When we restart a transaction because of an incompatible server, pick a new transaction ID.
This should increase compatibility with DNS servers that don't like if they get different requests with the same
transaction ID.
given that DNSSEC lookups may result in quite a number of auxiliary transactions, let's better be safe than sorry and
also enforce a limit on the number of total transactions, not just on the number of queries.
Previously, when getting notified about a completed auxiliary DNSSEC transaction we'd immediately act on it, and
possibly abort the main transaction. This is problematic, as DNS transactions that already completed at the time we
started using them will never get the notification event, and hence never be acted on in the same way.
Hence, introduce a new call dns_transaction_dnssec_ready() that checks the state of auxiliary DNSSEC transactions, and
returns 1 when we are ready for the actual DNSSEC validation step. Then, make sure this is invoked when the auxiliary
transactions are first acquired (and thus possibly reused) as well when the notifications explained above take place.
This fixes problems particularly when doing combined A and AAAA lookups where the auxiliary DNSSEC transactions get
reused between them, and where we got confused if we reused an auxiliary DNSSEC transaction from one when it already
got completed from the other.
Let's avoid thinking that a CNAME/DNAME chain traversal could be a good idea if QTYPE is already CNAME/DNAME.
(Also, let's bail out early when trying to see if some RR is a suitable CNAME/DNAME for some other RR).
Move IDNA logic out of the normal domain name processing, and into the bus frontend calls. Previously whenever
comparing two domain names we'd implicitly do IDNA conversion so that "pöttering.de" and "xn--pttering-n4a.de" would be
considered equal. This is problematic not only for DNSSEC, but actually also against he IDNA specs.
Moreover it creates problems when encoding DNS-SD services in classic DNS. There, the specification suggests using
UTF8 encoding for the actual service name, but apply IDNA encoding to the domain suffix.
With this change IDNA conversion is done only:
- When the user passes a non-ASCII hostname when resolving a host name using ResolveHostname()
- When the user passes a non-ASCII domain suffix when resolving a service using ResolveService()
No IDNA encoding is done anymore:
- When the user does raw ResolveRecord() RR resolving
- On the service part of a DNS-SD service name
Previously, IDNA encoding was done when serializing names into packets, at a point where information whether something
is a label that needs IDNA encoding or not was not available, but at a point whether it was known whether to generate a
classic DNS packet (where IDNA applies), or an mDNS/LLMNR packet (where IDNA does not apply, and UTF8 is used instead
for all host names). With this change each DnsQuery object will now maintain two copies of the DnsQuestion to ask: one
encoded in IDNA for use with classic DNS, and one encoded in UTF8 for use with LLMNR and MulticastDNS.
How to reproduce
$ systemctl set-default multi-user # https://github.com/systemd/systemd/issues/2298
$ systemctl preset-all
Failed to execute operation: Too many levels of symbolic links
$ systemctl poweroff
Fixes:
==1==
==1== HEAP SUMMARY:
==1== in use at exit: 65,645 bytes in 7 blocks
==1== total heap usage: 40,539 allocs, 40,532 frees, 30,147,547 bytes allocated
==1==
==1== 109 (24 direct, 85 indirect) bytes in 1 blocks are definitely lost in loss record 2 of 7
==1== at 0x4C2BBCF: malloc (in /usr/lib/valgrind/vgpreload_memcheck-amd64-linux.so)
==1== by 0x4C2DE2F: realloc (in /usr/lib/valgrind/vgpreload_memcheck-amd64-linux.so)
==1== by 0x23DA71: unit_file_changes_add (install.c:233)
==1== by 0x23E45D: remove_marked_symlinks_fd (install.c:453)
==1== by 0x23E267: remove_marked_symlinks_fd (install.c:405)
==1== by 0x23E641: remove_marked_symlinks (install.c:494)
==1== by 0x243A91: execute_preset (install.c:2190)
==1== by 0x244343: unit_file_preset_all (install.c:2351)
==1== by 0x18AAA2: method_preset_all_unit_files (dbus-manager.c:1846)
==1== by 0x1D8157: method_callbacks_run (bus-objects.c:420)
==1== by 0x1DA9E9: object_find_and_run (bus-objects.c:1257)
==1== by 0x1DB02B: bus_process_object (bus-objects.c:1373)
==1==
==1== LEAK SUMMARY:
==1== definitely lost: 24 bytes in 1 blocks
==1== indirectly lost: 85 bytes in 1 blocks
==1== possibly lost: 0 bytes in 0 blocks
==1== still reachable: 65,536 bytes in 5 blocks
==1== suppressed: 0 bytes in 0 blocks
==1== Reachable blocks (those to which a pointer was found) are not shown.
