Re: [Qemu-devel] coroutine-ucontext broken for x86-32

[Jan Kiszka <jan.kiszka@siemens.com>]

On 2012-05-09 14:31, Anthony Liguori wrote:
> On 05/09/2012 12:25 PM, Jan Kiszka wrote:
>> On 2012-05-09 14:17, Anthony Liguori wrote:
>>> On 05/09/2012 06:38 AM, Jan Kiszka wrote:
>>>> On 2012-05-09 08:15, Peter Maydell wrote:
>>>>> On 9 May 2012 11:11, Kevin Wolf<kwolf@redhat.com>   wrote:
>>>>>> Am 08.05.2012 21:35, schrieb Jan Kiszka:
>>>>>>> I hunted down a fairly subtle corruption of the VCPU thread signal mask
>>>>>>> in KVM mode when using the ucontext version of coroutines:
>>>>>>>
>>>>>>> coroutine_new calls getcontext, makecontext, swapcontext. Those
>>>>>>> functions get/set also the signal mask of the caller. Unfortunately,
>>>>>>> they only use the sigprocmask syscall on i386, not the rt_sigprocmask
>>>>>>> version. So they do not properly save/restore the blocked RT signals,
>>>>>>> namely our SIG_IPI - it becomes unblocke this way.
>>>>>>
>>>>>> If other coroutine backends work (sigaltstack?), we could try to detect
>>>>>> the situation in configure and set the right default. Not sure what the
>>>>>> condition is, glibc + i386?
>>>>>
>>>>> I don't think you can do a compile-time test for this short of
>>>>> just disabling use of the ucontext code on all i386/Linux platforms.
>>>>>
>>>>> I think it's becoming increasingly obvious that the setcontext/getcontext
>>>>> code path is not very well used and prone to nasty libc bugs. Trying
>>>>> to implement coroutines in C is just a really bad idea and I think
>>>>> we should be trying to reduce our use of them if we possibly can,
>>>>> presumably by switching to actually using threads where we really
>>>>> need the parallelism.
>>>>
>>>> I tend to agree.
>>>>
>>>> FWIW, sigaltstack works around the issue here, but I'm still looking s
>>>> bit skeptical at its implementation.
>>>
>>> Is there any downside to using SIGUSR1?
>>
>> You mean for SIG_IPI? I don't think so. But the point is that the, well,
>> limitation of ucontext will continue to break RT signals,
> 
> Yes, but we currently don't use RT signals, right?  So we could switch to 
> SIGUSR1, fix the problem in glibc, and call it a day, no?

That cures the current symptom but does not prevent future diseases
around RT signals. I would prefer to disable ucontext usage on those
platforms we identified as broken.

BTW, I'm starting to believe it's not a glibc but rather a Linux kernel
issue, only biting us on 32/64. sigprocmask should only manipulate those
signals, its masks can address. Digging deeper...

Jan

-- 
Siemens AG, Corporate Technology, CT T DE IT 1
Corporate Competence Center Embedded Linux