[Spice-devel] [RFC] server: add main_dispatcher

[Hans de Goede]

Hi,

On 09/08/2011 11:18 AM, Alon Levy wrote:
> On Thu, Sep 08, 2011 at 10:58:41AM +0200, Hans de Goede wrote:
>> Hi,
>>
>> On 09/08/2011 09:31 AM, Gerd Hoffmann wrote:
>>> Hi,
>>>
>>>>> simple object system, not unions). Just an idea, could be changed
>>>>> later perhaps. socketpair is probably fine.
>>>>
>>>> First of all I see I forgot to add a mutex, since I said "any non main thread", right
>>>> now we only have one, so no need to lock, but the idea is that it can also be called from
>>>> others, and so we will need it later. (unless that write can be considered atomic, but
>>>> I don't think it can)
>>>
>>> A single write system call is atomic.
>>
>> Yes and no, this is true, except when the write blocks after writing part of the
>> message, once that happens all bets are off. I think we may have several cases
>> where we don't check for this. At the minimum we should assert (and not ASSERT)
>> when a write returns a wrong number of bytes written (iow it managed
>> to write some but not all of the message).
>
> Can that happen in a normal order of events? or is that a thing that really merits
> crashing the process?

Define normal, I think the kernel will buffer up to 16k (or maybe only 4k ?) before blocking,
so if the producing side manages to write 16k before the consumer gets around to consuming
any you will hit this. This seems an exceptional circumstance, so I think crashing is an
ok answer to this.

We should either handle this gracefully (which means extra code + locking for something
which should never happen), or at least catch the should never happen case.

Regards,

Hans