bug#28211: Stack marking issue in multi-threaded code, 2020 edition

[Ludovic Courtès]

Hi,

Ludovic Courtès <address@hidden> skribis:

> What if ‘scm_i_vm_mark_stack’ walks the stack right before the ‘vp->fp’
> assignment?  It can determine that one of the just-assigned ‘sp[nargs]’
> is a dead slot, and thus set it to SCM_UNSPECIFIED.  Later, when we set
> ‘vp->fp’, that stack slot that we just initialized has been overwritten
> by ‘scm_i_vm_mark_stack’.  Down the road, we get something like:
>
>   Wrong type to apply: #<unspecified>

I believe the patch below solves this problem.  Also attached is a small
program that reproduces the bug (without the patch) after a couple of
minutes.

Thoughts?

The full grafting test case exposes another VM stack-corruption-looking
crash that I’m still investigating.

Ludo’.

diff --git a/libguile/vm.c b/libguile/vm.c
index b20c6eb5f..5cc934e99 100644
--- a/libguile/vm.c
+++ b/libguile/vm.c
@@ -1351,7 +1351,8 @@ scm_i_vm_emergency_abort (SCM *tag_and_argv, size_t n)
   scm_t_bits *prompt;
   scm_t_dynstack_prompt_flags flags;
   ptrdiff_t fp_offset, sp_offset;
-  union scm_vm_stack_element *fp, *sp;
+  union scm_vm_stack_element *fp;
+  volatile union scm_vm_stack_element *sp;
   SCM *argv;
   uint32_t *vra;
   uint8_t *mra;
@@ -1392,16 +1393,19 @@ scm_i_vm_emergency_abort (SCM *tag_and_argv, size_t n)
      to be jumping to an older part of the stack.  */
   if (sp < vp->sp)
     abort ();
-  sp[nargs].as_scm = cont;
-
-  while (nargs--)
-    sp[nargs].as_scm = *argv++;
 
   /* Restore VM regs */
   vp->fp = fp;
-  vp->sp = sp;
+  vp->sp = (union scm_vm_stack_element *) sp;
   vp->ip = vra;
 
+  /* Restore the arguments on SP.  This must be done after 'vp->fp' has
+     been set so that a concurrent 'scm_i_vm_mark_stack' does not
+     overwrite it (see <https://bugs.gnu.org/28211>).  */
+  sp[nargs].as_scm = cont;
+  while (nargs--)
+    sp[nargs].as_scm = *argv++;
+
   /* Jump! */
   vp->mra_after_abort = mra;
   longjmp (*registers, 1);
@@ -1417,7 +1421,8 @@ abort_to_prompt (scm_thread *thread, uint8_t *saved_mra)
   scm_t_bits *prompt;
   scm_t_dynstack_prompt_flags flags;
   ptrdiff_t fp_offset, sp_offset;
-  union scm_vm_stack_element *fp, *sp;
+  union scm_vm_stack_element *fp, *orig_sp;
+  volatile union scm_vm_stack_element *sp;
   uint32_t *vra;
   uint8_t *mra;
   jmp_buf *registers;
@@ -1452,6 +1457,7 @@ abort_to_prompt (scm_thread *thread, uint8_t *saved_mra)
   /* Recompute FP, as scm_dynstack_unwind may have expanded the stack.  */
   fp = vp->stack_top - fp_offset;
   sp = vp->stack_top - sp_offset;
+  orig_sp = vp->sp;
 
   /* Continuation gets nargs+1 values: the one more is for the cont.  */
   sp = sp - nargs - 1;
@@ -1460,15 +1466,19 @@ abort_to_prompt (scm_thread *thread, uint8_t *saved_mra)
      to be jumping to an older part of the stack.  */
   if (sp < vp->sp)
     abort ();
-  sp[nargs].as_scm = cont;
-  while (nargs--)
-    sp[nargs] = vp->sp[nargs];
 
   /* Restore VM regs */
   vp->fp = fp;
-  vp->sp = sp;
+  vp->sp = (union scm_vm_stack_element *) sp;
   vp->ip = vra;
 
+  /* Restore the arguments on SP.  This must be done after 'vp->fp' has
+     been set so that a concurrent 'scm_i_vm_mark_stack' does not
+     overwrite it (see <https://bugs.gnu.org/28211>).  */
+  sp[nargs].as_scm = cont;
+  while (nargs--)
+    sp[nargs] = orig_sp[nargs];
+
   /* If there are intervening C frames, then jump over them, making a
      nonlocal exit.  Otherwise fall through and let the VM pick up where
      it left off.  */

;; https://issues.guix.gnu.org/issue/28211

(use-modules (ice-9 threads))

(define (thunk)
  (catch 'foo
    (lambda ()
      (throw 'foo (iota 10)))
    (lambda (key lst)
      (unless (equal? lst (iota 10))
        (primitive-_exit 42)))))

(n-par-for-each (* 2 (current-processor-count))
                (lambda _
                  (let loop ()
                    (thunk)
                    (loop)))
                (iota 1000))