Current sleep services (nanosleep) provide sleep periods very far from the
expectations when scheuling microsecond-scale timers. On our testbed, using
rdtscp() before and after a nanosleep() syscall to measure the effective
elapsed time with a 1us timer, we got ~59us.
Even with larger timeout periods, the difference is still evident (e.g., with a
100us timer, we measured ~158us of elapsed time).
We believe that one of the reasons is the use of the timespec structure, that
needs to be copied for user to kernel and then converted into a single-value
representation.
In our work Metronome (https://dl.acm.org/doi/pdf/10.1145/3386367.3432730) we
had the need for a precise microsecond-granularity sleep service, as
nanosleep() was far from our needs, so we developed hr_sleep(), a new sleep
service. Since the sleep periods needed in our case are small, we don't want
our sleep service to re-schedule a timer in case of a signal interruption, so
it just returns -EINTR to the user. The user must be aware that this is a
best-effort sleep service, so the sleep period specified is an upper-bound of
the effective elapsed time.
We believe this patch can be useful in applications where fine-grained
granularity is requested for small sleep periods, and re-scheduling the timer
in case of a signal is not mandatory.
In the paper previously linked, Section 3.1 provides more details about
hr_sleep and Section 3.3 extensively evaluates hr_sleep() and compares it to
nanosleep(). For a 1us timeout, hr_sleep() elapses ~3.8us in mean vs. the ~59us
of nanosleep().
hr_sleep has been previously submitted at
https://lore.kernel.org/lkml/20210115180733.5663-1-marco.falte...@uniroma2.it/.
This commit answers to the previous feedback in
https://lore.kernel.org/lkml/CALCETrWfnL=3m3nmmhs-a3si5jptsctf6cethvtsdnwa5mh...@mail.gmail.com/
and applies the requested changes.
Signed-off-by: Marco Faltelli
---
arch/x86/entry/syscalls/syscall_64.tbl | 1 +
kernel/time/hrtimer.c | 67 ++
2 files changed, 68 insertions(+)
diff --git a/arch/x86/entry/syscalls/syscall_64.tbl
b/arch/x86/entry/syscalls/syscall_64.tbl
index 7bf01cbe582f..85b14dfa40fb 100644
--- a/arch/x86/entry/syscalls/syscall_64.tbl
+++ b/arch/x86/entry/syscalls/syscall_64.tbl
@@ -364,6 +364,7 @@
440common process_madvise sys_process_madvise
441common epoll_pwait2sys_epoll_pwait2
442common mount_setattr sys_mount_setattr
+443common hr_sleepsys_hr_sleep
#
# Due to a historical design error, certain syscalls are numbered differently
diff --git a/kernel/time/hrtimer.c b/kernel/time/hrtimer.c
index 4a66725b1d4a..887c01392e08 100644
--- a/kernel/time/hrtimer.c
+++ b/kernel/time/hrtimer.c
@@ -2006,6 +2006,73 @@ SYSCALL_DEFINE2(nanosleep_time32, struct old_timespec32
__user *, rqtp,
}
#endif
+
+
+#ifdef CONFIG_64BIT
+
+
+struct control_record {
+ struct task_struct *task;
+ int awake;
+ struct hrtimer hr_timer;
+};
+
+
+static enum hrtimer_restart hr_sleep_callback(struct hrtimer *timer)
+{
+ struct control_record *control;
+ struct task_struct *the_task;
+
+ control = (control_record *)container_of(timer, control_record,
hr_timer);
+ control->awake = 1;
+ the_task = control->task;
+ wake_up_process(the_task);
+
+ return HRTIMER_NORESTART;
+}
+
+
+
+/**
+ * hr_sleep - a high-resolution sleep service for fine-grained timeouts
+ * @nanoseconds: the requested sleep period in nanoseconds
+ *
+ * Returns:
+ * 0 when the sleep request successfully terminated
+ * -EINVAL if a sleep period < 0 is requested
+ * -EINTR if a signal interrupted the calling thread
+ */
+SYSCALL_DEFINE1(hr_sleep, long, nanoseconds)
+{
+ DECLARE_WAIT_QUEUE_HEAD(the_queue);
+ struct control_record control;
+ ktime_t ktime_interval;
+ struct restart_block *restart;
+
+ if (nanoseconds < 0)
+ return -EINVAL;
+
+ if (nanoseconds == 0)
+ return 0;
+
+ ktime_interval = ktime_set(0, nanoseconds);
+ hrtimer_init(&(control.hr_timer), CLOCK_MONOTONIC, HRTIMER_MODE_REL);
+ control.hr_timer.function = _sleep_callback;
+ control.task = current;
+ control.awake = 0;
+ hrtimer_start(&(control.hr_timer), ktime_interval, HRTIMER_MODE_REL);
+ wait_event_interruptible(the_queue, control.awake == 1);
+ hrtimer_cancel(&(control.hr_timer));
+ if (control.awake == 0)
+ //We have been interrupted by a signal
+ return -EINTR;
+ return 0;
+
+}
+
+#endif
+
+
/*
* Functions related to boot-time initialization:
*/
--
2.25.1