==1== To see them, rerun with: --leak-check=full --show-leak-kinds=all
==1==
==1== For counts of detected and suppressed errors, rerun with: -v
==1== ERROR SUMMARY: 1 errors from 1 contexts (suppressed: 0 from 0)
* Use $ROOTLIBDIR/systemd always
* Don't pass $ROOTLIBDIR/systemd as the first argument:
$ cat /proc/1/cmdline
/lib/systemd/systemd/lib/systemd/systemd...
If we have a signed DNAME RR response, there's no need to insist on a signature for a CNAME RR response, after all it
is unlikely to be signed, given the implicit synthethis of CNAME through DNAME RRs.
The functionality of SYSTEMD_USER_WANTS that attaches dependencies to device
units from udev rules was broken since commit b2c23da8. I guess it was due to
a mass replace s/SYSTEMD_USER/MANAGER_USER/.
Quite often we read the same RR key multiple times from the same message. Try to replace them by a single object when
we notice this. Do so again when we add things to the cache.
This should reduce memory consumption a tiny bit.
In some cases we learn something about a server's feature level through its responses. If we notice that after doing
basic checking of a response, and after collecting all auxiliary DNSSEC info the feature level of the server is lower
than where we started, restart the whole transaction.
This is useful to deal with servers that response rubbish when talked to with too high feature levels.
Let's make sure we first check if the OPT was lost in the reply, before we accept a reply as successful and use it for
verifying the current feature level.
When we restart a DNS transaction, remove all connections to any auxiliary DNSSEC transactions, after all we might
acquire completely different data this time, requiring different auxiliary DNSSEC transactions.
If we receive a reply that lacks the OPT RR, then this is reason to downgrade what was verified before, as it's
apparently no longer true, and the previous OPT RR we saw was only superficially OK.
Similar, if we realize that RRSIGs are not augmented, then also downgrade the feature level that was verified, as
DNSSEC is after all not supported. This check is in particular necessary, as we might notice the fact that RRSIG is not
augmented only very late, when verifying the root domain.
Also, when verifying a successful response, actually take in consideration that it might have been reported already
that RRSIG or OPT are missing in the response.
This adds logic to downgrade the feature level more aggressively when we have reason to. Specifically:
- When we get a response packet that lacks an OPT RR for a query that had it. If so, downgrade immediately to UDP mode,
i.e. don't generate EDNS0 packets anymore.
- When we get a response which we are sure should be signed, but lacks RRSIG RRs, we downgrade to EDNS0 mode, i.e.
below DO mode, since DO is apparently not really supported.
This should increase compatibility with servers that generate non-sensical responses if they messages with OPT RRs and
suchlike, for example the situation described here:
https://open.nlnetlabs.nl/pipermail/dnssec-trigger/2014-November/000376.html
This also changes the downgrade code to explain in a debug log message why a specific downgrade happened.
This validates OPT RRs more rigorously, before honouring them: if we any of the following condition holds, we'll ignore
them:
a) Multiple OPT RRs in the same message
b) OPT RR not owned by the root domain
c) OPT RR in the wrong section (Belkin routers do this)
d) OPT RR contain rfc6975 algorithm data (Belkin routers do this)
e) OPT version is not 0
f) OPT payload doesn't add up with the lengths
Note that d) may be an indication that the server just blindly copied OPT data from the response into the reply.
RFC6975 data is only supposed to be included in queries, and we do so. It's not supposed to be included in responses
(and the RFC is very clear on that). Hence if we get it back in a reply, then the server probably just copied the OPT
RR.
This new test case tries to resolve a couple of known domains, to verify the validation results. It talks to resolved
via the bus, thus comprehensively testing the whole shebang.
Of course, it requires network connectivity and a DNSSEC capable DNS server, hence this is a manual test.
This fills in the last few gaps:
- When checking if a domain is non-existing, also check that no wildcard for it exists
- Ensure we don't base "covering" tests on NSEC RRs from a parent zone
- Refuse to accept expanded wildcard NSEC RRs for absence proofs.
empty non-terminals generally lack NSEC RRs, which means we can deduce their existance only from the fact that there
are other RRs that contain them in their suffix. Specifically, the NSEC proof for NODATA on ENTs works by sending the
NSEC whose next name is a suffix of the queried name to the client. Use this information properly.
Having this information available is useful when we need to check whether various RRs are suitable for proofs. This
information is stored in the RRs as number of labels to skip from the beginning of the owner name to reach the
synthesizing source/signer. Simple accessor calls are then added to retrieve the signer/source from the RR using this
information.
This also moves validation of a a number of RRSIG parameters into a new call dnssec_rrsig_prepare() that as side-effect
initializes the two numeric values.
When proving NODATA DS lookups we need to insist on looking at the parent zone's NSEC RR, not the child zone's.
When proving any other NODATA lookups we need to insist on looking at the child zone's NSEC RR, not the parent's.
