Closed Bug 2084501 Opened 27 days ago Closed 26 days ago

Crash in [@ webrender::intern::impl$9::index]

Categories

(Core :: Graphics: WebRender, defect)

defect

Tracking

()

RESOLVED FIXED
160 Branch
Tracking Status
firefox-esr153 --- unaffected
firefox158 --- wontfix
firefox159 --- wontfix
firefox160 --- fixed

People

(Reporter: clouseau-bot, Assigned: gw)

References

(Blocks 1 open bug, Regression)

Details

(Keywords: crash, regression)

Crash Data

Attachments

(1 external file, 1 obsolete file)

Crash report: https://crash-stats.mozilla.org/report/index/46666b0e-ff91-4cda-a6b0-0d35b0261021

MOZ_CRASH Reason:

bad dl store lookup DlHandle(5, 2)

Top 10 frames:

0  xul.dll  RustMozCrash(char const*, int, char const*)  mozglue/static/rust/wrappers.cpp:17
1  xul.dll  mozglue_static::panic_hook(std::panic::PanicHookInfo*)  mozglue/static/rust/lib.rs:99
2  xul.dll  core::ops::function::FnOnce::call_once<void (*)(ref$<std::panic::PanicHookInfo>)...
3  xul.dll  std::panicking::panic_with_hook()  library/std/src/panicking.rs:833
4  xul.dll  std::panicking::panic_handler::closure$0()  library/std/src/panicking.rs:698
5  xul.dll  std::sys::backtrace::__rust_end_short_backtrace<std::panicking::panic_handler::c...  library/std/src/sys/backtrace.rs:182
6  xul.dll  std::panicking::panic_handler()  library/std/src/panicking.rs:689
7  xul.dll  core::panicking::panic_fmt()  library/core/src/panicking.rs:80
8  xul.dll  webrender::visibility::update_prim_visibility(webrender::prim_store::PictureInde...  gfx/wr/webrender/src/visibility.rs:727
9  xul.dll  webrender::frame_builder::FrameBuilder::build(webrender::scene::BuiltScene*, boo...  gfx/wr/webrender/src/frame_builder.rs:732

There is 1 crash in nightly 160 with buildid 20261020195026.

This signature is not new: its first report anywhere is in build 20260403140140 (2026-04-03), 200 days before the build above. (A 364-day crash-stats search only reaches 2026-05-06, because Socorro's search index keeps about six months — which is why the signature looks much newer there.)

Timing check: this signature was already being reported in build 20260506140522, 167 days before the changeset named below landed, so that changeset cannot be what INTRODUCED this SIGNATURE. It may still be relevant, and it may still be the cause of the crash in this particular report — an old signature can acquire a new cause, and a rare crash can be made frequent by a change that did not create it — but the signature's own history is not evidence that it is. Clouseau lowered its own confidence one step for that, and then put it back: an independent blind re-analysis, given none of the reasoning above, still agreed the changeset can cause this crash. The confidence stated below is that restored one.

Hardware-error share of this signature: of its 9 reports on this channel over the last year, 11% come from the known-buggy Intel Raptor Lake (family 6 model 183 stepping 1, bug 1975808; Firefox-nightly population: 4%). The report analysed above is not one of them — it has no bit-flip annotation and did not come from one of those CPUs — so if this signature is worth anything, it is worth it on crashes like that one.

Clouseau analysis (automated, 72% worth investigating — a calibrated estimate that this is worth someone's time, not that the changeset below caused it). The mechanism below fits the evidence but is not proven end-to-end:

The render backend probably panics in DlStore::index while frame-building an offscreen scene, because that transaction's rectangle store updates were held back when it built no main scene.

  1. 2a53f18f1858 moved rectangle template data into the follower store dl_stores.rectangles and made frame building read it for every rectangle through snap_policy -> prim_has_anti_aliasing -> as_common_data (render_backend.rs:341). It also passed &self.dl_stores to process_offscreen_scene.
  2. On the scene builder thread, SceneMsg::RenderOffscreen always builds an offscreen scene, but built_scene is set only under scene.has_root_pipeline() && rebuild_scene (scene_builder_thread.rs:1018). The offscreen pipeline's namespace open and rectangle adds still ship in the transaction's dl_updates.
  3. RenderBackend::process_transaction holds a transaction's dl_updates in held_dl_updates when !has_built_scene and the document exists (render_backend.rs:1307-1312, added in fc4f700fd37d).
  4. The offscreen loop (render_backend.rs:1411-1423) frame-builds txn.offscreen_scenes against &self.dl_stores regardless, so dl_stores.rectangles[DlHandle(5, 2)] finds namespace 5 not open or slot 2 empty, and panics.

This fits the crash reason bad dl store lookup DlHandle(5, 2) (only DlStore can print it, and rectangles is the only DlStore in this build). It also fits frame #10, which is the inlined Document::process_offscreen_scene call at render_backend.rs:1417. The older interners already get an offscreen-only flush at scene_builder_thread.rs:1065, while dl_updates get none. Not observed: that this transaction actually built no main scene. 2a53f18f1858 is most likely what exposed the hold gap from fc4f700fd37d by giving the store its first frame-time reader. Look first at the has_built_scene condition in render_backend.rs:1307: it should also apply held and current dl_updates when txn.offscreen_scenes is non-empty, keeping the closes deferred.

