softirq performance fixes, cleanups, 2.4.10.
Hello.
I am trying to understand the basics of softirq handling.
It seems to me that ksoftirqd()'s loop can be cleanuped a bit with following (untested) patch on top of 2.4.10-softirq-A7.
It also removes the 'mask' variable in do_softirq().
Oleg
kernel/softirq.c renamed from kernel/softirq.c.orig +7 -12
@@ -85,7 +85,7 @@ {int max_restart =MAX_SOFTIRQ_RESTART ;int cpu =smp_processor_id (); -__u32 pending, mask; +__u32 pending;long flags;if (in_interrupt ())
@@ -98,7 +98,6 @@if (pending) {struct softirq_action *h; - mask = ~pending;local_bh_disable (); restart:/* Reset the pending bitmask before enabling irqs */
@@ -381,26 +380,22 @@#endif current->nice =19 ; -schedule (); -__set_current_state (TASK_INTERRUPTIBLE );ksoftirqd_task (cpu) = current;for (;;) { -back: +schedule (); +__set_current_state (TASK_INTERRUPTIBLE ); +do {do_softirq ();if (current->need_resched) goto preempt; }while (softirq_pending (cpu)); -schedule (); -__set_current_state (TASK_INTERRUPTIBLE ); - } +continue ; preempt: -__set_current_state (TASK_RUNNING ); -schedule (); -__set_current_state (TASK_INTERRUPTIBLE ); - goto back; +__set_current_state (TASK_RUNNING ); + } }static __init int spawn_ksoftirqd (void )
Quoted 12 lines · Oleg Nesterov
kernel/softirq.c renamed from kernel/softirq.c.orig +1
-1@@ -85,7 +85,7 @@ {int max_restart =MAX_SOFTIRQ_RESTART ;int cpu =smp_processor_id (); -__u32 pending, mask; +__u32 pending;long flags;if (in_interrupt ())
kernel/softirq.c line 103
Quoted 7 lines
kernel/softirq.c renamed from kernel/softirq.c.orig +0
-1@@ -98,7 +98,6 @@if (pending) {struct softirq_action *h; - mask = ~pending;local_bh_disable (); restart:/* Reset the pending bitmask before enabling irqs */
correct.
kernel/softirq.c line 385
Quoted 6 lines
kernel/softirq.c renamed from kernel/softirq.c.orig +0
-2@@ -381,26 +380,22 @@#endif current->nice =19 ; -schedule (); -__set_current_state (TASK_INTERRUPTIBLE );
buggy (check cpus_allowed).
Quoted 13 lines
kernel/softirq.c renamed from kernel/softirq.c.orig +3
-4for (;;) { -back: +schedule (); +__set_current_state (TASK_INTERRUPTIBLE ); +do {do_softirq ();if (current->need_resched) goto preempt; }while (softirq_pending (cpu)); -schedule (); -__set_current_state (TASK_INTERRUPTIBLE ); - }
kernel/softirq.c line 399
Quoted 10 lines
kernel/softirq.c renamed from kernel/softirq.c.orig +3
-4+continue ; preempt: -__set_current_state (TASK_RUNNING ); -schedule (); -__set_current_state (TASK_INTERRUPTIBLE ); - goto back; +__set_current_state (TASK_RUNNING ); + } }
you dropped Ingo's optimization (but you resurrected the strictier /proc statistics).
Andrea
Quoted 4 lines · Oleg Nesterov
I am trying to understand the basics of softirq handling.
It seems to me that ksoftirqd()'s loop can be cleanuped a bit with following (untested) patch on top of 2.4.10-softirq-A7.
thanks - the patch looks mostly correct.
Quoted 1 line
It also removes the 'mask' variable in
do_softirq().
(right, applied.)
Review
Quoted 2 lines
(quoted diff) +0
-2- schedule(); - __set_current_state(TASK_INTERRUPTIBLE);
we need a single schedule to migrate the CPU to the right CPU. Btw, it's not a big offense to run on the wrong CPU: do_softirq() will run correctly. But there wont be a ksoftirqd exeuting on the 'right' CPU (for this short amount of time until it schedules once), and we might miss a single instance of softirq processing. So it's cleaner to do a schedule() that guarantees that we migrate to the target CPU.
kernel/softirq.c line 398
Quoted 2 lines
kernel/softirq.c renamed from kernel/softirq.c.orig +2
-0+__set_current_state (TASK_RUNNING ); + }
well, this basically reverts it to the original code and removes the optimization - but i dont have any strong feelings in either way. I think Andrea is right and it's more correct to set it TASK_RUNNING, for task-statistics reasons. Reverted it to the old code.
and in fact while doing that i found another bug in the old code:TASK_INTERRUPTIBLE was set in a non-IRQ-atomic way, so theoretically it could be possible that the compiler reorders the setting of TASK_INTERRUPTIBLE and the test for softirq_pending(), and if an IRQ gets inbetween the we could miss processing softirq for 1/HZ again. Added an optimization barrier(). (SMP-consistency is not needed, since we are per-CPU.)
Ingo