The report is from 2023: every bfcache back/forward navigation ran
CSSFontSelector::FontCacheInvalidated() on Windows, because
WebContentsImpl::DidFinishNavigation re-applies the same
WebPreferences and the font family setters reported a change
for unchanged values.
I probed that on current main first: re-applying identical preferences
through WebView::ApplyWebPreferences() four times in a row
produces invalidations only on the first apply. The list case the bug was
filed for was fixed by
b012decc740a9
("Tweak font family setting logic", 2024), which caches the first available
font per family list and compares against it.
One instance of the same bug survived that fix, on macOS.
GenericFontFamilySettings::UpdateFixed() compared the
incoming family against Fixed(), and Fixed()
translates a stored "Osaka" to "Osaka-Mono" on
macOS (crbug.com/40535332). The embedder passes "Osaka", or the
list that resolves to it, and the comparison is against
"Osaka-Mono", which it can never equal:
bool GenericFontFamilySettings::UpdateFixed(const AtomicString& family,
UScriptCode script) {
const AtomicString& old_family = Fixed(script); // "Osaka-Mono"
if (!ShouldUpdateFontFamily(old_family, family)) { // never false
return false;
}
So DevToolsEmulator::SetFixedFontFamily() calls
Settings::NotifyGenericFontFamilyChange(), which runs
Page::SettingsChanged(kFontFamily) →
StyleEngine::UpdateGenericFontFamilySettings() →
CSSFontSelector::FontCacheInvalidated() →
MarkFontsNeedUpdate() on every navigation, which is the stack
in the bug.
This is the configured default, not an exotic setting: the macOS fixed
font for Japanese is ,Osaka,BIZ UDGothic,Menlo in
locale_settings_mac.grd.
The fix compares against the stored family
(GenericFontFamilyForScript(fixed_font_family_map_, script))
instead. Fixed() keeps translating for its callers.
GenericFontFamilySettingsTest.OsakaFixedFontNotChanged is
guarded by BUILDFLAG(IS_MAC), so to see it fail on Linux I
built with the IS_MAC guard around the Osaka
translation removed:
second UpdateFixed("Osaka") | |
|---|---|
| without the fix | Actual: true, Expected: false |
| with the fix | false, test passes |
blink_platform_unittests --gtest_filter=*Font* and
blink_unittests --gtest_filter=*Font*:*Settings* pass. The
macOS bots run the real test; mac-rel and mac14-tests are on the CL.
An earlier version compared whole GenericFontFamilySettings
objects in CSSFontSelector::UpdateGenericFontFamilySettings()
and skipped the invalidation when they matched. That is a backstop at the
consumer, not a fix: the settings genuinely differ at the moment the
selector is called, since the map holds the list on one apply and the
resolved family after the next read. It also added an
operator== to a platform type for a single call site. The
setter is the thing that was wrong, so the fix is one line there.