2a53f18f1858 landed 2026-10-19, the day before build 20261020195026. It is the third landing of bug 2078462 after backouts fd966ad73701 and cdb74ad03c32. Before it, dl_interner.rs carried #![allow(dead_code)] with a note that the store's read side had no caller, so no frame build could hit DlStore::index and this panic message could not occur. The signature name's 200-day age comes from the misattributed intern::impl$9::index inline, which earlier reports share for real intern::DataStore panics. It does not date this message. The bug 2079474 changesets in the window touch none of the store, render-backend hold or scene-builder files. Signature reports carry 0% bit-flip annotations, and this report's CPU (family 6 model 190) is not the defective Raptor Lake.

Suspected regressor: 2a53f18f1858 (gh) (bug 2078462) by :gw.

Code references:

What the automated skeptic pass checked

In its own words (a pass means the check succeeded, which is not always support for the conclusion):

  • unverifiable mechanism — The hold-without-built-scene path (render_backend.rs:1307-1312, fc4f700fd37d) and the unconditional offscreen frame build (render_backend.rs:1411-1423) are both confirmed in pinned source, and frame #10 is that offscreen call. Nothing in the dump shows directly that this transaction had built_scene == None or that namespace 5 was among the held opens.
  • pass C1-panic-origin — The panic string exists only in DlStore::uid (dl_interner.rs:742) and DlStore::index (:782). enumerate_dl_stores! lists only rectangles at the build rev, so the NormalBorderPrim / intern::impl$9::index inline in the signature is a misattribution.
  • pass hunk0 — visibility.rs:727 -> snap_policy -> prim_has_anti_aliasing -> as_common_data -> &dl_stores.rectangles[handle].common (render_backend.rs:341). Blame at b268695b69da puts lines 289/296/340/341 and the &self.dl_stores argument at 1420 on 2a53f18f1858.
  • pass alt-backdrop-filter — Bug 2079474 (ea745deba60c, fb3b55d32e19, 17b22abfced1, 59fe71d7c538) is a backdrop-filter rewrite. It touches neither dl_interner.rs, render_backend.rs nor scene_builder_thread.rs, and cannot produce a DlStore lookup panic. Ruled out as the cause.

Filed as a new bug rather than a comment on bug 2067323 — which is open on this signature, but it was filed before the changeset above landed, so it cannot be about this regression. Please duplicate if that is wrong.

:gw, can you have a look please?

Filed automatically by Clouseau, which analyses nightly crashes with an LLM. Nothing above was written or checked by a human. Please close it as INVALID if it is wrong — that is useful feedback, not a nuisance.

This is not an independent Mozilla discovery: the crash is known here only because somebody submitted the crash report linked at the top. If this duplicates an existing report, please resolve THIS bug as the duplicate and leave the credit with the earlier reporter.

Flags: needinfo?(mozilla)

Since nightly and release are affected, beta will likely be affected too.
For more information, please visit BugBot documentation.

Set release status flags based on info from the regressing bug 2078462

The crash report
https://crash-stats.mozilla.org/report/index/46666b0e-ff91-4cda-a6b0-0d35b0261021
logged a device reset ("Present1 failed: 0x887a0005", "Handle DeviceReset")
4 seconds before the crash, and almost all crashes with this signature are on
YouTube. A device reset recreates WebRender the same way a GPU process restart
does, so the browser test here likely tracks the real crashes.

After WebRender is recreated, the root WebRenderBridgeParent sets the root
pipeline, but that pipeline has no display list until chrome paints again. If
content captures a view transition in that window it sends a transaction with
a temporary pipeline's display list, RenderOffscreen and RemovePipeline. The
scene builder builds the offscreen scene but no main scene, because the
document's root pipeline has no display list. The render backend holds the
dl_updates of a transaction that built no scene, so the temporary pipeline's
namespace is never opened and its rectangles are never inserted into the
render backend's dl stores. It frame-builds the offscreen scene against those
stores anyway, and the first rectangle in the snapshot picture panics in
DlStore::index. In debug builds the stale handle assert in DlStore::entry
fires first.

Offscreen frames are built with present = false, which skips the picture
cache visibility pass, so only snapshot pictures read the store. The offscreen
display list needs a snapshot to crash, which a view transition capture has.

The wrench raw test sends that transaction to a document whose root pipeline
has no display list. The browser test gets there in Gecko: it pauses the
chrome refresh driver, restarts the GPU process, and starts view transitions
in content before chrome paints again. Both crash with the stack from the
crash report.

A transaction that builds no main scene holds its dl ops back from the
stores, but its offscreen scenes were built against them. Frame-build
those against a copy of the stores with the ops applied; applying them to
the stores early could remove entries the current scene still reads.

Flags: needinfo?(mozilla)
Assignee: nobody → mozilla
Status: NEW → ASSIGNED
Severity: -- → S3
Status: ASSIGNED → RESOLVED
Closed: 26 days ago
Resolution: --- → FIXED
Target Milestone: --- → 160 Branch
Attachment #9665872 - Attachment is obsolete: true
You need to log in before you can comment on or make changes to this bug.

Attachment

General

Created:
Updated:
Size: