Assertion failure: obj == cacheResult, at js/src/vm/Compartment-inl.h:104 (nuked-realm bypass: AllowNewWrapper checked against Window's realm, not the reified WindowProxy)
Categories
(Core :: JavaScript Engine, defect, P1)
Tracking
()
People
(Reporter: decoder, Assigned: jandem)
References
(Blocks 1 open bug)
Details
(Keywords: ai-involved, sec-low, testcase, Whiteboard: [adv-main153-])
Attachments
(3 files, 2 external files)
Summary
Compartment::getNonWrapperObjectForCurrentCompartment checks the nukedIncomingWrappers guard (AllowNewWrapper) against the object that is passed in, but then performs a Window→WindowProxy identity switch and creates the wrapper for a different object. For a bare Window, the check tests the Window's realm, while the wrapper that is actually minted targets the WindowProxy. After a same-compartment navigation the WindowProxy lives in a different realm of the same compartment, and if that realm's incoming wrappers were nuked the function still mints a fresh, fully-live cross-compartment wrapper into the nuked realm — reviving read/write/inject access that nuking was supposed to sever.
This is the same nuked-realm-isolation invariant class Mozilla fixed in bug 2031152 (WeakRef.deref bypassing nukeCCW) and the bug 2032512/2032328 series; this instance is unfixed in the tested snapshot (the file is identical to current mozilla-central).
Root cause
js/src/vm/Compartment.cpp, Compartment::getNonWrapperObjectForCurrentCompartment (lines 212-292):
// js/src/vm/Compartment.cpp:237-242
// Disallow creating new wrappers if we nuked the object's realm ...
if (!AllowNewWrapper(this, obj)) { // checks obj's CURRENT realm
obj.set(NewDeadProxyObject(cx, obj));
return !!obj;
}
...
// js/src/vm/Compartment.cpp:244-263
if (IsWindow(obj)) {
obj.set(ToWindowProxyIfWindow(obj)); // IDENTITY SWITCH
obj.set(UncheckedUnwrap(obj));
...
}
AllowNewWrapper (js/src/proxy/CrossCompartmentWrapper.cpp:482) tests obj->nonCCWRealm()->nukedIncomingWrappers. For a bare Window this tests the Window's realm. The code then reifies obj into the Window's current WindowProxy. After a same-compartment navigation (browser: nsGlobalWindowOuter::SetNewDocument; shell: newGlobal({sameCompartmentAs, transplantWindowProxy})) the brain-transplanted WindowProxy belongs to a different realm of the same compartment. Neither the rest of getNonWrapperObjectForCurrentCompartment nor Compartment::getOrCreateWrapper (lines 294-342) re-checks AllowNewWrapper against the new identity object, so a fresh live CCW into the nuked realm is created and cached.
The wrapper map is then inconsistent with a recomputed wrap: a direct value-wrap of the same WindowProxy takes AllowNewWrapper → false → NewDeadProxyObject, while the debug cache-consistency check in JS::Compartment::wrap(JSContext*, MutableHandleValue) finds the revived live CCW:
// js/src/vm/Compartment-inl.h:104
MOZ_ASSERT_IF(cacheResult, obj == cacheResult); // FAILS
Reproduction
SpiderMonkey JS shell, debug ASan build, product_version = 20260603-5cffec5992d3.
Run the included testcase:
js --fuzzing-safe poc.js
poc.js (full step-by-step in the comment header):
g1 = newGlobal({newCompartment:true, useWindowProxy:true})— Window W1 / realm R1 / WindowProxy P; keepe1(surviving CCW to an R1 eval function) andwp1(CCW to P).g2 = newGlobal({sameCompartmentAs:g1, useWindowProxy:true, transplantWindowProxy:wp1})— same-compartment navigation: P is brain-swapped in place into new realm R2.g2.eval("nukeAllCCWs()")— setsR2.nukedIncomingWrappers; R1 untouched.- Wrap the bare Window W1 into the main compartment via
Debugger.Object.prototype.unsafeDereference():AllowNewWrapper(main, W1)passes (W1 in R1), the IsWindow block switches to P (R2, nuked), and a fresh live CCW main→P is minted. e1("globalThis")returns P as a value →JS::Compartment::wrap(Value): wrapper-map cache hits the revived live CCW while the recompute produces a dead proxy → assert fails.
