On Thu, 2012-10-04 at 11:02 -0400, Steven Rostedt wrote:
> void __init softirq_early_init(void)
> {
> local_irq_lock_init(local_softirq_lock);
> }
>
> Where:
>
> #define local_irq_lock_init(lvar) \
> do { \
> int __cpu; \
> for_each_possible_cpu(__cpu) \
> spin_lock_init(&per_cpu(lvar, __cpu).lock); \
> } while (0)
>
> As the softirq lock is a local_irq_lock, which is a per_cpu lock, the
> initialization is done to all per_cpu versions of the lock. But lets
> look at where the softirq_early_init() is called from.
>
> In init/main.c: start_kernel()
>
> /*
> * Interrupts are still disabled. Do necessary setups, then
> * enable them
> */
> softirq_early_init();
> tick_init();
> boot_cpu_init();
> page_address_init();
> printk(KERN_NOTICE "%s", linux_banner);
> setup_arch(&command_line);
> mm_init_owner(&init_mm, &init_task);
> mm_init_cpumask(&init_mm);
> setup_command_line(command_line);
> setup_nr_cpu_ids();
> setup_per_cpu_areas();
> smp_prepare_boot_cpu(); /* arch-specific boot-cpu hooks */
>
> One of the first things that is called is the initialization of the
> softirq lock. But if you look further down, we see the per_cpu areas
> have not been set up yet. Thus initializing a local_irq_lock() before
> the per_cpu section is set up, may not work as it is initializing the
> per cpu locks before the per cpu exists.
>
> By moving the softirq_early_init() right after setup_per_cpu_areas(),
> the kernel boots fine.
>
I investigated why this still works on x86, and found this. By adding
some printks:
void __init softirq_early_init(void)
{
int __cpu;
printk("init softirq locks\n");
local_irq_lock_init(local_softirq_lock);
printk("list locks\n");
for_each_possible_cpu(__cpu)
printk("local_softirq_lock[%d].node_list=%p\n", __cpu,
per_cpu(local_softirq_lock,__cpu).lock.lock.wait_list.node_list.prev);
}
The output was:
Initializing cgroup subsys cpu
init softirq locks
list locks
Linux version 3.2.30-test-rt45+ (rostedt@goliath) (gcc version 4.6.0 (GCC) )
#262 SMP PREEMPT RT Thu Oct 4 15:48:16 EDT 2012
Command line: ro root=/dev/mapper/VG01-F13x64 rd_LVM_LV=VG01/F13x64 rd_NO_LUKS
rd_NO_MD rd_NO_DM console=ttyS0,115200 ignore_loglevel selinux=0
earlyprintk=ttyS0,115200 ftrace_dump
_on_oops
Note, it printed "list locks" but never printed anything for that loop.
Seems that before the per_cpu area is initialized, the
for_each_possible_cpu() does not execute. To confirm this, I added that
same loop in spawn_ksoftirq() and it shows this:
... fixed-purpose events: 3
... event mask: 0000000700000003
local_softirq_lock[0].node_list= (null)
local_softirq_lock[1].node_list= (null)
local_softirq_lock[2].node_list= (null)
local_softirq_lock[3].node_list= (null)
NMI watchdog enabled, takes one hw-pmu counter.
Booting Node 0, Processors #1
smpboot cpu 1: start_ip = 98000
Yep, the node_list was never initialized.
This doesn't crash x86 because it is saved by:
static inline void init_lists(struct rt_mutex *lock)
{
if (unlikely(!lock->wait_list.node_list.prev))
plist_head_init(&lock->wait_list);
}
and the first time something blocks on the lock, the wait_list is
initialized.
The reason that it crashes on powerpc, is because the
for_each_possible_cpu() actually does loop:
(on powerpc box)
Initializing cgroup subsys cpuset^M
Initializing cgroup subsys cpu
init softirq locks
list locks^M
local_softirq_lock[0].node_list=c000000000781f00
local_softirq_lock[1].node_list=c000000000781f00
Linux version 3.2.30-test-rt45-dirty (rostedt@goliath) (gcc version 4.6.0 (GCC)
) #24 SMP PREEMPT RT Thu Oct 4 15:55:07 EDT 2012^M
[0000] : CF000012^M
The problem is that the per_cpu() returns the same pointer for each CPU
passed to it (as you can see, the node_list pointer is the same). As the
node_list was initialized, but to the wrong pointer, the init_lists()
above will not correct the problem as it did with x86. When the
wait_list starts to be used, it will soon become corrupted.
Moving the init to after the per_cpu setup, I get this:
pcpu-alloc: s84096 r0 d46976 u524288 alloc=1*1048576
pcpu-alloc: [0] 0 1
init softirq locks
list locks
local_softirq_lock[0].node_list=c000000001001f00
local_softirq_lock[1].node_list=c000000001081f00
Built 1 zonelists in Node order, mobility grouping on. Total pages: 16370
As you can see, the node_lists are now unique per_cpu.
-- Steve
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to [email protected]
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/
