler(sd_event_source *s, uint64_t usec, vo
SET_FOREACH(t, targets) {
log_notice("Memory pressure for %s is greater than %lu for more than %"PRIu64" seconds and there was reclaim activity",
t->path, LOAD_INT(t->mem_pressure_limit), PRESSURE_DURATION_USEC / USEC_PER_SEC);
t->path, LOAD_INT(t->mem_pressure_limit), m->default_mem_pressure_duration_usec / USEC_PER_SEC);
r = oomd_kill_by_pgscan(candidates, t->path, m->dry_run);
if (r == -ENOMEM)
@ -471,7 +471,7 @@ static int manager_connect_bus(Manager *m) {
return 0;
}
int manager_start(Manager *m, bool dry_run, int swap_used_limit, int mem_pressure_limit) {
int manager_start(Manager *m, bool dry_run, int swap_used_limit, int mem_pressure_limit, usec_t mem_pressure_usec) {
unsigned long l;
int r;
@ -487,6 +487,8 @@ int manager_start(Manager *m, bool dry_run, int swap_used_limit, int mem_pressur
if (r < 0)
return r;
m->default_mem_pressure_duration_usec = mem_pressure_usec ?: DEFAULT_MEM_PRESSURE_DURATION_USEC;
r = manager_connect_bus(m);
if (r < 0)
return r;
@ -505,6 +507,7 @@ int manager_start(Manager *m, bool dry_run, int swap_used_limit, int mem_pressur
int manager_get_dump_string(Manager *m, char **ret) {
_cleanup_free_ char *dump = NULL;
_cleanup_fclose_ FILE *f = NULL;
char buf[FORMAT_TIMESPAN_MAX];
OomdCGroupContext *c;
size_t size;
char *key;
@ -521,10 +524,12 @@ int manager_get_dump_string(Manager *m, char **ret) {
"Dry Run: %s\n"
"Swap Used Limit: %u%%\n"
"Default Memory Pressure Limit: %lu%%\n"
"Default Memory Pressure Duration: %s\n"
"System Context:\n",
yes_no(m->dry_run),
m->swap_used_limit,
LOAD_INT(m->default_mem_pressure_limit));
LOAD_INT(m->default_mem_pressure_limit),
format_timespan(buf, sizeof(buf), m->default_mem_pressure_duration_usec, USEC_PER_SEC));
oomd_dump_system_context(&m->system_context, f, "\t");
fprintf(f, "Swap Monitored CGroups:\n");

View File

@ -16,7 +16,7 @@
* percentage of time all tasks were delayed (i.e. unproductive).
* Generally 60 or higher might be acceptable for something like system.slice with no memory.high set; processes in
* system.slice are assumed to be less latency sensitive. */
#define PRESSURE_DURATION_USEC (30 * USEC_PER_SEC)
#define DEFAULT_MEM_PRESSURE_DURATION_USEC (30 * USEC_PER_SEC)
#define DEFAULT_MEM_PRESSURE_LIMIT 60
#define DEFAULT_SWAP_USED_LIMIT 90
@ -33,6 +33,7 @@ struct Manager {
bool dry_run;
unsigned swap_used_limit;
loadavg_t default_mem_pressure_limit;
usec_t default_mem_pressure_duration_usec;
/* k: cgroup paths -> v: OomdCGroupContext
* Used to detect when to take action. */
@ -53,7 +54,7 @@ DEFINE_TRIVIAL_CLEANUP_FUNC(Manager*, manager_free);
int manager_new(Manager **ret);
int manager_start(Manager *m, bool dry_run, int swap_used_limit, int mem_pressure_limit);
int manager_start(Manager *m, bool dry_run, int swap_used_limit, int mem_pressure_limit, usec_t mem_pressure_usec);
int manager_get_dump_string(Manager *m, char **ret);

View File

@ -31,6 +31,7 @@ struct OomdCGroupContext {
/* These are only used by oomd_pressure_above for acting on high memory pressure. */
loadavg_t mem_pressure_limit;
usec_t mem_pressure_duration_usec;
usec_t last_hit_mem_pressure_limit;
};

View File

@ -19,11 +19,13 @@
static bool arg_dry_run = false;
static int arg_swap_used_limit = -1;
static int arg_mem_pressure_limit = -1;
static usec_t arg_mem_pressure_usec = 0;
static int parse_config(void) {
static const ConfigTableItem items[] = {
{ "OOM", "SwapUsedLimitPercent", config_parse_percent, 0, &arg_swap_used_limit },
{ "OOM", "DefaultMemoryPressureLimitPercent", config_parse_percent, 0, &arg_mem_pressure_limit },
{ "OOM", "DefaultMemoryPressureDurationSec", config_parse_sec, 0, &arg_mem_pressure_usec },
{}
};
@ -160,7 +162,7 @@ static int run(int argc, char *argv[]) {
if (r < 0)
return log_error_errno(r, "Failed to create manager: %m");
r = manager_start(m, arg_dry_run, arg_swap_used_limit, arg_mem_pressure_limit);
r = manager_start(m, arg_dry_run, arg_swap_used_limit, arg_mem_pressure_limit, arg_mem_pressure_usec);
if (r < 0)
return log_error_errno(r, "Failed to start up daemon: %m");

View File

@ -14,3 +14,4 @@
[OOM]
#SwapUsedLimitPercent=90%
#DefaultMemoryPressureLimitPercent=60%
#DefaultMemoryPressureDurationSec=30s

View File

@ -14,12 +14,15 @@ if [[ "$cgroup_type" != *"cgroup2"* ]] && [[ "$cgroup_type" != *"0x63677270"* ]]
fi
[[ -e /skipped ]] && exit 0 || true
echo "DefaultMemoryPressureDurationSec=5s" >> /etc/systemd/oomd.conf
systemctl start testsuite-56-testbloat.service
systemctl start testsuite-56-testchill.service
# Verify systemd-oomd is monitoring the expected units
oomctl | grep "/testsuite-56-workload.slice"
oomctl | grep "1%"
oomctl | grep "Default Memory Pressure Duration: 5s"
# systemd-oomd watches for elevated pressure for 30 seconds before acting.
# It can take time to build up pressure so either wait 5 minutes or for the service to fail.