Bug 235384 – AXIsolatedTree::updateChildren childrenIDs and children local variables could get out of sync

• Status: RESOLVED
• Resolution: FIXED
• Component: Accessibility

[Tyler Wilcock]

In AXIsolatedTree::updateChildren, we have this:

const auto& axChildren = axAncestor->children();
auto axChildrenIDs = axAncestor->childrenIDs();

Because the current version of AXCoreObject::childrenIDs always updates the underlying children if necessary, these two variables could get out of sync if childrenIDs actually performs an update after we already got children().

[Radar WebKit Bug Importer]

<rdar://problem/87791765>

[Tyler Wilcock]

Created attachment 449519 [details]
Patch

[Andres Gonzalez]

(In reply to Tyler Wilcock from comment #2)
> Created attachment 449519 [details]
> Patch

No need to add a bool param to ChildrenIDs since it is never used with value true. Keeping it simple, we can say that childrenIDs returns the IDs of the current children.

[Tyler Wilcock]

(In reply to Andres Gonzalez from comment #3)
> (In reply to Tyler Wilcock from comment #2)
> > Created attachment 449519 [details]
> > Patch
> 
> No need to add a bool param to ChildrenIDs since it is never used with value
> true. Keeping it simple, we can say that childrenIDs returns the IDs of the
> current children.
Well, it's used in places outside this patch in ways that do expect childrenIDs to update if necessary. For example, this:

void AXIsolatedTree::updateNode(AXCoreObject& axObject)
{
    ...truncated...
    auto newObject = AXIsolatedObject::create(axObject, this, parentID);
    newObject->m_childrenIDs = axObject.childrenIDs();
    ...truncated...
}

I did try to make childrenIDs be non-updating always, but that causes ~10 additional ITM failures. So I opted to give childrenIDs() the same bool API that children() has and fix the definitely wrong usage of it in AXIsolatedTree::updateChildren.

[Andres Gonzalez]

(In reply to Tyler Wilcock from comment #4)
> (In reply to Andres Gonzalez from comment #3)
> > (In reply to Tyler Wilcock from comment #2)
> > > Created attachment 449519 [details]
> > > Patch
> > 
> > No need to add a bool param to ChildrenIDs since it is never used with value
> > true. Keeping it simple, we can say that childrenIDs returns the IDs of the
> > current children.
> Well, it's used in places outside this patch in ways that do expect
> childrenIDs to update if necessary. For example, this:
> 
> void AXIsolatedTree::updateNode(AXCoreObject& axObject)
> {
>     ...truncated...
>     auto newObject = AXIsolatedObject::create(axObject, this, parentID);
>     newObject->m_childrenIDs = axObject.childrenIDs();
>     ...truncated...
> }
> 
> I did try to make childrenIDs be non-updating always, but that causes ~10
> additional ITM failures. So I opted to give childrenIDs() the same bool API
> that children() has and fix the definitely wrong usage of it in
> AXIsolatedTree::updateChildren.

OK, thanks.

[Tyler Wilcock]

Thank you both for the review!

[EWS]

Committed r288313 (246229@main): <https://commits.webkit.org/246229@main>

All reviewed patches have been landed. Closing bug and clearing flags on attachment 449519 [details].

[Darin Adler]

Comment on attachment 449519 [details]
Patch

View in context: https://bugs.webkit.org/attachment.cgi?id=449519&action=review

> Source/WebCore/accessibility/AccessibilityObjectInterface.h:1612
> +    auto& kids = children(updateChildrenIfNecessary);

It’s more elegant to use Vector::map instead of a loop and easier to get the vector size allocation right; something to consider in future.

> Source/WebCore/accessibility/AccessibilityObjectInterface.h:1614
> +    childrenIDs.reserveCapacity(kids.size());

This should use reserveInitialCapacity, slightly more efficient.

> Source/WebCore/accessibility/AccessibilityObjectInterface.h:1615
> +    for (const auto& child : kids)

This can just use auto&, no need for the const.

> Source/WebCore/accessibility/AccessibilityObjectInterface.h:1616
>          childrenIDs.append(child->objectID());

This can use the slightly more efficient uncheckedAppend because of the reserveCapacity above.