372309 - Crash in [@SetArrayElement] using canvas

Status: VERIFIED FIXED

(Core :: JavaScript Engine, defect)

Description

[Eli Friedman]

Attachment: Testcase (not simplified)

WARNING: Frame IP not in any known module. Following frames may be wrong.
0xcdcdcd20
js3250!SetArrayElement(struct JSContext * cx = 0x02d80d58, struct JSObject * obj = 0x03afa5c0, unsigned long index = 0x12e5dc, long v = 511)+0x5e [c:\mozilla\mozilla\js\src\jsarray.c @ 299]
js3250!InitArrayElements(struct JSContext * cx = 0x000c57d1, struct JSObject * obj = 0x03afa5c0, unsigned long start = 0x62be8, unsigned long end = 0x9c400, long * vector = 0x03afae9c)+0x39 [c:\mozilla\mozilla\js\src\jsarray.c @ 738]
js3250!InitArrayObject(struct JSContext * cx = 0x03afae9c, struct JSObject * obj = 0x03afa5c0, unsigned long length = 0x9c400, long * vector = 0x08490040)+0x61 [c:\mozilla\mozilla\js\src\jsarray.c @ 762]
js3250!js_NewArrayObject(struct JSContext * cx = 0x002c5dd7, unsigned long length = 0xff000000, long * vector = 0x0012e7e4)+0x31 [c:\mozilla\mozilla\js\src\jsarray.c @ 1997]
gklayout!nsCanvasRenderingContext2D::GetImageData(void)+0x2fc [c:\mozilla\mozilla\content\canvas\src\nscanvasrenderingcontext2d.cpp @ 3077]
xpcom_core!NS_InvokeByIndex(class nsISupports * that = 0x063ff9c8, unsigned int methodIndex = 0x37, unsigned int paramCount = 0, struct nsXPTCVariant * params = 0x0012e6b8)+0x27 [c:\mozilla\mozilla\xpcom\reflect\xptcall\src\md\win32\xptcinvoke.cpp @ 102]
xpc3250!XPCWrappedNative::CallMethod(class XPCCallContext * ccx = 0x00b78aa8, XPCWrappedNative::CallMode mode = 12028584 (No matching enumerant))+0x7f2 [c:\mozilla\mozilla\js\src\xpconnect\src\xpcwrappednative.cpp @ 2251]

Seems like an evil enough crash that it might be worth filing as a security bug.

To reproduce, load up two images into the testcase (don't know of any images off the top of my head on b.m.o to refer to; you can save the testcase and put the urls of two local files in.)  Then, click the compare button.  Then, click it again.  Watch it explode.  (Testcase could probably be reduced more, but I haven't done it yet.)

Comment 1

[Eli Friedman]

Attachment: Simplified testcase (WARNING: CRASHES INSTANTLY)

Hmm, should have simplified before posting originally... pretty simple :(

Comment 2

[Brian Crowder]

bz thinks, and I agree, that this is a problem of the new array object not having a root and InitArrayElements CHECK_OPERATION_LIMIT potentially yielding a GC (or many other) operations which smash the array's newborn root (if there ever is one in this case??).  I'll take a stab at a patch...

Since it is a GC-hazard, I feel that it most certainly should be considered a critical security bug.

Comment 3

[Brendan Eich [:brendan]]

Trunk-only, though.

/be

Comment 4

[Brian Crowder]

Attachment: temp root for array object

Not sure if this is correct, but seems to be the common technique.  Whoever the array object is passed to will be responsible for rooting it again.

Comment 5

[Igor Bukanov]

(In reply to comment #4)
> Created an attachment (id=256968) [details]
> temp root for array object
> 
> Not sure if this is correct, but seems to be the common technique.  Whoever the
> array object is passed to will be responsible for rooting it again.
> 

For complete safety it is necessary to put obj into newborn array to restore invariant that the callback can break.

Comment 6

[Brian Crowder]

Attachment: better behavior

Thanks for the feedback, Igor.

Comment 7

[Igor Bukanov]

Comment on attachment 256975 [details] [diff] [review]
better behavior

> JSObject *
> js_NewArrayObject(JSContext *cx, jsuint length, jsval *vector)
> {
...
>+    tvr.u.value = OBJECT_TO_JSVAL(obj);
>     if (!InitArrayObject(cx, obj, length, vector)) {
>         cx->weakRoots.newborn[GCX_OBJECT] = NULL;
>         return NULL;
>     }
>+    JS_POP_TEMP_ROOT(cx, &tvr);
>+
>+    /* Set newborn root, in case we lost it.  */
>+    cx->weakRoots.newborn[GCX_OBJECT] = obj;
>     return obj;

Nit: the if (!InitArrayObject) part can be written as:

if (!InitArrayObject(cx, obj, length, vector))
  obj = NULL;

Comment 8

[Igor Bukanov]

Comment on attachment 256975 [details] [diff] [review]
better behavior

>+    JS_PUSH_TEMP_ROOT_OBJECT(cx, obj, &tvr);
>+    tvr.u.value = OBJECT_TO_JSVAL(obj);
>     if (!InitArrayObject(cx, obj, length, vector)) {
>         cx->weakRoots.newborn[GCX_OBJECT] = NULL;
>         return NULL;

Non-nit: this will leave tvr pushed.

