RE: starpack vfs can't be accessed from thread

[email protected] wrote:
> However, while this enables the thread to access the vfs, the 
> package mechanism does not find packages inside the vfs from 
> within the new thread.  If you cd into the directory 
> directly, files can be "source"'d from the thread so the vfs 
> is (sort of) working.  I'm setting the auto_path variable for 
> the thread, is there anything else that's necessary to get 
> the package mechanism to work?

Thanks for continuing to bang your head against this.  I have now identified
the issue.  It was a follow on complication from the changes to vfs to make it
thread-safe (which meant mount points are no longer shared between threads).

This requires a change in the starkit boot.tcl for how it determined whether
to mount the exe for access.  It previously used a check that wasn't really
mount-info based, but made sure it only occured once per process.  I have
modified it to do the mount anytime it doesn't exist.  This guards against any
future eventuality that the mount is re-shared in threads, or even possibly
that each individual interp must remount the exe.  The patch is:

@@ -11,8 +11,11 @@
       # get rid of a build residue
       unset -nocomplain ::tclDefaultLibrary
   
<      # the following code only gets executed once on startup
<      if {[info exists tcl_rcFileName]} {
->     # The following code only gets executed if we don't have our exe
->     # already mounted.  This should only happen once per thread.
->     # We could use [vfs::filesystem info], but that would require
->     # loading vfs into every interp.
->     if {![file isdirectory $noe]} {
   	load {} pwb
   	load {} vfs

You can modify this in the existing 8.4.14 basekit that you have.  We will
have it corrected as well in the next release.

> (8.4.13) when it crashed.  Perhaps there's still something in 
> the vfs code that's not thread-safe?

Possibly.  I had clear conditions to make it crash before, and those are all
addressed.  The vfs C code no longer has any global vars being used, but that
doesn't mean something else isn't amiss.  Please let me know if you have a
reproducable test case.

  Jeff Hobbs, The Tcl Guy, http://www.ActiveState.com/