Migration of Standard Timers
Subsystems (2)
PATCH 1/1 Migration of Standard Timers
I'm reposting this with some suggested changes.
This patch allows the user to migrate currently queued standard timers from one cpu to another. Migrating timers off of select cpus allows those cpus to run time critical threads with minimal timer induced latency (which can reach 100's of usec for a single timer as shown on an X86_64 test machine), thereby improving overall determinance on those selected cpus.
This patch considers timers placed with add_timer_on() to have 'cpu affinity' and does not move them, unless the timers are being migrated off of a hotplug cpu that is going down.
The changes in drivers/base/cpu.c provide a clean and convenient interface for triggering the migration through sysfs, via writing the destination cpu number to an owner writeable file (0200 permissions) associated with the source cpu. In additon, this functionality is available for kernel module use.
Note that migrating timers will not, by itself, keep new timers off of the chosen cpu. But with careful control of thread affinity, one can control the affinity of new timers and keep timer induced latencies off of the chosen cpu.
This particular patch does not affect the hrtimers. That could be addressed later.
kernel/timer.c +42 -25
@@ -147,6 +147,7 @@ void fastcall init_timer(struct timer_li { timer->entry.next =NULL ; timer->base =__raw_get_cpu_var (tvec_bases); + timer->is_bound_to_cpu =0 ; }EXPORT_SYMBOL (init_timer);
@@ -250,6 +251,7 @@ void add_timer_on(struct timer_list *timBUG_ON (timer_pending (timer) || !timer->function);spin_lock_irqsave (&base->lock, flags); timer->base = base; + timer->is_bound_to_cpu =1 ;/* Don't migrate if cpu online */ internal_add_timer (base, timer);spin_unlock_irqrestore (&base->lock, flags); }
@@ -1616,50 +1618,65 @@ static int __devinit init_timers_cpu(intreturn 0 ; } -#ifdef CONFIG_HOTPLUG_CPU -static void migrate_timer_list (tvec_base_t *new_base,struct list_head *head) +static void migrate_timer_list (tvec_base_t *new_base,struct list_head *head,int cpu_down) { -struct timer_list *timer; +struct timer_list *timer, *t; -while (!list_empty (head)) { - timer = list_entry(head->next, struct timer_list, entry); +list_for_each_entry_safe (timer, t, head, entry) { +if (!cpu_down && timer->is_bound_to_cpu) +continue ;detach_timer (timer,0 ); timer->base = new_base;internal_add_timer (new_base, timer); } } -static void __devinit migrate_timers (int cpu) +int migrate_timers (int dest_cpu,int source_cpu,int cpu_down) {tvec_base_t *old_base;tvec_base_t *new_base; +spinlock_t *lock1, *lock2; +unsigned long flags;int i; -BUG_ON (cpu_online (cpu)); - old_base =per_cpu (tvec_bases, cpu); - new_base =get_cpu_var (tvec_bases); - -local_irq_disable (); -spin_lock (&new_base->lock); -spin_lock (&old_base->lock); +if (source_cpu == dest_cpu) +return -EINVAL ; + +if (!cpu_online (dest_cpu)) +return -EINVAL ; + +if (cpu_down) +BUG_ON (cpu_online (source_cpu)); +else if (!cpu_online (source_cpu)) +return -EINVAL ; + + old_base =per_cpu (tvec_bases, source_cpu); + new_base =per_cpu (tvec_bases, dest_cpu); -BUG_ON (old_base->running_timer); +/* Order locking based on relative cpu number */ + lock1 = dest_cpu > source_cpu ? &old_base->lock : &new_base->lock; + lock2 = dest_cpu > source_cpu ? &new_base->lock : &old_base->lock; + +spin_lock_irqsave (lock1, flags); +spin_lock (lock2); + +if (cpu_down) +BUG_ON (old_base->running_timer);for (i =0 ; i <TVR_SIZE ; i++) -migrate_timer_list (new_base, old_base->tv1.vec + i); +migrate_timer_list (new_base, old_base->tv1.vec + i, cpu_down);for (i =0 ; i <TVN_SIZE ; i++) { -migrate_timer_list (new_base, old_base->tv2.vec + i); -migrate_timer_list (new_base, old_base->tv3.vec + i); -migrate_timer_list (new_base, old_base->tv4.vec + i); -migrate_timer_list (new_base, old_base->tv5.vec + i); +migrate_timer_list (new_base, old_base->tv2.vec + i, cpu_down); +migrate_timer_list (new_base, old_base->tv3.vec + i, cpu_down); +migrate_timer_list (new_base, old_base->tv4.vec + i, cpu_down); +migrate_timer_list (new_base, old_base->tv5.vec + i, cpu_down); } -spin_unlock (&old_base->lock); -spin_unlock (&new_base->lock); -local_irq_enable (); -put_cpu_var (tvec_bases); +spin_unlock (lock2); +spin_unlock_irqrestore (lock1, flags); +return 0 ; } -#endif /* CONFIG_HOTPLUG_CPU */ +EXPORT_SYMBOL_GPL (migrate_timers);static int __cpuinit timer_cpu_notify (struct notifier_block *self,unsigned long action,void *hcpu)
@@ -1672,7 +1689,7 @@ static int __cpuinit timer_cpu_notify(stbreak ;#ifdef CONFIG_HOTPLUG_CPU case CPU_DEAD : -migrate_timers (cpu); +migrate_timers (smp_processor_id (), cpu,1 );break ;#endif default :
drivers/base/cpu.c +23 -0
@@ -54,6 +54,25 @@ static ssize_t store_online(struct sys_d } static SYSDEV_ATTR(online, 0600, show_online, store_online); +static ssize_t store_timer_migrate (struct sys_device *dev,const char *buf, +size_t count) +{ +unsigned int source_cpu = dev->id; +unsigned long dest_cpu; +int rc; + + dest_cpu =simple_strtoul (buf,NULL ,10 ); +if (dest_cpu >INT_MAX ) +return -EINVAL ; + + rc =migrate_timers (dest_cpu, source_cpu,0 ); +if (rc <0 ) +return rc; +else +return count; +} +static SYSDEV_ATTR(timer_migrate, 0200, NULL, store_timer_migrate); +static void __devinit register_cpu_control (struct cpu *cpu) {sysdev_create_file (&cpu->sysdev, &attr_online);
@@ -62,6 +81,8 @@ void unregister_cpu(struct cpu *cpu) {int logical_cpu = cpu->sysdev.id; +sysdev_remove_file (&cpu->sysdev, &attr_timer_migrate); +unregister_cpu_under_node (logical_cpu,cpu_to_node (logical_cpu));sysdev_remove_file (&cpu->sysdev, &attr_online);
@@ -124,6 +145,8 @@ int __devinit register_cpu(struct cpu *c cpu_sys_devices[num] = &cpu->sysdev;if (!error)register_cpu_under_node (num,cpu_to_node (num)); +if (!error) +sysdev_create_file (&cpu->sysdev, &attr_timer_migrate);#ifdef CONFIG_KEXEC if (!error)
include/linux/timer.h +4 -0
@@ -15,6 +15,8 @@ struct timer_list {unsigned long data;struct tvec_t_base_s *base; + +char is_bound_to_cpu; };extern struct tvec_t_base_s boot_tvec_bases;
@@ -24,6 +26,7 @@ extern struct tvec_t_base_s boot_tvec_ba .expires = (_expires), \ .data = (_data), \ .base = &boot_tvec_bases, \ + .is_bound_to_cpu = 0, \ }#define DEFINE_TIMER (_name, _function, _expires, _data) \
@@ -95,6 +98,7 @@ static inline void add_timer(struct timeextern void init_timers (void );extern void run_local_timers (void ); +extern int migrate_timers (int dest_cpu,int source_cpu,int cpu_down);struct hrtimer ;extern int it_real_fn (struct hrtimer *);
Quoted 7 lines · Dimitri Sivanich
This patch allows the user to migrate currently queued standard timers from one cpu to another. Migrating timers off of select cpus allows those cpus to run time critical threads with minimal timer induced latency (which can reach 100's of usec for a single timer as shown on an
X86_64test machine), thereby improving overall determinance on those selected cpus.
