Running Ghost sample crashes with Segmentation fault from time to time #3484
Comments
|
Here is the stack trace from |
|
The |
|
BTW, the correct English word is "examples" not "samples". |
|
HMm. OK, I can reproduce the crash, but somehow you disabled the normal crash-processing code? I'm not able to debug this in the usual way, something is preventing a normal crash dump... |
|
SO, in gdb: which is confusing. I don't understand where the backtrace printing is coming from. Who is priting that? |
|
Hmm, I see this: So |
|
Assorted checks in various places show consistently, reproducibly, that both TV and AV use-counts become too low, leading to access of deleted memory and other crazy symptoms. How, exactly that comes to be remains mysterious to me, but is presumably because the |
|
Found it. But only after 8+ hours of difficult debugging. Someone .. not sure who -- either this example demo, or ghost itself, has issued the string What is happening is that Here's how I debugged this, in case you ever want to do the same thing. First, notice that the use-count of So edit and the custom routine We now have a stack trace of whodunnit: Why the heck would OK, right. Someone types in |
|
Anyway .. I'm done debugging. I leave it up to you to figure out who is telling guile to quit and force-run exit handlers (thus shutting down not just truth values, but all of the atomspace) while the atomspace is still in use. |
|
I believe that the problem is the following. (ghost-run) calls (psi-run ghost-component) and it runs a loop by call-with-new-thread to spawn a new thread. (use-modules
(opencog)
(opencog nlp)
(opencog nlp relex2logic)
(opencog openpsi)
(opencog ghost)
(opencog ghost procedures)
)
(ghost-parse "
r: (where [do does] _* work) '_0 work in SomeCompany company.
")
(ghost-run)There is the method (ghost-halt) which aim is to stop the running loop. It actually sets the false value to the (Predicate "run-loop") node. However, ghost-halt has the same lack. It sets the flag to terminate the loop and after that guile can frees Atomspace while the run-loop is still running. When the run-loop tries to read a value from the predicate node it is possible that it does not exists anymore. I can implement a solution there psi-halt sets the false value to the predicate node and then joins to the loop thread to wait it is finished. (define run #t)
(define (repl)
(display "run-loop\n")
(if run
(begin
(sleep 1)
(repl))
(display "Exit!\n")))
(define repl-thread
(call-with-new-thread
(lambda () (repl))))
(sleep 3)
(set! run #f)
(join-thread repl-thread 5)
(sleep 3)The solution requires that (ghost-halt) is always called before exiting from guile. |
|
I'm guessing that ghost starts one or more threads. So, The problem with exit-hooks is that they are run in arbitrary order. There is no particular way to make one exit hook run before another; there's no priority ordering. To be clear: even if you did use |
|
@stellarspot, just want to make sure, are you getting this segfault only when exiting Guile when the loop was still running? Or does it crash even without exiting Guile? |
I prepared the suggested fix #3487 It also requires that a user always needs to call psi-halt for each created by psi-run function thread |
I have not seen the crash during working in Guile, only when it is exiting. |
I run the code sample-ghost-notebook.scm and it crashes with Segmentation fault error sometimes.
Here is the result of call
catchsegv guile samples/ghost/sample-ghost-notebook.scmThe text was updated successfully, but these errors were encountered: