Open Bug 2050826 Opened 4 months ago Updated 1 day ago

Crash in [@ js::GCData<T>::get]

Categories

(Core :: JavaScript: GC, defect, P5)

Unspecified
All
defect

Tracking

()

Tracking Status
firefox-esr115 --- unaffected
firefox-esr140 --- unaffected
firefox152 --- unaffected
firefox153 --- wontfix
firefox154 --- wontfix

People

(Reporter: dmeehan, Unassigned)

References

(Regression)

Details

(Keywords: crash, regression)

Crash Data

Crash report: https://crash-stats.mozilla.org/report/index/d7333ef0-4383-47f5-af1e-e7cb20260626

Reason:

EXCEPTION_ACCESS_VIOLATION_READ

Top 10 frames:

0  xul.dll  js::GCData<unsigned int>::get() const  js/src/gc/Barrier.h:579
0  xul.dll  js::GCData<unsigned int>::operator unsigned int() const  js/src/gc/Barrier.h:586
0  xul.dll  js::ObjectElements::isFixed() const  js/src/vm/NativeObject.h:395
0  xul.dll  js::IsNativeObjectFixedElements(js::HeapSlot*)  js/src/vm/NativeObject.h:631
0  xul.dll  js::IsNativeObjectDynamicElements(js::HeapSlot*)  js/src/vm/NativeObject.h:635
0  xul.dll  js::gc::MarkingTracerT<2>::processMarkStackTop(JS::SliceBudget&)  js/src/gc/Marking.cpp:1903
0  xul.dll  js::GCMarker::markCurrentColorInParallel(js::gc::ParallelMarkTask*, JS::Slice...  js/src/gc/Marking.cpp:1503
1  xul.dll  js::gc::ParallelMarkTask::tryMarking(js::AutoLockHelperThreadState&)  js/src/gc/ParallelMarking.cpp:198
1  xul.dll  js::gc::ParallelMarkTask::run(js::AutoLockHelperThreadState&)  js/src/gc/ParallelMarking.cpp:165
2  xul.dll  js::GCParallelTask::runTask(JS::GCContext*, js::AutoLockHelperThreadState&)  js/src/gc/GCParallelTask.cpp:219
Flags: needinfo?(jcoppeard)
Keywords: regression
Regressed by: 2044578

Likely a signature shift following the addition of GCData<T>.

Severity: -- → S4
Flags: needinfo?(jcoppeard)
Priority: -- → P5

The bug is linked to a topcrash signature, which matches the following criteria:

  • Top 20 desktop browser crashes on beta
  • Top 10 content process crashes on beta

:gtonietto, could you consider increasing the severity of this top-crash bug?

For more information, please visit BugBot documentation.

Flags: needinfo?(gtonietto)
Keywords: topcrash
Flags: needinfo?(gtonietto)

This is a signature shift from bug 2024333 (which was itself a shift from an existing bad-memory type crash).

Based on the topcrash criteria, the crash signature linked to this bug is not a topcrash signature anymore.

For more information, please visit BugBot documentation.

Keywords: topcrash
OS: Windows 11 → All

Crash report: https://crash-stats.mozilla.org/report/index/27f056eb-186c-4e70-98e2-1c4030261116

Crash Reason:

EXCEPTION_ACCESS_VIOLATION_READ at 0x0000000000003c07

Top 10 frames:

0  xul.dll  js::SavedStacks::insertFrames(JSContext*, JS::MutableHandle<js::SavedFrame *>, m...  js/src/vm/SavedStacks.cpp:1583
1  xul.dll  js::ErrorToException(JSContext*, JSErrorReport*, JSErrorFormatString const* (*)(...  js/src/vm/ErrorObject.cpp:1375
2  xul.dll  js::ReportErrorNumberVA(JSContext*, js::IsWarning, JSErrorFormatString const* (*...  js/src/vm/ErrorReporting.cpp:484
3  xul.dll  JS_ReportErrorNumberLatin1(JSContext*, JSErrorFormatString const* (*)(void*, con...  js/src/jsapi.cpp:3999
4  xul.dll  js::ReportNotObjectArg(JSContext*, char const*, char const*, JS::Handle<JS::Valu...  js/src/vm/JSObject.cpp:109
5  xul.dll  js::proxy(JSContext*, unsigned int, JS::Value*)  js/src/proxy/ScriptedProxyHandler.cpp:1551
6
7
8
9

Filed because this signature's crash rate rose with Firefox 160. Firefox 160, in its first 5 days on release (2026-11-11 to 2026-11-15): this signature crashed 175 of the 35,301 installations with a crash report (226 reports). Firefox 159, from its release on 2026-10-28 to 160's: 442 of 129,518 (499 reports). That is 1.45x the share of installations (binomial excess z = 4.2, bar 3.62). The higher share shows from its first days on release. This signature is a new name for an older crash: until build 20260715202819 (2026-07-15) it reported as js::ScriptWarmUpData::getTaggedPtr (21 reports on the builds of the 28 days before, 4 since; the new frame js::GCData<T>::get took over the name from js::ScriptWarmUpData::getTaggedPtr); js::ScriptWarmUpData::getTaggedPtr was first recorded in build 20221006191955 (2022-10-06).

Crash-stats counts for this signature on release when this was posted:

Firefox 159 (2026-10-28 to 2026-11-10) Firefox 160 (2026-11-11 to 2026-11-15)
Reports 499 226
Distinct install times at least 442 at least 175
Reports from the busiest install time 16 (3%) 11 (5%)
Reports from Intel Raptor Lake CPUs (family 6 model 183 stepping 1) 59 (12%) 63 (28%)

Clouseau analysis (automated -- nothing below was checked by a human; a claim it could not ground is deliberately absent):

This signature is a catch-all SpiderMonkey frame (js::GCData<T>::get, inlined field reads such as ObjectElements flags, getDenseInitializedLength and ScriptWarmUpData::getTaggedPtr) that faults on a wild or near-null pointer across many unrelated GC and VM paths. It was called js::ScriptWarmUpData::getTaggedPtr until a rename in July 2026, and has existed since 2022. In Firefox 160's first five days on release, crashes from known-defective Intel Raptor Lake CPUs (family 6 model 183 stepping 1) rose from 10% to 28% of reports, a second Intel model rose from outside the top 20 to 9% (with a few machines crashing repeatedly), and Linux all but disappeared. No new stack cohort appeared. The representative report decodes a nonsensical or byte [rax-0x7d], cl instruction inside an inlined field read, which fits corrupted execution on that CPU. The rise is best explained by which machines adopted 160 first and by unstable hardware, not by a Firefox change. Before changing any code, recompute the rate per version without Raptor Lake and the repeat-crash installations once 160 is fully rolled out.

Possible path to the crash:

Mechanism (observed): a read access violation inside an inlined js::GCData<T>::get field read. In the representative report the chain is ScriptWarmUpData::getTaggedPtr -> BaseScript::jitScript -> JSScript::baselineScript, reached from JSJitFrameIter during SavedStacks::insertFrames, while an error raised by js::proxy / ReportNotObjectArg was capturing the stack. The decoded instruction is or byte [rax-0x7d], cl with a memory access at 0x3b82, while the top-level address is 0x3c07; the two values disagree.

Path (observed): the signature collects many unrelated callers (tenuring, structured clone, property enumeration, SpreadCallOperation). No proto-signature reaches more than 2% before or after the split, so no single code path drives the spike.

Trigger (inferred, medium): hardware instability, mainly on Raptor Lake CPUs. A read-modify-write or instruction makes no sense for an inlined 64-bit field load, which suggests execution went off track (the CPU decoded from the wrong place). Raptor Lake's share rose from 10% to 28% in 160 builds, and the brief reports that 71% of this signature's reports over the past year carry a bit-flip annotation.

Population (inferred, medium): the version step comes from the early-adopter mix. 160's early reports are 92% Windows (78% before), with very few from Linux, and several installations crashed repeatedly. Meanwhile 159's remaining installations dropped to 0.72x over the same dates, consistent with crash-prone machines moving to 160 rather than new crashes appearing.

Checked:

  • Raptor Lake (family 6 model 183 stepping 1) was 133/1275 (10%) of reports before build 20261104181552 and 63/228 (28%) from it on. Family 6 model 186 stepping 3 rose from outside the top 20 to 20/228 (9%). (crashstats facets cpu_info, split_at_build 20261104181552, 30 days, release)
  • No proto-signature cohort dominates either population. The largest is 24/1275 before and 3/228 after, and the after-split cohorts are the same families as before (ObjectElements, getDenseInitializedLength, ScriptWarmUpData). (crashstats facets proto_signature, split_at_build 20261104181552, 30 days)
  • OS mix shifted: Windows 989/1275 (78%) before vs 210/228 (92%) after. Linux distributions are each 0-1% after the split, but Mint, Ubuntu and Flatpak were each 1-4% before. (crashstats facets platform_pretty_version, split_at_build 20261104181552, 30 days)
  • Repeat crashers in 160: one install_time value has 11 reports and another has 8. Microcode 0x6134 accounts for 20 reports (9%), probably the few model-186 machines. (crashstats facets install_time and cpu_microcode_version, split_at_build 20261104181552, 30 days)
  • Representative report: Raptor Lake with microcode 0x132, content process, uptime 114 s, plenty of free memory. The decoded instruction or byte [rax - 0x7d], cl reads 0x3b82, while the top-level address is 0x3c07. Frame #0 inlines GCData<uint64_t>::get -> ScriptWarmUpData::getTaggedPtr -> JSScript::baselineScript -> JSJitFrameIter, in SavedStacks::insertFrames, called from ErrorToException / ReportNotObjectArg / js::proxy. (crashstats report 27f056eb-186c-4e70-98e2-1c4030261116)
  • Reports per 1000 same-version reports: 160.0 is 4.47, against 2.17-2.5 for 157.0.1-159.0.1. The brief's test says this is within 3x of the previous version, so not a step. (brief, CRASH RATE BY VERSION)
  • The existing bug 2050826 (Core :: JavaScript: GC) has regressed_by 2044578 ('Make GC barrier wrappers atomic under concurrent marking'). That bug fits the July 2026 rename that introduced the GCData frame, not a 160 rate change. [inferred, medium confidence] (bugzilla 2050826, 2044578; brief SIGNATURE RENAME (handoff at release build 20260715202819))
  • The decoded or instruction makes no sense for the inlined source expression (a plain 64-bit field load), which points to execution going off track rather than a bad pointer in correct code. [inferred, medium confidence] (crashstats report 27f056eb crash_info.instruction plus inline chain)

Alternatives considered:

  • A new code path or call site introduced in 160: disfavored -- the after-split proto-signatures are the same diverse families as before, with no new dominant cohort.
  • Bug 2044578 as the cause of the 160 step: disfavored -- the rename to this signature happened at build 20260715202819, months before 160. The diff was not read, so it is not ruled out as a source of the underlying crash.
  • A real per-installation rate regression in 160 code: unresolved -- the excess is concentrated in Raptor Lake CPUs, repeat-crash installations and a Windows-heavy early population, and the rate per version is within the brief's 3x bar.

Worth checking first:

  • Is there any excess left after removing hardware-flagged reports? Recompute 160 vs 159 reports per 1000 (or installations) excluding family 6 model 183/191 and bit-flip-annotated reports, after 160 has reached the Linux population.
  • Do non-Raptor-Lake 160 reports show sensible decoded instructions? Sample a few 160.0 reports on other CPUs and check whether crash_info.instruction matches the inlined source load, which would point to genuine wild pointers rather than corrupted execution.
  • Is the model-186/microcode 0x6134 cluster one machine? Look at reports for install_time 1794340882 and 1794496569.

Posted automatically by Clouseau, which analyses release crashes with an LLM. Nothing above was written or checked by a human.

You need to log in before you can comment on or make changes to this bug.