RTKit does not work on musl libc as a result of lacking sched_getscheduler and sched_setscheduler
Opened this issue · 5 comments
As in the title, musl libc does not implement these functions, since Linux's syscall behaviour is seen as non-compliant with the POSIX standard as written (it takes a thread ID rather than a normal pid, and returns 0 on success, rather than returning the previous policy).
Void Linux currently uses the attached patch to implement sched_getscheduler and sched_setscheduler for musl.
This patch vendors in the implementations of sched_getscheduler() and
sched_setscheduler() from
https://git.musl-libc.org/cgit/musl/commit/?id=61be1cfec1f5da66c68f92a6939e3a38e673c9d6
that were subsequently dropped from higher POSIX compatibility. However, rtkit
needs those functions to work. Since each function is ~ 1 line of code, we
can just vendor them in here.
For more details, see discussion in
https://gitlab.alpinelinux.org/alpine/aports/-/merge_requests/19915
--- rtkit-0.13.orig/rtkit-daemon.c
+++ rtkit-0.13/rtkit-daemon.c
@@ -32,6 +32,7 @@
#include <sys/stat.h>
#include <string.h>
#include <sched.h>
+#include <syscall.h>
#include <sys/time.h>
#include <sys/resource.h>
#include <stdlib.h>
@@ -196,6 +197,15 @@
static pthread_t canary_thread_id = 0, watchdog_thread_id = 0;
static volatile uint32_t refuse_until = 0;
+static int _sched_setscheduler(pid_t pid, int sched, const struct sched_param *param) {
+ static const struct sched_param def;
+ return syscall(SYS_sched_setscheduler, pid, 0, &def);
+}
+
+static int _sched_getscheduler(pid_t pid) {
+ return syscall(SYS_sched_getscheduler, pid);
+}
+
static const char *get_proc_path(void) {
/* Useful for chroot environments */
@@ -439,7 +449,7 @@
if (st != t->starttime)
return FALSE;
- if ((r = sched_getscheduler(t->pid)) < 0) {
+ if ((r = _sched_getscheduler(t->pid)) < 0) {
/* Maybe it died right now? */
if (errno == ESRCH)
@@ -564,7 +574,7 @@
memset(¶m, 0, sizeof(param));
param.sched_priority = priority;
- if (sched_setscheduler(0, sched_policy|SCHED_RESET_ON_FORK, ¶m) < 0) {
+ if (_sched_setscheduler(0, sched_policy|SCHED_RESET_ON_FORK, ¶m) < 0) {
r = -errno;
syslog(LOG_ERR, "Failed to make ourselves RT: %s\n", strerror(errno));
goto finish;
@@ -581,7 +591,7 @@
memset(¶m, 0, sizeof(param));
- if (sched_setscheduler(0, SCHED_OTHER, ¶m) < 0)
+ if (_sched_setscheduler(0, SCHED_OTHER, ¶m) < 0)
syslog(LOG_WARNING, "Warning: Failed to reset scheduling to SCHED_OTHER: %s\n", strerror(errno));
if (setpriority(PRIO_PROCESS, 0, nice_level) < 0)
@@ -696,7 +706,7 @@
memset(¶m, 0, sizeof(param));
param.sched_priority = 0;
- if (sched_setscheduler(tid, SCHED_OTHER, ¶m) < 0) {
+ if (_sched_setscheduler(tid, SCHED_OTHER, ¶m) < 0) {
if (errno != ESRCH)
syslog(LOG_WARNING, "Warning: Failed to reset scheduling to SCHED_OTHER for thread %llu: %s\n", (unsigned long long) tid, strerror(errno));
r = -1;
@@ -748,7 +758,7 @@
/* Ok, everything seems to be in order, now, let's do it */
memset(¶m, 0, sizeof(param));
param.sched_priority = (int) priority;
- if (sched_setscheduler(t->pid, sched_policy|SCHED_RESET_ON_FORK, ¶m) < 0) {
+ if (_sched_setscheduler(t->pid, sched_policy|SCHED_RESET_ON_FORK, ¶m) < 0) {
r = -errno;
syslog(LOG_ERR, "Failed to make thread %llu RT: %s\n", (unsigned long long) t->pid, strerror(errno));
goto finish;
@@ -809,7 +819,7 @@
/* Ok, everything seems to be in order, now, let's do it */
memset(¶m, 0, sizeof(param));
param.sched_priority = 0;
- if (sched_setscheduler(t->pid, SCHED_OTHER|SCHED_RESET_ON_FORK, ¶m) < 0) {
+ if (_sched_setscheduler(t->pid, SCHED_OTHER|SCHED_RESET_ON_FORK, ¶m) < 0) {
r = -errno;
syslog(LOG_ERR, "Failed to make process %llu SCHED_NORMAL: %s\n", (unsigned long long) t->pid, strerror(errno));
goto finish;
@@ -943,7 +953,7 @@
if (errno != 0 || !e || *e != 0)
continue;
- if ((r = sched_getscheduler(tid)) < 0) {
+ if ((r = _sched_getscheduler(tid)) < 0) {
if (errno != ESRCH)
syslog(LOG_WARNING, "Warning: sched_getscheduler() failed: %s\n", strerror(errno));
continue;
This is not a particularly elegant fix, so if you would like me to I can make one with #ifdefs and the like to check for musl and only use the syscall / alternative implementation in that case.
+static int _sched_setscheduler(pid_t pid, int sched, const struct sched_param *param) {
+ static const struct sched_param def;
+ return syscall(SYS_sched_setscheduler, pid, 0, &def);
+}Why is the struct sched_param parameter ignored?
For SCHED_OTHER, SCHED_BATCH, and SCHED_IDLE, it must always contain .sched_priority=0, so it is not actually used by the kernel. For SCHED_FIFO and SCHED_RR, it does matter, and I am as-of-yet unsure as to why musl did not handle this possibility. The patch implements the sched_[set/get]scheduler functions exactly as they were in musl prior to being dropped.
Looking at this mailing list thread: https://www.openwall.com/lists/musl/2016/03/01/5
The reason it doesn't do anything is that Linux does not provide a way
to set scheduling parameters for a process, only for threads. The
sched_setscheduler syscall is documented as taking a pid but actually
takes a thread id and only operates on that thread. glibc just ignores
this and provides sched_* functions that do the wrong thing.Fortunately there's an easy fix: use pthread_setschedparam, and
pthread_self to get the pthread_t value you need to pass to it.
A more elegant solution would be for rtkit to switch to pthread_setschedparam?
Also see this commit description here: https://git.musl-libc.org/cgit/musl/commit/src/sched/sched_setparam.c?id=1e21e78bf7a5c24c217446d8760be7b7188711c2