Comment 9

[Brian Crowder]

Attachment: igor's nit

This is the patch I will land this evening, barring further feedback.  Thanks, Igor.

Comment 10

[Brian Crowder]

Attachment: tweak to kill a compiler warning

Comment 11

[Brian Crowder]

Attachment: with the right patch this time

Sorry for the bugspam

Comment 12

[Igor Bukanov]

(In reply to comment #10)
> Created an attachment (id=256977) [details]
> tweak to kill a compiler warning
> 

The diff shows that this patch and the previous are the same.

Comment 13

[Brian Crowder]

The last patch doesn't have the additional issue of leaving tvr pushed, either.

Comment 14

[Brendan Eich [:brendan]]

Comment on attachment 256978 [details] [diff] [review]
with the right patch this time

>+    JS_PUSH_TEMP_ROOT_OBJECT(cx, obj, &tvr);
>+    tvr.u.value = OBJECT_TO_JSVAL(obj);

The macro sets u.value for you.

>+    if (!InitArrayObject(cx, obj, length, vector))
>+        obj = NULL;
>+    JS_POP_TEMP_ROOT(cx, &tvr);
>+
>+    /* Set/clear newborn root, in case we lost it.  */
>+    cx->weakRoots.newborn[GCX_OBJECT] = (JSGCThing *) obj;

Need to get rid of newborn roots, but that's another bug (on Igor's list IIRC).

/be

Comment 15

[Brian Crowder]

Attachment: trust the macro!

ugh, I went and found that macro so I could get rid of that line then didn't get rid of the line.  Juggling too many patches today.  Thanks, Brendan

Comment 16

[Brian Crowder]

jsarray.c: 3.107

Comment 17

[Bob Clary [:bc] (inactive)]

Attachment: js1_5/extensions/regress-372309.js

Comment 18

[Bob Clary [:bc] (inactive)]

verified fixed 1.9.0 20070320 win/mac*/linux

Comment 19

[Bob Clary [:bc] (inactive)]

/cvsroot/mozilla/js/tests/js1_5/extensions/regress-372309.js,v  <--  regress-372309.js
initial revision: 1.1

Comment 20

[Igor Bukanov]

The bug exists on 1.8.1 branch, for details see bug 360701.

Comment 21

[Igor Bukanov]

Comment on attachment 256989 [details] [diff] [review]
trust the macro!

The patch must go to 1.8.1 branch to fix the GC hazard there.

Comment 22

[Daniel Veditz [:dveditz]]

Comment on attachment 256989 [details] [diff] [review]
trust the macro!

approved for 1.8.1.8, a=dveditz for release-drivers

Comment 23

[Igor Bukanov]

I checked in the patch from comment 15 to MOZILLA_1_8_BRANCH:

Checking in jsarray.c;
/cvsroot/mozilla/js/src/jsarray.c,v  <--  jsarray.c
new revision: 3.58.2.26; previous revision: 3.58.2.25
done

Comment 24

[Carsten Book [:Tomcat]]

verified fixed 1.8.1.8 using the testcase from comment #1 and Mozilla/5.0 (X11; U; Linux i686; en-US; rv:1.8.1.8pre) Gecko/2007092903 BonEcho/2.0.0.8pre on Fedora F7 and Mozilla/5.0 (Windows; U; Windows NT 5.2; en-US; rv:1.8.1.8pre) Gecko/20070927 BonEcho/2.0.0.8pre ID:2007092704 - no crash on testcase

-> Adding verified keyword

Comment 25

[Alexander Sack]

Comment on attachment 256989 [details] [diff] [review]
trust the macro!

a=asac for 1.8.0.15

Comment 26

[Alexander Sack]

please land on 1.8.0 branch.

Comment 27

[Reed Loden [:reed]]

MOZILLA_1_8_0_BRANCH:

Checking in js/src/jsarray.c;
/cvsroot/mozilla/js/src/jsarray.c,v  <--  jsarray.c
new revision: 3.58.2.10.2.13; previous revision: 3.58.2.10.2.12
done