User-Agent: Mozilla/5.0 (X11; U; Linux i686; en-US; rv:220.127.116.11) Gecko/2009042708 Fedora/3.0.10-1.fc10 Firefox/3.0.10 Build Identifier: Mozilla/5.0 (X11; U; Linux i686; en-US; rv:18.104.22.168) Gecko/2009042708 Fedora/3.0.10-1.fc10 Firefox/3.0.10 Filing under Networking because the ultimate bug is in a missing res_init() (I suppose) somewhere. When using GSSAPI negotiate authentication, a failure may occur if /etc/resolv.conf has changed (eg. you connected to a VPN network). Looking at network traces queries about the host we are connecting to shows that the main engine is correctly using the new DNSs while the gssapi resolve calls are still going to the old DNS server specified in resolv.conf Apparently a res_init() call is missing somewhere. Setting: export NSPR_LOG_MODULES= negotiateauth:4 I see in the logs that the failing call is: gss_init_se_context_ptr() in nsAuthGSSAPI::GetNextToken() It fails with: 245614352[7fae0e754040]: gss_init_sec_context() failed: An invalid name was supplied Hostname cannot be canonicalized this happens in mozilla-1.9.1/extensions/auth/nsAuthGSSAPI.cpp line 463 Although I can't tell whether it is ok to just add a res_init() call before it or if res_init() should be triggered elsewhere. HTH Reproducible: Sometimes Steps to Reproduce: The best way I could reproduce this failure was to connect to my copmany's network over a VPN, kinit and auth to an internal web site using GSSAPI negotiate auth. Then let the ticket expire and disconnect from the network. Wait till the next day, try again to connect and see that I fail to connect (can't resolve the host name at all). Reconnect to the VPN (this changes /etc/resolve.conf to query internal DNS servers), kinit again and reload the page and see that firefox now finds the web server, but auth fails because the gssapi queries are still going out to the old DNS server. Actual Results: SSO auth fails. Expected Results: Auth works. As a work around you can save&quit firefox and restart it, but that's annoying.
Also seeing this: Mozilla/5.0 (X11; U; Linux x86_64; en-US; rv:22.214.171.124) Gecko/20090803 Fedora/3.5.2-2.fc11 Firefox/3.5.2
The original testcase doesn't work for me reliably. However, the following does: - start firefox with vpn disconnected - load any public page, and firefox will use the public dns - connect the vpn, which will change the dns settings - try kerberos auth and it fails, despite having a valid ticket
Firefox never calls res_init and thus never refreshes dns config explicitly. There is a single reference to this function in the Mozilla code, within a comment in nsHostResolver.cpp: "Use a persistent thread pool in order to avoid spinning up new threads all the time. In particular, thread creation results in a res_init() call from libc which is quite expensive."
Ok, res_init is deprecated, one should use res_ninit instead. I see that Mozilla indeed calls res_ninit on failures, at most once per second. This is why we actually succeed loading the intranet page after vpn connect. However, the kerberos service code (contained in system libs) doesn't seem to do that, but probably calls the DNS resolver directly. Mozilla's res_ninit() calls inside the host-resolver code don't help us, as Mozilla's resolver runs on its own separate thread. This means, on the thread where we call the gssapi functions, Mozilla never calls res_ninit, and the system gssapi implementation apparently doesn't do it on its own. I can confirm that a call to res_ninit(), just before calling gss_import_name, fixes this bug. I'm not yet sure, is it fine to add this directly to the extensions/auth code. In Linux res_ninit is implemented inside libc. On other systems, could it be implemented by some networking lib, that only gets linked by Mozilla's libnecko XPCOM module? If this may happen, then we'd need to implement/extend a service inside necko for the call to res_ninit, and have extension/auth call this service remotely.
Created attachment 405350 [details] [diff] [review] Patch v1 This is the patch I used and tested on Linux. As explained, I don't know whether res_ninit can be found on all platforms using the libs currently referenced from libauth, or whether some platforms require libs referenced by necko, only.
8 years ago
Comment on attachment 405350 [details] [diff] [review] Patch v1 Proposing correctness fix for stable branches.
reopening, patch broke the tree...
Created attachment 408486 [details] [diff] [review] Patch v1 fixed Sigh, how comes I added an #ifdef to include, but failed to use an #ifdef to the actual function call? Thanks to Stephen Gallagher and blassey for their messages who suggested the same fix. Attaching new patch, which built fine on TryServer.
Second checkin attempt: http://hg.mozilla.org/mozilla-central/rev/c814dffbd980
Comment on attachment 408486 [details] [diff] [review] Patch v1 fixed This checkin attempt succeeded, tinderboxes look good. Nominating for stable branches.
Comment on attachment 408486 [details] [diff] [review] Patch v1 fixed We'll look at this for the next cycle, after it's gotten baking on 1.9.2. Kai: You should ping a 1.9.2 driver for approval.
kaie: you'll need to lobby for the bug on 1.9.2 through email or IRC, the approval requests for non=blockers are getting missed.
Comment on attachment 408486 [details] [diff] [review] Patch v1 fixed Approved for 126.96.36.199 and 188.8.131.52, a=dveditz for release-drivers
Checking in extensions/auth/nsAuthGSSAPI.cpp; /cvsroot/mozilla/extensions/auth/nsAuthGSSAPI.cpp,v <-- nsAuthGSSAPI.cpp new revision: 1.15; previous revision: 1.14
Comment on attachment 408486 [details] [diff] [review] Patch v1 fixed Approved for 184.108.40.206, a=dveditz for release-drivers