summaryrefslogtreecommitdiff
diff options
context:
space:
mode:
authorJeff Dike <jdike@addtoit.com>2007-11-29 03:21:51 +0300
committerLinus Torvalds <torvalds@woody.linux-foundation.org>2007-11-29 20:24:53 +0300
commit364e3a3d8a26aae058cf1c257457ad1f6b1cfe4c (patch)
tree1e5c04d671c5fcc8c2e25c37b08378e95a7e4e11
parent9fc89c2dea7ca7915e6606e49167cdca2f3c4e30 (diff)
downloadlinux-364e3a3d8a26aae058cf1c257457ad1f6b1cfe4c.tar.xz
uml: fix !NO_HZ busy-loop
With NO_HZ disabled, the UML idle loop effectively becomes a busy loop, as it will sleep for no time. The cause was forgetting to restart the tick after waking up from sleep. It was disabled before sleeping, and the remaining time used as the interval to sleep. So, the tick needs to be restarted when nanosleep finishes. This is done by introducing after_sleep_interval, which is empty in the NO_HZ case, but which sets the tick starting in the !NO_HZ case. Signed-off-by: Jeff Dike <jdike@linux.intel.com> Cc: Miklos Szeredi <miklos@szeredi.hu> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
-rw-r--r--arch/um/os-Linux/time.c54
1 files changed, 51 insertions, 3 deletions
diff --git a/arch/um/os-Linux/time.c b/arch/um/os-Linux/time.c
index e34e1effe0f5..ef02d941c2ad 100644
--- a/arch/um/os-Linux/time.c
+++ b/arch/um/os-Linux/time.c
@@ -59,7 +59,7 @@ long long disable_timer(void)
{
struct itimerval time = ((struct itimerval) { { 0, 0 }, { 0, 0 } });
- if(setitimer(ITIMER_VIRTUAL, &time, &time) < 0)
+ if (setitimer(ITIMER_VIRTUAL, &time, &time) < 0)
printk(UM_KERN_ERR "disable_timer - setitimer failed, "
"errno = %d\n", errno);
@@ -74,13 +74,61 @@ long long os_nsecs(void)
return timeval_to_ns(&tv);
}
+#ifdef UML_CONFIG_NO_HZ
+static int after_sleep_interval(struct timespec *ts)
+{
+}
+#else
+static inline long long timespec_to_us(const struct timespec *ts)
+{
+ return ((long long) ts->tv_sec * UM_USEC_PER_SEC) +
+ ts->tv_nsec / UM_NSEC_PER_USEC;
+}
+
+static int after_sleep_interval(struct timespec *ts)
+{
+ int usec = UM_USEC_PER_SEC / UM_HZ;
+ long long start_usecs = timespec_to_us(ts);
+ struct timeval tv;
+ struct itimerval interval;
+
+ /*
+ * It seems that rounding can increase the value returned from
+ * setitimer to larger than the one passed in. Over time,
+ * this will cause the remaining time to be greater than the
+ * tick interval. If this happens, then just reduce the first
+ * tick to the interval value.
+ */
+ if (start_usecs > usec)
+ start_usecs = usec;
+ tv = ((struct timeval) { .tv_sec = start_usecs / UM_USEC_PER_SEC,
+ .tv_usec = start_usecs % UM_USEC_PER_SEC });
+ interval = ((struct itimerval) { { 0, usec }, tv });
+
+ if (setitimer(ITIMER_VIRTUAL, &interval, NULL) == -1)
+ return -errno;
+
+ return 0;
+}
+#endif
+
extern void alarm_handler(int sig, struct sigcontext *sc);
void idle_sleep(unsigned long long nsecs)
{
- struct timespec ts = { .tv_sec = nsecs / UM_NSEC_PER_SEC,
- .tv_nsec = nsecs % UM_NSEC_PER_SEC };
+ struct timespec ts;
+
+ /*
+ * nsecs can come in as zero, in which case, this starts a
+ * busy loop. To prevent this, reset nsecs to the tick
+ * interval if it is zero.
+ */
+ if (nsecs == 0)
+ nsecs = UM_NSEC_PER_SEC / UM_HZ;
+ ts = ((struct timespec) { .tv_sec = nsecs / UM_NSEC_PER_SEC,
+ .tv_nsec = nsecs % UM_NSEC_PER_SEC });
if (nanosleep(&ts, &ts) == 0)
alarm_handler(SIGVTALRM, NULL);
+ after_sleep_interval(&ts);
}