Re: [Qemu-devel] chardev's and fd's in monitors

[Daniel P. Berrange]

On Wed, Oct 19, 2016 at 09:12:11AM +0100, Dr. David Alan Gilbert wrote:
> * Markus Armbruster (address@hidden) wrote:
> > "Daniel P. Berrange" <address@hidden> writes:
> > 
> > > On Tue, Oct 18, 2016 at 02:52:13PM +0100, Dr. David Alan Gilbert wrote:
> > >> * Daniel P. Berrange (address@hidden) wrote:
> > >> > On Tue, Oct 18, 2016 at 02:25:25PM +0100, Dr. David Alan Gilbert wrote:
> > >> > > * Daniel P. Berrange (address@hidden) wrote:
> > >> > > > On Tue, Oct 18, 2016 at 12:32:02PM +0100, Dr. David Alan Gilbert 
> > >> > > > wrote:
> > >> > > > > * Daniel P. Berrange (address@hidden) wrote:
> > >> > > > > > On Wed, Oct 12, 2016 at 08:15:02PM +0100, Dr. David Alan 
> > >> > > > > > Gilbert wrote:
> > >> > > > > > > Hi,
> > >> > > > > > >   I had a look at a couple of readline like libraries;
> > >> > > > > > > editline and linenoise.  A difficulty with using them is that
> > >> > > > > > > they both want fd's or FILE*'s; editline takes either but
> > >> > > > > > > from a brief look I think it's expecting to extract the fd.
> > >> > > > > > > That makes them tricky to integrate into qemu, where
> > >> > > > > > > the chardev's hide a whole bunch of non-fd things; in 
> > >> > > > > > > particular
> > >> > > > > > > tls, mux, ringbuffers etc.
> > >> > > > > > > 
> > >> > > > > > > If we could get away with just a FILE* then we could use 
> > >> > > > > > > fopencookie,
> > >> > > > > > > but that's GNU only.
> > >> > > > > > > 
> > >> > > > > > > Is there any sane way of shepherding all chardev's into 
> > >> > > > > > > having an
> > >> > > > > > > fd?
> > >> > > > > > 
> > >> > > > > > The entire chardev abstraction model exists precisely because 
> > >> > > > > > we cannot
> > >> > > > > > make all chardevs look like a single fd. Even those which are 
> > >> > > > > > fd based
> > >> > > > > > may have separate FDs for input and output.
> > >> > > > > 
> > >> > > > > Note that editline takes separate in/out streams, but it does 
> > >> > > > > want those streams
> > >> > > > > to be FILE*'s.
> > >> > > > > 
> > >> > > > > > IMHO the only viable approach would be to enhance 
> > >> > > > > > linenoise/editline to
> > >> > > > > > not assume use of fd* or FILE * abstractions.
> > >> > > > > 
> > >> > > > > I think if it came to that then we'd probably end up sticking 
> > >> > > > > with what we
> > >> > > > > had for a very long time; I'd assume it would take a long time 
> > >> > > > > before
> > >> > > > > any mods we made to the libraries would come around to be 
> > >> > > > > generally useful.
> > >> > > > > 
> > >> > > > > > BTW, what is the actual thread issue you are facing ? Chardevs 
> > >> > > > > > at least
> > >> > > > > > ought to be usable from a separate thread, as long as each 
> > >> > > > > > distinct
> > >> > > > > > chardev object instance was only used from one thread at a 
> > >> > > > > > time ?
> > >> > > > > 
> > >> > > > > Marc-André pointed that out; I hadn't realised they were thread 
> > >> > > > > safe.
> > >> > > > > But what are the rules? You say 'only used from one thread at a 
> > >> > > > > time' -
> > >> > > > > what happens if we have a mux and the different streams to the 
> > >> > > > > mux come
> > >> > > > > from different threads?
> > >> > > > 
> > >> > > > Well there is no mutex locking on the CharDriverState objects, so 
> > >> > > > the
> > >> > > > exact rule is "you mustn't do anything from multiple threads that 
> > >> > > > will
> > >> > > > race on contents of CharDriverState". That's too fuzzy to be 
> > >> > > > useful to
> > >> > > > developers though, so I think the only sensible option right now 
> > >> > > > is to
> > >> > > > say any "top level" CharDriverState should only be touch from one 
> > >> > > > thread
> > >> > > > at a time. IOW, if you have a mux, that that rule would apply to 
> > >> > > > the
> > >> > > > mux itself and the various children it owns as if they were a 
> > >> > > > single
> > >> > > > unnit.
> > >> > > 
> > >> > > OK; I think we're probably saved by the big lock at the moment, so 
