Bug 171231

<rdar://problem/31800404>

I handled cache responses in bug 171394. I'll use this one to investigate why non-cached responses for cross-origin resources were somehow allowing these properties.

Created attachment 308453 [details] [IMAGE] Web Inspector - bbc.com Resource Timing Entries Hmm, I'm not able to reproduce this with trunk (and with my changes on bug 171394). See attached screenshot of: > console.table(performance.getEntriesByType("resource"), ["name", "fetchStart", "connectStart", "domainLookupStart"]) Performed on bbc.com. This was load without using the cache (Web Inspector shows all resources did not use the memory cache). In this case only a few resources had connect/dns information and each of those domains explicitly had Timing-Allow-Origin: "*". Each of the others are zero'd indicating they are cross-origin without TAO.

Ahh. So here is my theory. The same resource often shows up a bunch of times, so maybe when you ran: js> Array.from(performance.getEntriesByType("resource")).filter(r => r.connectEnd).map(r => r.name); You were seeing the cache responses in the resulting list (which will at the time had their connectStart/domainLookupStart falling back to fetchStart like a normal request). Those cache responses are being corrected, so now these false positives don't show up. Let me try to verify that.

Hmm, that doesn't seem to fully explain things. On STP 28 I was able to reproduce the issue with a single cross-origin resource that claimed to not be from a cache yet had non-zero'd data. However, I'm unable to reproduce that on my trunk builds, so it seems this will be fixed by the change on 171394.

Created attachment 308462 [details] List of resources

I ran the same test on cnn.com which I've never visited in STP 28. Many uncached resources had connectStart values and did not have a TAO header. I'm attaching some screenshots showing what I saw.

Created attachment 308463 [details] CSS file 1 - no header - non-zero timing

Created attachment 308464 [details] CSS file 2 - no header - non-zero timing

Created attachment 308465 [details] JS file 1 - no header - zeroed timing

Created attachment 308466 [details] JS file 2 - no header - zeroed timing

Yep, as I'm discovering in bug 171394 that code path might have been possible for ongoing requests. So the more I look into it, I think that should fix this issue.

Yes, I'm going to mark this as a duplicate of 171394. Sorry for unnecessarily creating an extra bug, I had just originally thought there were two distinct issues! I'm unable to reproduce after that change: <http://trac.webkit.org/changeset/215981> That change fixed two issues. (1) Loads from the cache for an earlier completed request. - The typical cache case. This will appear in Web Inspector as cached (from Memory or Disk cache) (2) Loads that would use the cache result for an "ongoing request". - This can happen internally within WebCore making multiple requests for a resource - This could eventually appear in Web Inspector as a Network request when the load actually completes These examples are of case (2). In cases where there was an ongoing request we were incorrectly creating a PerformanceResourceTiming entry before the request completed (and therefore before we knew if there was a TAO or not). So, the information for these requests, with and without TAO should be more accurate. -- Thanks for reporting this issue! We were reporting inaccurate / misleading in these cases. You should be able to verify in a WebKit Nightly, or a Safari Technology Preview in a few weeks. *** This bug has been marked as a duplicate of bug 171394 ***

Checked out in WebKit Nightly and things look good to me.