starting a swank server segfaults when core files are used

[Attila Lendvai]

a memory fault reliably happens when a swank server is started when sbcl was loaded from a previously saved core file, which had swank initialized.

Nikodemus identified the issue:

I'm thinking you're holding on to a live stream from a previous invocation, and then trying to write to the mmap'ed stream buffer which doesn't exist anymore.

Looking at SIMPLE-ANNOUNCE-FUNCTION, it writes to *LOG-OUTPUT*, which is conditionally initialized in INIT-LOG-OUTPUT. If you accidentally initialize it before saving the core, bad things are bound to happen.

a possible fix:

(in-package :swank)

(defun after-saved-image-initialized ()
  (setf *log-output* nil)
  (init-log-output))

#+sbcl(pushnew 'after-saved-image-initialized sb-ext:*init-hooks*)

[Nikodemus Siivola (nikodemus)]

Is the fault address deterministic?

I'm OK with the tarball.

[Nikodemus Siivola (nikodemus)]

HA! I have a guess.

I'm thinking you're holding on to a live stream from a previous invocation, and then trying to write to the mmap'ed stream buffer which doesn't exist anymore.

Looking at SIMPLE-ANNOUNCE-FUNCTION, it writes to *LOG-OUTPUT*, which is conditionally initialized in INIT-LOG-OUTPUT. If you accidentally initialize it before saving the core, bad things are bound to happen.

Try this:

(defun deinit-swank ()
  ;; Dunno if you need to do anything else?
  (setf swank:*log-output* nil))

(push 'deinit-swank *save-hooks*)

[Attila Lendvai (attila-lendvai)]

brilliant remote debugging skills there Nikodemus!

when i started the image, and i simply

(setf swank:*log-output* nil)

then starting the server works fine!

i think this is a slime bug, or at least a deficiency... but on the other hand, if there's any chance for it, then a better error message from SBCL would be great!

thanks again!

[Attila Lendvai (attila-lendvai)]

for the record, putting this somewhere in swank.lisp fixes the issue:

(defun after-saved-image-initialized ()
  (setf *log-output* nil)
  (init-log-output))

#+sbcl(pushnew 'after-saved-image-initialized sb-ext:*init-hooks*)

[Attila Lendvai (attila-lendvai)]

i'm changing it to be a slime bug.

[Tobias C. Rittweiler (tcr)]

Fixed in Slime 2010-10-08. Fix for SBCL only.

[R. Matthew Emerson (rme)]

CCL could use a similar fix.

 ccl:*save-exit-functions* is where one would push the function to reset *log-output*.

[Stas Boukarev (stassats)]

CCL has now the same fix.