Which driver or subsystem is doing 100s of usecs of work in a timer? Shouldn't another mechanism like a workqueue be used instead?
Lee
Quoted 1 line · Lee Revell
Which driver or subsystem is doing 100s of usecs of work in a timer?
The longest one I've captured so far results from:
rsp rip Function (args)
======================= <nmi>0xffff810257822fd8 0xffffffff803a0e94 rt_check_expire+0x8c
======================= <interrupt>0xffff81025781fee8 0xffffffff803a0e08 rt_check_expire 0xffff81025781ff08 0xffffffff802386b3 run_timer_softirq+0x133 0xffff81025781ff38 0xffffffff80235262 __do_softirq+0x5e 0xffff81025781ff68 0xffffffff8020a958 call_softirq+0x1c 0xffff81025781ff80 0xffffffff8020bea7 do_softirq+0x2c 0xffff81025781ff90 0xffffffff80235142 irq_exit+0x48
Quoted 1 line
Shouldn't another mechanism like a workqueue be used instead?
Not quite sure what you're asking here.
Quoted 5 lines · Dimitri Sivanich
Quoted 1 line · Lee Revell
Which driver or subsystem is doing 100s of usecs of work in a timer?
The longest one I've captured so far results from:
Quoted 11 lines
rsp rip Function (args)
======================= <nmi>0xffff810257822fd80xffffffff803a0e94rt_check_expire+0x8c
======================= <interrupt>0xffff81025781fee80xffffffff803a0e08rt_check_expire0xffff81025781ff080xffffffff802386b3run_timer_softirq+0x1330xffff81025781ff380xffffffff80235262__do_softirq+0x5e0xffff81025781ff680xffffffff8020a958call_softirq+0x1c0xffff81025781ff800xffffffff8020bea7do_softirq+0x2c0xffff81025781ff900xffffffff80235142irq_exit+0x48
Ah, I remember that one. Eric Dumazet had some suggestions to fix it 6-12 months ago which never went anywhere - the thread was called "RCU latency regression in 2.6.16-rc1".
That one is especially annoying as there's no workaround (shrinking the route cache or reducing the GC interval via net.ipv4.route.* sysctls has no effect)
Lee
Quoted 20 lines · On Tue, 19 Sep 2006 10:29:42 -0500 Dimitri Sivanich <[email protected]> wrote:
This patch allows the user to migrate currently queued standard timers from one cpu to another. Migrating timers off of select cpus allows those cpus to run time critical threads with minimal timer induced latency (which can reach 100's of usec for a single timer as shown on an
X86_64test machine), thereby improving overall determinance on those selected cpus.This patch considers timers placed with
add_timer_on()to have 'cpu affinity' and does not move them, unless the timers are being migrated off of a hotplug cpu that is going down.The changes in
drivers/base/cpu.cprovide a clean and convenient interface for triggering the migration through sysfs, via writing the destination cpu number to an owner writeable file (0200 permissions) associated with the source cpu. In additon, this functionality is available for kernel module use.
Quoted 7 lines
Note that migrating timers will not, by itself, keep new timers off of the chosen cpu. But with careful control of thread affinity, one can control the affinity of new timers and keep timer induced latencies off of the chosen cpu.
This particular patch does not affect the hrtimers. That could be addressed later.
I can't say I like this, sorry.
- It adds another word to the timer_list structure for a very obscure application. And a lot of kernel data structures aggregate timer_lists.
- There are places in the kernel which assume that once a timer is added on a CPU, it will stay there. The timer handler re-arms the timer, confident in the knowledge that everything stays on this CPU.
I recall working on such code a couple of years ago, but I now forget where it was.
It is reasonable to do add_timer() within the CPU_ONLINE handler (for example), in the expectation that that timer will fire on this CPU.
The proposed change permits the administrator to break that assumption. We would need to hunt down any code which makes that assumption and convert it to add_timer_on(). And we'd need to be very vigilant in the future, since people could easily add new code which had the old assumption which worked just fine for them in testing.