On 12 Nov 2024, at 9:41, Chuck Lever wrote: > On Tue, Nov 12, 2024 at 09:27:45AM -0500, Benjamin Coddington wrote: >> On 11 Nov 2024, at 17:49, Philip Rowlands wrote: >> >>> If a host dies after nsm_make_temp_pathname but before rename(temp, path) we may be left with paths resembling .../server.example.com.new >>> >>> Some clever person has registered and installed a wildcard DNS record for *.com.new. >>> >>> $ host server.example.com.new >>> server.example.com.new has address 104.21.68.132 >>> server.example.com.new has address 172.67.195.202 >>> >>> You can see where this is going... >>> >>> Our firewall scanners tripped on outbound access to this address, port 111, I assume due to NSM reboot notifications. >>> >>> Suggested workarounds include: >>> * explicitly skip over paths matching the expect tempname pattern in nsm_load_dir() >>> * use a different tmp suffix than .new, e.g. one which won't work in DNS >>> >>> Steps to reproduce: >>> >>> # cat /var/lib/nfs/statd/sm/server.example.com.new >>> 0100007f 000186b5 00000003 00000010 89ae3382e989d91800000000dc00ed000000ffff 1.2.3.4 my-client-name >>> # sm-notify -d -f -n >>> sm-notify: Version 2.7.1 starting >>> sm-notify: Retired record for mon_name server.example.com.new >>> sm-notify: Added host server.example.com.new to notify list >>> sm-notify: Initializing NSM state >>> sm-notify: Failed to open /proc/sys/fs/nfs/nsm_local_state: No such file or directory >>> sm-notify: Effective UID, GID: 29, 29 >>> sm-notify: Sending PMAP_GETPORT for 100024, 1, udp >>> sm-notify: Added host server.example.com.new to notify list >>> sm-notify: Host server.example.com.new due in 2 seconds >>> sm-notify: Sending PMAP_GETPORT for 100024, 1, udp >>> # etc. >>> >>> tcpdump shows the outbound traffic: >>> 22:42:31.940208 IP 192.168.0.131.819 > 172.67.195.202.sunrpc: UDP, length 56 >>> 22:42:33.942440 IP 192.168.0.131.819 > 172.67.195.202.sunrpc: UDP, length 56 >>> 22:42:37.946903 IP 192.168.0.131.819 > 172.67.195.202.sunrpc: UDP, length 56 >>> >>> The client statd was artificially placed for the purposes of showing the problem, but I hope it's close enough to make sense. >> >> Makes sense.. yikes! >> >> Maybe we could just prepend '.' since nsm_load_dir() ignores those - Chuck, you were in here last any thoughts? > > The problem with a leading dot is, of course, the file becomes > hidden, which might be surprising to administrators who are trying > to diagnose a problem. I used to be one of those, and would say this isn't a big issue for any competent admin. It has another advantage of also never being a valid DNS name because it has an "empty label". > Note that a domain label can contain only the letters A-Z (or a-z), > the digits 0-9, hyphen (-), and dot (.). So replace ".new" with > something that contains an invalid character like ".<new>" Hmm.. I thought (goes to dig it up) that any binary string can serve as a name representation. https://datatracker.ietf.org/doc/html/rfc2181#section-11 ..that's been updated by a number of other RFCs.. (gah! - case insensitive comparisons!) I admit to not knowing or wanting to keep digging through RFCs for the current domain label specification. Do you have a current reference and feel like we can depend on it? Thanks for chiming in here! Ben