In a debug build this trips the assert (poc.js). In a release/opt build the assert is compiled out and the revived wrapper stays live: demo-release.js demonstrates read/write/inject through it after the nuke (read through revived wrapper: sensitive-value-in-g2, write through revived wrapper: 42).
Crash signature
Assertion failure: obj == cacheResult, at js/src/vm/Compartment-inl.h:104
==ERROR: AddressSanitizer: SEGV on unknown address 0x000000000000 (WRITE)
#0 MOZ_CrashSequence mozilla/Assertions.h:261
#1 JS::Compartment::wrap(JSContext*, JS::MutableHandle<JS::Value>) js/src/vm/Compartment-inl.h:104
#2 js::CrossCompartmentWrapper::call(...) js/src/proxy/CrossCompartmentWrapper.cpp:233
#3 js::Proxy::call(...) js/src/proxy/Proxy.cpp:700
#4 js::InternalCallOrConstruct(...) js/src/vm/Interpreter.cpp:574
Full output in crash-output.txt.
Browser relevance
Same-compartment, multi-realm windows model same-origin navigations that reuse the outer window; transplantWindowProxy (bug 2037348) was added to the shell to model nsGlobalWindowOuter::SetNewDocument. Realm-targeted nuking runs on window/document teardown (WindowDestroyedEvent::Run, xpc::NukeAllWrappersForRealm), and the browser commonly nukes with DontNukeWindowReferences, keeping window-proxy CCWs alive across the flag-setting nuke — making the missed re-check more reachable there than in the shell. Wrapping bare Window objects is routine in Gecko (the IsWindow path exists for it); the Debugger route in the PoC is just the shell's way of producing such a wrap.
Suggested fix
Re-evaluate AllowNewWrapper(this, obj) against the reified identity object — after the IsWindow/ToWindowProxyIfWindow/UncheckedUnwrap sequence, and again in getOrCreateWrapper before creating a wrapper — returning a dead proxy when the WindowProxy's current realm has nukedIncomingWrappers set.
Artifacts
poc.js— debug-build assertion testcase.crash-output.txt— full crash output (debug ASan shell).demo-release.js— opt-build demonstration of the nuke-isolation bypass.
| Reporter | ||
Comment 1•5 months ago
|
||
| Reporter | ||
Comment 2•5 months ago
|
||
| Reporter | ||
Comment 3•5 months ago
|
||
Comment 4•5 months ago
|
||
This bugs severity seems to have been set during creation. I am resetting the severity to --- so the SpiderMonkey team will see this as un-triaged and go through our triage process. Otherwise we might miss this bug.
Comment 5•5 months ago
|
||
Jan, could help me look at who might be the right person to look at this bug?
Updated•5 months ago
|
| Assignee | ||
Updated•5 months ago
|
| Assignee | ||
Comment 6•5 months ago
|
||
After a same-compartment navigation the WindowProxy lives in a different realm of the same compartment, and if that realm's incoming wrappers were nuked the function still mints a fresh, fully-live cross-compartment wrapper into the nuked realm — reviving read/write/inject access that nuking was supposed to sever.
Nuking is used to avoid memory leaks for chrome => content wrappers; it's not really a security boundary.
Updated•5 months ago
|
| Assignee | ||
Comment 7•5 months ago
|
||
| Assignee | ||
Updated•5 months ago
|
Comment 9•5 months ago
|
||
Comment 10•5 months ago
|
||
Please nominate this for Beta uplift when you get a chance.
| Assignee | ||
Updated•4 months ago
|
Comment 11•4 months ago
|
||
firefox-beta Uplift Approval Request
- User impact if declined/Reason for urgency: Fixes an issue in the wrapping code. Uplifting only to have the fix in ESR153.
- Code covered by automated testing?: yes
- Fix verified in Nightly?: yes
- Needs manual QE testing?: no
- Steps to reproduce for manual QE testing:
- Risk associated with taking this patch: low
- Explanation of risk level: Just adds an extra check to handle a corner case.
- String changes made/needed?: N/A
- Is Android affected?: yes
| Assignee | ||
Comment 12•4 months ago
|
||
Original Revision: https://phabricator.services.mozilla.com/D307475
Updated•4 months ago
|
Updated•4 months ago
|
Comment 13•4 months ago
|
||
| uplift | ||
Updated•4 months ago
|
Updated•4 months ago
|
Updated•2 months ago
|
Description
•