Re: [RFC][PATCH v2 5/5] sched: User Mode Concurency Groups

On Mon, Jan 24, 2022 at 03:29:56PM +0100, Peter Zijlstra wrote:

On Fri, Jan 21, 2022 at 04:18:46PM +0100, Peter Zijlstra wrote:

quoted

Something like this, still yuck though. Also still need to write me a
test for this.

quoted

--- a/kernel/sched/umcg.c
+++ b/kernel/sched/umcg.c

@@ -232,6 +232,8 @@ static int umcg_update_state(struct task
 /* Called from syscall enter path and exceptions that can schedule */
 void umcg_sys_enter(struct pt_regs *regs, long syscall)
 {
+	current->umcg_timeout = 0;
+
 	/* avoid recursion vs our own syscalls */
 	if (syscall == __NR_umcg_wait ||
 	    syscall == __NR_umcg_ctl)

@@ -519,6 +521,7 @@ void umcg_notify_resume(struct pt_regs *
 	struct umcg_task __user *self = tsk->umcg_task;
 	bool worker = tsk->flags & PF_UMCG_WORKER;
 	u32 state;
+	int ret;
 
 	/* avoid recursion vs schedule() */
 	if (worker)

@@ -554,12 +557,17 @@ void umcg_notify_resume(struct pt_regs *
 		umcg_unpin_pages();
 	}
 
-	switch (umcg_wait(0)) {
+	ret = umcg_wait(tsk->umcg_timeout);

Oh how I hate signals... this can get scribbled by a syscall/fault from
sigcontext :/

Maybe I can recover the timo argument from the original syscall
pt_regs.. let me try.

Urgh, recursive hell... If the signal does *anything* that tickles
notify-resume it'll find RUNNABLE and go wait there --- ad infinitum.

I need to go cook dinner, I'll prod more at this later