Re: [RFC] MIPS: smp: Sleeping func called from start_secondary()

["Jiaxun Yang" <jiaxun.yang@flygoat.com>]

在2024年8月2日八月 上午10:53，Serge Semin写道：
> Hi Jiaxun
>
> On Fri, Aug 02, 2024 at 05:43:45PM +0800, Jiaxun Yang wrote:
>> 
>> 
>> 在2024年8月2日八月 下午4:45，Serge Semin写道：
>> > Hi folks,
>> >
>> > To debug some another problem I recently enabled the
>> > CONFIG_DEBUG_ATOMIC_SLEEP config for my SoC based on the 2xP5600 cores
>> > with MIPS GIC as IRQ-controller. That caused the next BUG backtrace
>> > started being printed to the system log during the device boot-up:
>> 
>> Hi Serge,
>> 
>> Thanks for reporting the problem!
>> 
>> I actually have a patch lying around somewhere to convert cevt-r4k to use CPUHP
>> interface and avoid requesting interrupt on secondary CPUs.
>> 
>> Will post after getting more platform tests.
>
> Great! Thanks. I'll test it out upon submission.
>
> * Although I doubt that such (presumably) complicated change might be
> considered as a backportable fix. But at least it shall fix the problem for the
> mainline kernel.

It's actually harder than I expected, as managing it with percpu interrupt API would
break system sharing PCI with TI.

Maybe we need to come up a driver handle IRQ MUX.

Avoid calling get_c0_compare_int on secondary CPU init is easy, will post a patch.

Thanks

>
> Thanks,
> -Serge(y)
>
>> 
>> Thanks
>> - Jiaxun
>> 
>> >
>> > [    0.118053] BUG: sleeping function called from invalid context at 
>> > kernel/locking/mutex.c:283
>> > [    0.118062] in_atomic(): 1, irqs_disabled(): 1, non_block: 0, pid: 
>> > 0, name: swapper/1
>> > [    0.118069] preempt_count: 1, expected: 0
>> > [    0.118074] RCU nest depth: 0, expected: 0
>> > [    0.118078] Preemption disabled at:
>> > [    0.118082] [<80105040>] arch_dup_task_struct+0x20/0x118
>> > [    0.118116] CPU: 1 UID: 0 PID: 0 Comm: swapper/1 Not tainted 
>> > 6.11.0-rc1-bt1-00312-gca6f9469050c-dirty #2460
>> > [    0.118134] Stack : 816d0000 801b89bc 81870000 00000004 a02d0c08 
>> > 00000000 84ae7dcc 801b84c0
>> > [    0.118167]         00000000 00000002 00000000 00000000 00000000 
>> > 00000001 84ae7d70 84aaf200
>> > [    0.118195]         00000000 00000000 81115acc 00000000 00000059 
>> > 84ae7bec 00000000 00000000
>> > [    0.118222]         00000000 ffffffff ffffffea 00000020 816d0000 
>> > 00000000 81115acc 00000002
>> > [    0.118250]         00000000 04240021 a02d0c08 00000000 00000000 
>> > 00000000 30400000 ac24242a
>> > [    0.118277]         ...
>> > [    0.118286] Call Trace:
>> > [    0.118289] [<8010af2c>] show_stack+0x88/0x120
>> > [    0.118309] [<80f09dfc>] dump_stack_lvl+0x78/0xb0
>> > [    0.118323] [<801778d0>] __might_resched+0x190/0x1f0
>> > [    0.118349] [<80f10f80>] mutex_lock+0x20/0x74
>> > [    0.118365] [<801c8590>] irq_create_mapping_affinity+0x48/0x104
>> > [    0.118390] [<801121ec>] r4k_clockevent_init+0x58/0x38c
>> > [    0.118402] [<80116658>] start_secondary+0x34/0x134
>> > [    0.118411] [<80f0b1e4>] smp_bootstrap+0x68/0x98
>> > [    0.118424] 
>> >
>> > I managed to investigate the problem a bit and the reason turned out to be
>> > in the irq_create_mapping() method eventually invoked in the framework of the
>> > secondary CPU start-up procedure. The more detailed calls chain is:
>> > start_secondary()
>> > +-> mips_clockevent_init()
>> >     +-> r4k_clockevent_init()
>> >         +-> get_c0_compare_int()
>> >             +-> gic_get_c0_compare_int()
>> >                 +-> irq_create_mapping()
>> >                     +-> irq_create_mapping_affinity()
>> >                         +-> mutex_lock()
>> >                             +-> might_sleep()
>> >                                 +-> might_resched()
>> >                                     +-> __might_resched()
>> >                                         +-> Print the log above
>> >
>> > I currently lack of free time for further investigation in an attempt to
>> > find a reasonable solution. So here is just a report and a short summary
>> > of the problem. Should you have any idea of how to fix the problem I'll be
>> > very glad to test it out on a short notice.
>> >
>> > -Serge(y)
>> 
>> -- 
>> - Jiaxun

-- 
- Jiaxun