Hi all,

The attached patch fixes #1337, which has exactly the same root cause as
the "Salmonella crashes a lot" issue we struggled with back in 2015:
http://lists.gnu.org/archive/html/chicken-hackers/2015-10/msg00001.html

I guess we never noticed this because the circumstances are relatively
rare:  You have to use a foreign callback which calls code that triggers
a GC, and this GC needs to happen before the argvector is copied onto
the stack by an intermediate function that happens to call another
function with a larger number of arguments.

I don't know how to make a test case for this that reliably triggers the
error.  The test case from the ticket happens to trigger it now, but it
might not in the future.

I also noticed that when the stack is full in C_allocate_vector, we call
C_reclaim(allocate_vector_2, c), but we save one more thing than the
number of arguments allocate_vector takes.

Cheers,
Peter
From 27925ad4cb1b0ab0e6445116623096fbaf52c3b2 Mon Sep 17 00:00:00 2001
From: Peter Bex <pe...@more-magic.net>
Date: Wed, 28 Dec 2016 19:11:35 +0100
Subject: [PATCH] Fix crashes in callbacks after GC (#1337).

This was caused by keeping argvector in temp stack, similar to 9eed2742.
Instead, we also copy the argvector into the stack before re-invoking
the trampoline when in a callback context, just like a regular CPS
context.

Also, fix an off-by-one in the number of arguments which
allocate_vector_2 accepts; simply passing the original argcount doesn't
allow for the extra 'mode' argument which gets added onto the front of
the stack.
---
 NEWS      |  1 +
 runtime.c | 15 ++++++++++++---
 2 files changed, 13 insertions(+), 3 deletions(-)

diff --git a/NEWS b/NEWS
index 79403b9..38af2d2 100644
--- a/NEWS
+++ b/NEWS
@@ -71,6 +71,7 @@
 
 - Runtime system:
   - "time" macro now shows peak memory usage (#1318, thanks to Kooda).
+  - Avoid crashes in ffi callbacks after GC (#1337, thanks to cosarara).
 
 - Core libraries:
   - Keywords are more consistently read/written, like symbols (#1332).
diff --git a/runtime.c b/runtime.c
index 146da1b..bfaae1a 100644
--- a/runtime.c
+++ b/runtime.c
@@ -1559,6 +1559,9 @@ C_word CHICKEN_run(void *toplevel)
   serious_signal_occurred = 0;
 
   if(!return_to_host) {
+    /* We must copy the argvector onto the stack, because
+     * any subsequent save() will otherwise clobber it.
+     */
     int argcount = C_temporary_stack_bottom - C_temporary_stack;
     C_word *p = C_alloc(argcount);
 
@@ -2134,8 +2137,11 @@ C_word C_fcall C_callback(C_word closure, int argc)
   serious_signal_occurred = 0;
 
   if(!callback_returned_flag) {
-    C_word *p = C_temporary_stack;
-    
+    /* We must copy the argvector onto the stack, because
+     * any subsequent save() will otherwise clobber it.
+     */
+    C_word *p = C_alloc(C_restart_c);
+    C_memcpy(p, C_temporary_stack, C_restart_c * sizeof(C_word));
     C_temporary_stack = C_temporary_stack_bottom;
     ((C_proc)C_restart_trampoline)(C_restart_c, p);
   }
@@ -9975,7 +9981,10 @@ void C_ccall C_allocate_vector(C_word c, C_word *av)
       C_fromspace_top = C_fromspace_limit; /* trigger major GC */
   
     C_save(C_SCHEME_TRUE);
-    C_reclaim((void *)allocate_vector_2, c);
+    /* We explicitly pass 7 here, that's the number of things saved.
+     * That's the arguments, plus one additional thing: the mode.
+     */
+    C_reclaim((void *)allocate_vector_2, 7);
   }
 
   C_save(C_SCHEME_FALSE);
-- 
2.1.4

From a48ce1449a873203f2b8c11c0f1d9facf7b57468 Mon Sep 17 00:00:00 2001
From: Peter Bex <pe...@more-magic.net>
Date: Wed, 28 Dec 2016 19:04:12 +0100
Subject: [PATCH] Fix crashes in callbacks after GC (#1337).

This was caused by keeping argvector in temp stack, similar to 9eed2742.
Instead, we also copy the argvector into the stack before re-invoking
the trampoline when in a callback context, just like a regular CPS
context.

Also, fix an off-by-one in the number of arguments which
allocate_vector_2 accepts; simply passing the original argcount doesn't
allow for the extra 'mode' argument which gets added onto the front of
the stack.
---
 NEWS      |  1 +
 runtime.c | 15 ++++++++++++---
 2 files changed, 13 insertions(+), 3 deletions(-)

diff --git a/NEWS b/NEWS
index 6a34708..ac51659 100644
--- a/NEWS
+++ b/NEWS
@@ -11,6 +11,7 @@
 
 - Runtime system:
   - "time" macro now shows peak memory usage (#1318, thanks to Kooda).
+  - Avoid crashes in ffi callbacks after GC (#1337, thanks to cosarara).
 
 - Core libraries:
   - Keywords are more consistently read/written, like symbols (#1332).
diff --git a/runtime.c b/runtime.c
index 4b6300f..d9ee000 100644
--- a/runtime.c
+++ b/runtime.c
@@ -1504,6 +1504,9 @@ C_word CHICKEN_run(void *toplevel)
   serious_signal_occurred = 0;
 
   if(!return_to_host) {
+    /* We must copy the argvector onto the stack, because
+     * any subsequent save() will otherwise clobber it.
+     */
     int argcount = C_temporary_stack_bottom - C_temporary_stack;
     C_word *p = C_alloc(argcount);
     C_memcpy(p, C_temporary_stack, argcount * sizeof(C_word));
@@ -2007,8 +2010,11 @@ C_word C_fcall C_callback(C_word closure, int argc)
   serious_signal_occurred = 0;
 
   if(!callback_returned_flag) {
-    C_word *p = C_temporary_stack;
-    
+    /* We must copy the argvector onto the stack, because
+     * any subsequent save() will otherwise clobber it.
+     */
+    C_word *p = C_alloc(C_restart_c);
+    C_memcpy(p, C_temporary_stack, C_restart_c * sizeof(C_word));
     C_temporary_stack = C_temporary_stack_bottom;
     ((C_proc)C_restart_trampoline)(C_restart_c, p);
   }
@@ -7508,7 +7514,10 @@ void C_ccall C_allocate_vector(C_word c, C_word *av)
       C_fromspace_top = C_fromspace_limit; /* trigger major GC */
   
     C_save(C_SCHEME_TRUE);
-    C_reclaim((void *)allocate_vector_2, c);
+    /* We explicitly pass 7 here, that's the number of things saved.
+     * That's the arguments, plus one additional thing: the mode.
+     */
+    C_reclaim((void *)allocate_vector_2, 7);
   }
 
   C_save(C_SCHEME_FALSE);
-- 
2.1.4

Attachment: signature.asc
Description: Digital signature

_______________________________________________
Chicken-hackers mailing list
Chicken-hackers@nongnu.org
https://lists.nongnu.org/mailman/listinfo/chicken-hackers

Reply via email to