Request for comment for async resolution of DNS names

[Jeremy Allison]

On Fri, Feb 10, 2017 at 11:50:05PM -0800, Matthieu Patou wrote:
> On 02/10/2017 02:30 AM, Stefan Metzmacher wrote:
> > Hi Matthieu,
> >
> >> On 02/09/2017 12:33 AM, Volker Lendecke wrote:
> >>> On Wed, Feb 08, 2017 at 11:10:11AM -0800, Matthieu Patou wrote:
> >>> <snip>
> >>> Can you restructure the code to only call getaddrinfo and nothing else
> >>> in the threaded job.
> >> Please find attached an updated version of the patch, it should do very
> >> minimal things in the thread function.
> >> It seems to work for me both when I have additional records in the SRV
> >> response and when I haven't, valgrind didn't seems to complain and net
> >> joins seems to work.
> >> I need more tests but so
> >>
> >> There was an alternative solution (that I have saved somewhere) where I
> >> cheat with the struct dns_rr_srv structure to pass the struct addrinfo
> >> back to the main thread by using the pointer to struct sockaddr_storage
> >> to store the point to struct addrinfo.
> >> But Jeremy was not very keen on it, so instead here is the variant where
> >> there is a separate state structure for the resolve_one_name function.
> > It seems we already have getaddrinfo_send/recv can't we just use them?
> > There're currently based on fncall_send/recv and are also available
> > in older releases.
> >
> > metze
> >
> I guess this is just the replacement for doing with pthreadpool_tevent +
> getaddrinfo that I have in the patch.
> 
> Jeremy can you confirm before I do the 4th implementation ?

Yep. getaddrinfo_send() uses fncall_send() interally which
wraps the pthreadpool use.