> > >> > > that
> > >> > > all device emulation that outputs text is probably holding it and 
> > >> > > the monitor
> > >> > > is also.  What about something like an error_report from a different 
> > >> > > thread
> > >> > > while something is happening in the monitor?
> > >> > 
> > >> > If we moved execution of monitor commands to separate thread from the
> > >> > thread handling monitor I/O, then we'd have to modify error_report so
> > >> > that it queued the text in some manner, such that it was only then
> > >> > fed back to the client once the command thread completed. Alternatively
> > >> > we'd have to introduced locking in the Monitor object, that serialized
> > >> > access to the underling CharDriverState I/O funcs.
> > >> 
> > >> I already use error_report's in places in migration threads of various
> > >> types; I'm not sure if that's a problem.
> > >
> > > Unless those places are protected by the big qemu lock, that sounds
> > > not good. error_report calls into error_vprintf which checks the
> > > 'cur_mon' global "Monitor" pointer. This variable is updated at
> > > runtime - eg in qmp_human_monitor_command(), monitor_qmp_read(),
> > > monitor_read(), etc. So if migration threads outside the BQL are
> > > calling error_report() that could well cause problems. If you
> > > are lucky messages will merely end up going to stderr instead of
> > > the monitor, but in worst case I wouldn't be surprised if there
> > > is a crash possibility in some race conditions.
> > 
> > cur_mon dates back to single-threaded times.
> > 
> > The idea is to print to the monitor when running within an HMP command,
> > else to stderr.
> > 
> > The current solution is to set cur_mon around monitor commands.  Fine
> > with a single thread, not fine at all with multiple threads.
> > 
> > Making cur_mon thread-local should fix things.
> > 
> > If you do want to report errors from another thread in a monitor, you
> > should use error_setg() & friends to get them into the monitor, in my
> > opinion.  Asynchronously barfing output to a monitor doesn't strike me
> > as a sensible design.  Not least because it doesn't work at all with
> > QMP!  If an error message is important enough for the human monitor's
> > user to make use route it to the human monitor, why is hiding it from
> > the QMP client okay?
> > 
> > If I'm wrong and it is sensible, we need locking.
> 
> The difficulty is that we've long tried to be consistent and use error_report
> rather than fprintf's; now that is turning out to be wrong if we're in
> other threads.  It's even trickier for the cases of routines that might
> be called in either the main thread or another thread - we have no
> right answer as to how they should print na error.

Pretty much all code should be using error_setg together with an Error **errp
parameter to the method. The only places that should use error_report are at
top of the call chain where they unambigously know that the printed result
is going to the right place.

Historically the migration hasn't use error_setg because where was no
mechanism to feed the Error **errp object back to the caller, due to the
"migrate" QMP command always being in the background. Similarly in the
incoming migration side, we've always wanted errors to go to stderr.

I fixed the outgoing side in

  commit d59ce6f34434bf47a9b26138c908650bf9a24be1
  Author: Daniel P. Berrange <address@hidden>
  Date:   Wed Apr 27 11:05:00 2016 +0100

    migration: add reporting of errors for outgoing migration

So we could useful convert the migration code to use error_setg everywhere
and then use error_report_err() on the incoming side to spit that Error
onto stderr.

The block layer has almost no use of Error **errp parameters either and
this has long been an issue preventing useful error reporting there.


Regards,
Daniel
-- 
|: http://berrange.com      -o-    http://www.flickr.com/photos/dberrange/ :|
|: http://libvirt.org              -o-             http://virt-manager.org :|
|: http://entangle-photo.org       -o-    http://search.cpan.org/~danberr/ :|