Bug 35373 – XSSAuditor is super super super slow

• Component: New Bugs
• Status: RESOLVED FIXED

[James Robinson]

A page that contained a large number of <input> elements with inline onblur handlers and that POSTs a lot of data is super super super slow with the XSSAuditor enabled.  This is a really huge perf regression.

[James Robinson]

Created attachment 49456 [details]
shark profile of load

[James Robinson]

Created attachment 49457 [details]
test page (PHP)

Test page, from the original Chromium bug.  It has directions for running it in the page.  A bit messy but shows the problem clearly.

[James Robinson]

Here are the two really bad callstacks on the second load:

	0.0%	91.8%	WebCore	                WebCore::HTMLTokenizer::write(WebCore::SegmentedString const&, bool)
	0.1%	91.7%	WebCore	                 WebCore::HTMLTokenizer::parseTag(WebCore::SegmentedString&, WebCore::HTMLTokenizer::State)
	0.0%	91.5%	WebCore	                  WebCore::HTMLTokenizer::processToken()
	0.0%	91.5%	WebCore	                   WebCore::HTMLParser::parseToken(WebCore::Token*)
	0.0%	90.3%	WebCore	                    WebCore::Element::setAttributeMap(WTF::PassRefPtr<WebCore::NamedNodeMap>, WebCore::FragmentScriptingPermission)
	0.0%	90.2%	WebCore	                     WebCore::StyledElement::attributeChanged(WebCore::Attribute*, bool)
	0.0%	58.2%	WebCore	                      WebCore::HTMLInputElement::parseMappedAttribute(WebCore::MappedAttribute*)
	0.0%	58.1%	WebCore	                       WebCore::HTMLTextFormControlElement::parseMappedAttribute(WebCore::MappedAttribute*)
	0.0%	51.5%	WebCore	                        WebCore::createAttributeEventListener(WebCore::Node*, WebCore::Attribute*)
	0.0%	51.5%	WebCore	                         WebCore::XSSAuditor::canCreateInlineEventListener(WebCore::String const&, WebCore::String const&) const

and

	0.0%	91.8%	WebCore	                WebCore::HTMLTokenizer::write(WebCore::SegmentedString const&, bool)
	0.1%	91.7%	WebCore	                 WebCore::HTMLTokenizer::parseTag(WebCore::SegmentedString&, WebCore::HTMLTokenizer::State)
	0.0%	91.5%	WebCore	                  WebCore::HTMLTokenizer::processToken()
	0.0%	91.5%	WebCore	                   WebCore::HTMLParser::parseToken(WebCore::Token*)
	0.0%	90.3%	WebCore	                    WebCore::Element::setAttributeMap(WTF::PassRefPtr<WebCore::NamedNodeMap>, WebCore::FragmentScriptingPermission)
	0.0%	90.2%	WebCore	                     WebCore::StyledElement::attributeChanged(WebCore::Attribute*, bool)
	0.0%	58.2%	WebCore	                      WebCore::HTMLInputElement::parseMappedAttribute(WebCore::MappedAttribute*)
	0.0%	25.5%	WebCore	                      WebCore::HTMLSelectElement::parseMappedAttribute(WebCore::MappedAttribute*)
	0.0%	25.5%	WebCore	                       WebCore::createAttributeEventListener(WebCore::Node*, WebCore::Attribute*)
	0.0%	25.4%	WebCore	                        WebCore::XSSAuditor::canCreateInlineEventListener(WebCore::String const&, WebCore::String const&) const

They don't make sense to me - why would the XSSAuditor have to check anything coming out of the HTMLTokenizer?  This is HTML served up by the server, not coming from any user-controlled data.

[Dimitri Glazkov (Google)]

Are you sure this warrants three supers? It looks only super super slow.

[Adam Barth]

> They don't make sense to me - why would the XSSAuditor have to check anything
> coming out of the HTMLTokenizer?  This is HTML served up by the server, not
> coming from any user-controlled data.

The XSS auditor is checking whether the event handlers are present in the post request.  If they are, it gets suspicious that there might be a reflected XSS attack.

I need to spend some time with this in the debugger to see how we can speed this up.  In the worse case, we might need to build some sort of index over the post data to make substring queries faster.

[James Robinson]

Created attachment 49513 [details]
Patch

[James Robinson]

While you're thinking, could you please review this patch?

[Adam Barth]

Comment on attachment 49513 [details]
Patch

If you want to disable the auditor, we can do that via the Settings object.  In particular, you can look in Chromium at how the --disable-xss-auditor command line switch is implemented.

[Adam Barth]

Patch in hand, benchmarking.

[Adam Barth]

Created attachment 49593 [details]
Patch

[Darin Adler]

Comment on attachment 49593 [details]
Patch

This patch needs a "why" comment explaining why it's important to use two caches.

[Darin Adler]

Comment on attachment 49593 [details]
Patch

The code, not the change log.

[Adam Barth]

I've added the following comment to the header:

        // A state store to help us avoid canonicalizing the same URL repeated.
        // When a page has form data, we need two caches: one to store the
        // canonicalized URL and another to store the cannonicalized form
        // data. If we only had one cache, we'd always generate a cache miss
        // and load some pages extremely slowly.
        // https://bugs.webkit.org/show_bug.cgi?id=35373

We're still slower on loading this page by a factor of four, but that's better than the factor of infinity slower we were before this patch.  (Ok, it's not really infinity, I just got bored waiting for the timer to finish.)

[Adam Barth]

http://trac.webkit.org/changeset/55290