First step for T102367 is to find unsecure resources.
How do we do this? Options I see:
| operations/puppet : production | dynamicproxy: custom log schema (http/https) for tools |
| Status | Assigned | Task | ||
|---|---|---|---|---|
| Open | None | T131288 Make labs proxies https only | ||
| Open | None | T102367 Migrate tools.wmflabs.org to https only (and set HSTS) | ||
| Resolved | • bd808 | T128409 Detect tools.wmflabs.org tools which are HTTP-only |
I don't think it is feasible to detect those tools this way because there would be way too many code paths that are only triggered under special circumstances. What is easier:
step 0. figure out which instance(s) are "on the proxy"
step -1. what is the relevant role class? i think "dynamicproxy" module but no role? is "description="Kubernetes to dynamicproxy syncronizer" involved?
The proxy are the instances tools-proxy-01 and tools-proxy-02. They use the role class role::labs::tools::proxy which includes toollabs::proxy which includes dynamicproxy which uses modules/dynamicproxy/templates/urlproxy.conf for configuration. So in the latter, some directives to log the scheme would need to be added.
Change 274161 had a related patch set uploaded (by Dzahn):
dynamicproxy: custom log schema (http/https) for tools
@scfc thank you for the details. i uploaded a WIP patch. can i use gzip like that? log path ok? how long do we want to run it,hmm.. also comments on gerrit
Change 274161 merged by Dzahn:
dynamicproxy: custom log schema (http/https) for tools
works now:)
root@tools-proxy-01:~# tail -f /var/log/nginx/access-scheme.log
shows first results
here's a first list of tools using http, status 200
add-information admin anagrimes anomiebot archaeo autodesc autolist awb bambots bawolff betacommand-dev bibleversefinder book2scroll catfood catnap catscan2 catscan3 checkdictation-fa checker checkwiki citationhunt citations cluebot commonshelper copyvios croptool crosswatch denkmalliste dewikinews-rss de.wikipedia.org dexbot dibot dispenser dnbtools dplbot dupdet dykfeed dykstats enwp10 eranbot erwin85 expose-data farhangestan fengtools filedupes file-siblings fist flickr2commons ftl geocommons geograph2commons geohack glamtools guc heritage hewiki-tools hoo icalendar icommons ifttt imagemapedit intuition ircredirect isbn isbn2wiki isin itwikinews-rss jackbot jira-bugimport joanjoc kasparbot kmlexport krdbot languagetool listeria locator magnustools magnus-toolserver magog makeref manypedia matvaretabellen meta metamine missingtopics mix-n-match most-wanted multidesc musikanimal nagf newwebtest ninihil not-in-the-other-language nullzero os osm osm4wiki otrsreports pagepile pageviews para parliamentdiagram paste persondata phabricator-bug-status phetools pltools potd-feed projektneuheiten-feed ptwikis pywikibot quentinv57-tools quick-intersection random-featured reasonator refill reftoolbar sdbot seealsology sighting sigma skipenwp10 slumpartikel spellcheck static stewardbots stimmberechtigung suggestbot supercount svgtranslate templatecount templatetiger templatetransclusioncheck templator test-lighttpd-precise test-lighttpd-trusty test-webservice-generic timescale toolschecker toolscript tusc typoscan ukbot url2commons usersearch usualsuspects vcat video2commons videoconvert w watchr whois widar wiki wikidata-exports wikidata-game wikidata-primary-sources wikidata-reconcile wikidata-terminator wikidata-todo wikiedudashboard wikihistory wikisense wikishootme wikisoba wikitest-rtl wiki-todo wikitrends wikiviewstats wikiviewstats2 wikivoyage wikiwatchdog wiwosm wlm-de-utils wlm-stats wmcounter wp-signpost wp-world wsexport xtools xtools-articleinfo xtools-ec yifeibot ytcleaner zoomable-images zoomviewer
@Magnus: In this context, it means that a request came in with http and was not redirected or failed. This could be by someone typing in the URL manually, a link from Wikipedia/etc. or an intra-application link.
This seems not particularly useful as a data point, since it is just counting:
By "this" I assume you mean the list above? I'd like more comments on the methods proposed in the description.
- make a list of tools.wmflabs.org URLs and test them all for unsecure resources with a simple URL fetching script;
- some smart ruby mechanize crawler to test all the domain recursively.
These are not unlikely to kill tool labs or the replicas, as many tools are not scraping-friendly, and happily send off long sql queries from a GET request.
- set some dark magic JavaScript site-wide logging sending all such occurrences of unsecure resources to Sentry or something;
We can probably do this with CSP: T130748: Add Content-Security-Policy header enforcing 3rd party web interaction restrictions to proxy responses
- naively grep all the projects' PHP and JavaScript code looking for hardcoded http:// URLs;
Sounds like a good idea to me (but probably on the NFS server and not on a bastion ;-))
As a side note, I set up a VM for my PetScan tool:
http://petscan.wmflabs.org/
This does not require http, but works for either, as I couldn't find a way to forward http to https on the proxy level.
Do I need to do that within the server I run?
@Magnus: https://petscan.wmflabs.org/ seems to work fine. Did you mean something else?
In Tools the tool never 'sees' http because everything goes through the proxy:
So all tools should be compatible if http is redirected to https, as per @Magnus's suggestion.
The problem is on the client side. There are probably many prominent non-browser client tools of these services that will fail to follow the HTTPS redirect on GET/HEAD (just fail to follow a redirect in general, or fail to have proper HTTPS support).
For clients that mostly use POST the problem is worse, as there's no hope for a universally-compatible secure redirect method for HTTP POST to HTTPS, and even if you had one, the client has already leaked critical POST data in the initial request before the redirect happens. The best you can do in terms of enforcement there is 403 the insecure POST requests and break the clients to make it obvious they need to switch to HTTPS URLs. That's a big breaking change that needs announcement and will make users hate you (which is something we're facing with production APIs, too).
What you can do to help modern browsers, though, without taking the redirect and/or POST-breakage risks, is universally add a long-duration HSTS header so that they lock onto HTTPS if they've ever visited once via HTTPS.
They do get an X-Forwarded-Proto header, and that's how video2commons currently does the detection and redirect.
I imagine most tools should use relative uris within the tool (because it could take more time to type/code with absolute urls), so http: posts go to http:, https: posts go to https:, so they should be okay.
AFAIK, the problems are:
We have been sending out Strict-Transport-Security: max-age:86400 headers following a 302 redirect for GET and HEAD requests for nearly 3 months now. There have been no widespread reports of "OMG! Tool <x> is totally broken" so I think we can actually call this resolved. There is still the matter of POST requests to deal with, but that is a matter of enforcement rather than detection.