Re: [Qemu-devel] Linux SIGIO handling changes

[Anthony Liguori <anthony@codemonkey.ws>]

Avi Kivity wrote:
> Anthony Liguori wrote:
>> Anders Melchiorsen wrote:
>>> I am resending this patch, split into two parts now.
>>>
>>> This cleans up SIGIO handling to improve latency:
>>> - SIGALRM for alarm timers
>>> - enable SIGIO on qemu_set_fd_handler2()
>>>
>>> The issue was found in KVM, where it is much more visible,
>>> because there is no periodic timer. However, it has been
>>> confirmed (by Aurelien Jarno) that even for qemu, this
>>> approach "improves network transfers in a huge way".
>>>
>>> Please apply, or give a firm rejection so I can stop resending.
>>>   
>>
>> Probably the right thing to do is the direction KVM is moving toward, 
>> i.e. have a separate IO thread.
>>
>> Setting SIGIO on every file descriptor is really just a hack to break 
>> out of the cpu exec loop.  It's unclear to me whether it's really 
>> always the right thing to do for every file descriptor.
>>
>
> Even with a separate iothread one needs the signals, as there is no 
> other race-free way to poll for both aio completions and fd readiness.

Unless you emulate signalfd() using a thread.  FWIW, I've been thinking 
of implementing something similar to posix-aio (using a thread-pool for 
AIO) for QEMU to get a bit more control for this sort of thing.

posix-aio is a pretty unfortunate interface as it doesn't provide a 
mechanism to do asynchronous fdatasync() nor individual vector requests.

Regards,

Anthony Liguori

Regards,

Anthony Liguori

> [well, pselect works, but I dislike it and it isn't present on all 
> kernels]
>