From 64c7c8f88559624abdbe12b5da6502e8879f8d28 Mon Sep 17 00:00:00 2001 From: Nick Piggin Date: Tue, 8 Nov 2005 21:39:04 -0800 Subject: [PATCH] sched: resched and cpu_idle rework Make some changes to the NEED_RESCHED and POLLING_NRFLAG to reduce confusion, and make their semantics rigid. Improves efficiency of resched_task and some cpu_idle routines. * In resched_task: - TIF_NEED_RESCHED is only cleared with the task's runqueue lock held, and as we hold it during resched_task, then there is no need for an atomic test and set there. The only other time this should be set is when the task's quantum expires, in the timer interrupt - this is protected against because the rq lock is irq-safe. - If TIF_NEED_RESCHED is set, then we don't need to do anything. It won't get unset until the task get's schedule()d off. - If we are running on the same CPU as the task we resched, then set TIF_NEED_RESCHED and no further action is required. - If we are running on another CPU, and TIF_POLLING_NRFLAG is *not* set after TIF_NEED_RESCHED has been set, then we need to send an IPI. Using these rules, we are able to remove the test and set operation in resched_task, and make clear the previously vague semantics of POLLING_NRFLAG. * In idle routines: - Enter cpu_idle with preempt disabled. When the need_resched() condition becomes true, explicitly call schedule(). This makes things a bit clearer (IMO), but haven't updated all architectures yet. - Many do a test and clear of TIF_NEED_RESCHED for some reason. According to the resched_task rules, this isn't needed (and actually breaks the assumption that TIF_NEED_RESCHED is only cleared with the runqueue lock held). So remove that. Generally one less locked memory op when switching to the idle thread. - Many idle routines clear TIF_POLLING_NRFLAG, and only set it in the inner most polling idle loops. The above resched_task semantics allow it to be set until before the last time need_resched() is checked before going into a halt requiring interrupt wakeup. Many idle routines simply never enter such a halt, and so POLLING_NRFLAG can be always left set, completely eliminating resched IPIs when rescheduling the idle task. POLLING_NRFLAG width can be increased, to reduce the chance of resched IPIs. Signed-off-by: Nick Piggin Cc: Ingo Molnar Cc: Con Kolivas Signed-off-by: Andrew Morton Signed-off-by: Linus Torvalds --- Documentation/sched-arch.txt | 89 ++++++++++++++++++++++++++++++++++++++++++++ 1 file changed, 89 insertions(+) create mode 100644 Documentation/sched-arch.txt (limited to 'Documentation') diff --git a/Documentation/sched-arch.txt b/Documentation/sched-arch.txt new file mode 100644 index 000000000000..941615a9769b --- /dev/null +++ b/Documentation/sched-arch.txt @@ -0,0 +1,89 @@ + CPU Scheduler implementation hints for architecture specific code + + Nick Piggin, 2005 + +Context switch +============== +1. Runqueue locking +By default, the switch_to arch function is called with the runqueue +locked. This is usually not a problem unless switch_to may need to +take the runqueue lock. This is usually due to a wake up operation in +the context switch. See include/asm-ia64/system.h for an example. + +To request the scheduler call switch_to with the runqueue unlocked, +you must `#define __ARCH_WANT_UNLOCKED_CTXSW` in a header file +(typically the one where switch_to is defined). + +Unlocked context switches introduce only a very minor performance +penalty to the core scheduler implementation in the CONFIG_SMP case. + +2. Interrupt status +By default, the switch_to arch function is called with interrupts +disabled. Interrupts may be enabled over the call if it is likely to +introduce a significant interrupt latency by adding the line +`#define __ARCH_WANT_INTERRUPTS_ON_CTXSW` in the same place as for +unlocked context switches. This define also implies +`__ARCH_WANT_UNLOCKED_CTXSW`. See include/asm-arm/system.h for an +example. + + +CPU idle +======== +Your cpu_idle routines need to obey the following rules: + +1. Preempt should now disabled over idle routines. Should only + be enabled to call schedule() then disabled again. + +2. need_resched/TIF_NEED_RESCHED is only ever set, and will never + be cleared until the running task has called schedule(). Idle + threads need only ever query need_resched, and may never set or + clear it. + +3. When cpu_idle finds (need_resched() == 'true'), it should call + schedule(). It should not call schedule() otherwise. + +4. The only time interrupts need to be disabled when checking + need_resched is if we are about to sleep the processor until + the next interrupt (this doesn't provide any protection of + need_resched, it prevents losing an interrupt). + + 4a. Common problem with this type of sleep appears to be: + local_irq_disable(); + if (!need_resched()) { + local_irq_enable(); + *** resched interrupt arrives here *** + __asm__("sleep until next interrupt"); + } + +5. TIF_POLLING_NRFLAG can be set by idle routines that do not + need an interrupt to wake them up when need_resched goes high. + In other words, they must be periodically polling need_resched, + although it may be reasonable to do some background work or enter + a low CPU priority. + + 5a. If TIF_POLLING_NRFLAG is set, and we do decide to enter + an interrupt sleep, it needs to be cleared then a memory + barrier issued (followed by a test of need_resched with + interrupts disabled, as explained in 3). + +arch/i386/kernel/process.c has examples of both polling and +sleeping idle functions. + + +Possible arch/ problems +======================= + +Possible arch problems I found (and either tried to fix or didn't): + +h8300 - Is such sleeping racy vs interrupts? (See #4a). + The H8/300 manual I found indicates yes, however disabling IRQs + over the sleep mean only NMIs can wake it up, so can't fix easily + without doing spin waiting. + +ia64 - is safe_halt call racy vs interrupts? (does it sleep?) (See #4a) + +sh64 - Is sleeping racy vs interrupts? (See #4a) + +sparc - IRQs on at this point(?), change local_irq_save to _disable. + - TODO: needs secondary CPUs to disable preempt (See #1) + -- cgit v1.2.3-59-g8ed1b