Tcl Source Code

Under some circumstances a coroutine can be invoked via ::yieldto while the namespace in which it resides is dying (nsPtr->flags&NS_DYING)

This may occur, and is occuring, for example, when a TclOO object is being destroyed, and a coroutine within it is invoked by an event.

The call to TclGetNamespaceFromObj() at tclBasic.c:8532 will indicate an error, because the current frame's namespace is not considered to match itself in the case that it is dying.

I am unsure what the purpose of the test is, or under what circumstances the two namespaces might differ, but this abort occurs as a result of script evaluation, and hence should be considered a bug.

This is now fixed and the ticket can be closed?

In [9198c16407f3] namespace handling is removed from TclNRTailcallObjCmd, allowing TclNRTailcallObjCmd to handle namespace issues later. This aligns the behaviour of the non-compiled variant with the compiled one: An error is produced if the namespace is not found.

If {*} from the example the compiled path for tailcall is taken, and in the compiled variant an error is raised:

namespace "::ns" not found

A slightly smaller example:

namespace eval ns {
	proc p args {
		::tailcall [namespace current] {*}$args
	}
	namespace delete [namespace current]
	p
}

The following script reproduces the issue:

namespace eval ns {
    proc p args {
        ::tailcall [namespace current] {*}$args
    }
    namespace eval [namespace current] {
        namespace delete [namespace current]
        p
    }
}

TIP 498 is one solution to this issue.

This is why it's great to have a test, so we can confirm when the bug is dead…

Is this bug still present?

The problem is indeed with [tailcall] whose guts are reused by [yieldto]. The current namespace was being cached in an nsNameType Tcl_Obj in order to insure its survival, and as a failure to resolve "could not happen" it was signaled by a panic. As colin is wont to do, he triggered the impossible! Fixed by using a better lifetime management of the namespace.

Note: the problem is likely to appear also with [tailcall].

1. no script to repro?
2. any suggestion as to what should happen in that case? Would an error be in order?