Re: [Qemu-devel] [PATCH v2] block: avoid SIGUSR2

[Paolo Bonzini]

On 10/28/2011 02:31 PM, Stefan Hajnoczi wrote:
> Zhi Yong: We were just talking about converting devices to aio.  If
> you have time to do that for fdc, sd, or any other synchronous API
> users in hw/ that would be helpful.  Please let us know which device
> you are refactoring so we don't duplicate work.

The problem is not really fdc or sd themselves, but whoever uses the 
result of the synchronous reads---respectively DMA and the SD clients.

Some SD clients talk to the SD card in a relatively confined way and 
have interrupts that they set when the operation is done, so these 
confined parts that talk to the card could be changed to a coroutine and 
locked with a CoMutex.  However, not even all of these can do it (in 
particular I'm not sure about ssi-sd.c cannot).

I'm thinking that the problem with the floppy is really that it mixes 
synchronous and asynchronous parts.  As long as you're entirely 
synchronous you should not have any problem, but as soon as you add 
asynchronicity (via bottom halves) you now have to deal with reentrancy.

"git grep _bh hw/" suggests that this should not be a huge problem; most 
if not all occurrences are related to ptimers, or are in entirely 
asynchronous code (IDE, SCSI, virtio).  Floppy+DMA seems to be the only 
problematic occurrence, and any fix (switch to timers, drop idle BH in 
qemu_aio_wait, reschedule if DMA reenters during I/O, drop BH completely 
and just loop) is as good as the others.

(Actually, another one worth checking is ATAPI, but I don't know the 
code and the standards well enough).

Paolo