Re: [Qemu-devel] [PATCH 17/34] qerror: drop QERR_SOCKET_CONNECT_IN_PROGRESS

[Luiz Capitulino <lcapitulino@redhat.com>]

On Fri, 3 Aug 2012 13:26:27 -0500
Michael Roth <mdroth@linux.vnet.ibm.com> wrote:

> On Thu, Aug 02, 2012 at 02:08:48PM -0300, Luiz Capitulino wrote:
> > On Thu, 2 Aug 2012 11:54:11 -0500
> > Michael Roth <mdroth@linux.vnet.ibm.com> wrote:
> > 
> > > On Wed, Aug 01, 2012 at 10:02:37PM -0300, Luiz Capitulino wrote:
> > > > This error is currently returned by inet_connect_opts(), however
> > > > it causes the follow spurious message on HMP:
> > > > 
> > > >     (qemu) migrate tcp:0:4444
> > > >     migrate: Connection can not be completed immediately
> > > >     (qemu)
> > > > 
> > > > But migration succeeds.
> > > 
> > > I think the core issue is that inet_connect_opts() passes back the
> > > QERR_SOCKET_CONNECT_IN_PROGRESS via Error (which is fine), but that
> > > we have users that erroneous pass this error up the stack, when really,
> > > when specifying blocking=on as one of the options, they should be
> > > expecting and doing specific handling for this error.
> > 
> > You're right here.
> > 
> > > So if we fix that (by simply using a local Error when doing the call and
> > > using error_propagate() for non QSCIP errors), I think we can basically
> > > drop patches 14-17 by fixing the callers in that manner and just giving QSCIP
> > > it's own error class.
> > 
> > I don't think QSCIP errors is something we should report to QMP clients, at
> > least not for the use-case this patch is about, hence we should not have
> > a specific error class for this.
> 
> But we do have internal users besides QMP, and in this case they're
> interested in a specific error. What if we generalized it to EAGAIN or
> something? It's seems to me a fairly reasonable exception since it's one
> of the few errno-style errors that we don't generally propagate up the
> stack and need to check for explicitly...

The fact that we don't propagate it up and that this is used only for
communication between two functions are the reasons for not creating an
error class for it.

For this specific case, my in_progress change is a better solution IMO.

> > As pointed out by Markus in his review, keeping the in_progress flag introduced
> > by patch 14/34 should be enough to drop patches 15 and 16.
> 
> Although, being an exceptional case I guess having an "in_progress" field
> to functions would use it is reasonable...
> 
> I think I'd still prefer a class for QSCIP/EAGAIN that we could use for
> socket utility functions, but I'm okay with an in_progress param.
> 
> > 
> > > Relying on the errno result was something these socket errors were
> > > specifically meant to fix, since errno is set multiple times
> > > throughout the function and extracting an errno reliably requires
> > > callers to examine all the possible error paths and errno setters. So I
> > > think it's a regression to go back to the old behavior, and these were
> > > issues found in inet_connect() when we attempted to generalize it's
> > > usage for non-blocking connections.
> > 
> > I'm not completely sure I agree because the new error format doesn't allow
> > callers to programatically know the cause of an failure. That's what errno
> 
> Is error_get_class() not to be used for this purpose? It seems like a
> good thing to allow for in the odd circumstances where we do end up
> adding new error classes (unless the notion of error classes is purely
> legacy support for ones that libvirt is dependent on, and new ones will
> never be added?)

I think I didn't express myself correctly here.

Yes, it's possible to use error_get_class() to programatically check for an
error class. There's nothing wrong with that.

The problem though, is that this series shrinks the number of errors classes
from 70+ to a few. So, most errors causes will be GenericError.

Take the block layer for example. Several block layer functions check for
errno and take different actions depending on the error cause. When we
Errorify the block layer (and we'll do it to propagate the error cause to
users) we won't be able to drop errno in favor of error_get_class(), otherwise
we'll be back to dozens of error classes (worse, most of them will only
be useful for the block layer or any subsystem doing similar things).

For the block layer case, I think we'll need to introduce error_sete() and
error_get_errno().

> > callers to programatically know the cause of an failure. That's what errno
> > is for, though.
> 
> But it's just simply unusable when calling into a function that has
> multiple paths that can set it (or clobber it). Errno values frequently
> require the context of the function that set it to do anything intelligent,
> which is why QSCIP was added to remove that burden from users of
> inet_connect_opts() and friends.

That's case by case. If there's a case where adding a new class make sense,
then we can do it. I don't think this is the case, though. Besides, the current
usage of QSCIP is buggy.

> It's good that errors are no longer tethered to the errors
> descriptions/parameters and that that has amounted to a big reduction
> in the number of error classes we have, but that doesn't mean we shouldn't
> be open to added new error classes in the future, where it makes sense.
> 
> But, again, an in_progress param seems like a workable compromise here, I
> just think prefering this approach over new error classes may lead to
> unecessary code churn in the future.
> 
> > 
> > But I'll drop the patch that changes inet_connect() to return errno,
> > so it's not worth it to discuss this specific case.
> > 
>