On Thu, 16 Mar 2023 at 13:31, Marco Elver elver@google.com wrote:
From: Dmitry Vyukov dvyukov@google.com
POSIX timers using the CLOCK_PROCESS_CPUTIME_ID clock prefer the main thread of a thread group for signal delivery. However, this has a significant downside: it requires waking up a potentially idle thread.
Instead, prefer to deliver signals to the current thread (in the same thread group) if SIGEV_THREAD_ID is not set by the user. This does not change guaranteed semantics, since POSIX process CPU time timers have never guaranteed that signal delivery is to a specific thread (without SIGEV_THREAD_ID set).
The effect is that we no longer wake up potentially idle threads, and the kernel is no longer biased towards delivering the timer signal to any particular thread (which better distributes the timer signals esp. when multiple timers fire concurrently).
Signed-off-by: Dmitry Vyukov dvyukov@google.com Suggested-by: Oleg Nesterov oleg@redhat.com Reviewed-by: Oleg Nesterov oleg@redhat.com Signed-off-by: Marco Elver elver@google.com
v6:
- Split test from this patch.
- Update wording on what this patch aims to improve.
v5:
- Rebased onto v6.2.
v4:
- Restructured checks in send_sigqueue() as suggested.
v3:
- Switched to the completely different implementation (much simpler) based on the Oleg's idea.
RFC v2:
- Added additional Cc as Thomas asked.
kernel/signal.c | 25 ++++++++++++++++++++++--- 1 file changed, 22 insertions(+), 3 deletions(-)
diff --git a/kernel/signal.c b/kernel/signal.c index 8cb28f1df294..605445fa27d4 100644 --- a/kernel/signal.c +++ b/kernel/signal.c @@ -1003,8 +1003,7 @@ static void complete_signal(int sig, struct task_struct *p, enum pid_type type) /* * Now find a thread we can wake up to take the signal off the queue. *
* If the main thread wants the signal, it gets first crack.* Probably the least surprising to the average bear.
* Try the suggested task first (may or may not be the main thread). */ if (wants_signal(sig, p)) t = p;@@ -1970,8 +1969,23 @@ int send_sigqueue(struct sigqueue *q, struct pid *pid, enum pid_type type)
ret = -1; rcu_read_lock();
/** This function is used by POSIX timers to deliver a timer signal.* Where type is PIDTYPE_PID (such as for timers with SIGEV_THREAD_ID* set), the signal must be delivered to the specific thread (queues* into t->pending).** Where type is not PIDTYPE_PID, signals must just be delivered to the* current process. In this case, prefer to deliver to current if it is* in the same thread group as the target, as it avoids unnecessarily* waking up a potentially idle task.*/ t = pid_task(pid, type);
if (!t || !likely(lock_task_sighand(t, &flags)))
if (!t)goto ret;if (type != PIDTYPE_PID && same_thread_group(t, current))t = current;if (!likely(lock_task_sighand(t, &flags))) goto ret; ret = 1; /* the signal is ignored */@@ -1993,6 +2007,11 @@ int send_sigqueue(struct sigqueue *q, struct pid *pid, enum pid_type type) q->info.si_overrun = 0;
signalfd_notify(t, sig);
/** If the type is not PIDTYPE_PID, we just use shared_pending, which* won't guarantee that the specified task will receive the signal, but* is sufficient if t==current in the common case.*/ pending = (type != PIDTYPE_PID) ? &t->signal->shared_pending : &t->pending; list_add_tail(&q->list, &pending->list); sigaddset(&pending->signal, sig);--
One last semi-gentle ping. ;-)
1. We're seeing that in some applications that use POSIX timers heavily, but where the main thread is mostly idle, the main thread receives a disproportional amount of the signals along with being woken up constantly. This is bad, because the main thread usually waits with the help of a futex or really long sleeps. Now the main thread will steal time (to go back to sleep) from another thread that could have instead just proceeded with whatever it was doing.
2. Delivering signals to random threads is currently way too expensive. We need to resort to this crazy algorithm: 1) receive timer signal, 2) check if main thread, 3) if main thread (which is likely), pick a random thread and do tgkill. To find a random thread, iterate /proc/self/task, but that's just abysmal for various reasons. Other alternatives, like inherited task clock perf events are too expensive as soon as we need to enable/disable the timers (does IPIs), and maintaining O(#threads) timers is just as horrible.
This patch solves both the above issues.
We acknowledge the unfortunate situation of attributing this patch to one clear subsystem and owner: it straddles into signal delivery and POSIX timers territory, and perhaps some scheduling. The patch itself only touches kernel/signal.c.
If anyone has serious objections, please shout (soon'ish). Given the patch has been reviewed by Oleg, and scrutinized by Dmitry and myself, presumably we need to find a tree that currently takes kernel/signal.c patches?
Thanks!
-- Marco