This effectively reverts most of 51967f98 ("SERVFAIL is an expected
error return, don't try all servers.") and 4ace25c5 ("Treat REFUSED (not
SERVFAIL) as an unsuccessful upstream response").

With the current behaviour, as soon as dnsmasq receives a SERVFAIL from an
upstream server, it stops trying to resolve the query and simply returns
SERVFAIL to the client.  With this commit, dnsmasq will instead try to
query other upstream servers upon receiving a SERVFAIL response.

According to RFC 1034 and 1035, the semantic of SERVFAIL is that of a
temporary error condition.  Recursive resolvers are expected to encounter
network or resources issues from time to time, and will respond with
SERVFAIL in this case.  Similarly, if a validating DNSSEC resolver [RFC
4033] encounters issues when checking signatures (unknown signing
algorithm, missing signatures, expired signatures because of a wrong
system clock, etc), it will respond with SERVFAIL.

Note that all those behaviours are entirely different from a negative
response, which would provide a definite indication that the requested
name does not exist.  In our case, if an upstream server responds with
SERVFAIL, another upstream server may well provide a positive answer for
the same query.

Thus, this commit will increase robustness whenever some upstream servers
encounter temporary issues or are misconfigured.

Quoting RFC 1034, Section 4.3.1. "Queries and responses":

    If recursive service is requested and available, the recursive response
    to a query will be one of the following:

       - The answer to the query, possibly preface by one or more CNAME
         RRs that specify aliases encountered on the way to an answer.

       - A name error indicating that the name does not exist.  This
         may include CNAME RRs that indicate that the original query
	  name was an alias for a name which does not exist.

       - A temporary error indication.

Here is Section 5.2.3. of RFC 1034, "Temporary failures":

    In a less than perfect world, all resolvers will occasionally be unable
    to resolve a particular request.  This condition can be caused by a
    resolver which becomes separated from the rest of the network due to a
    link failure or gateway problem, or less often by coincident failure or
    unavailability of all servers for a particular domain.

And finally, RFC 1035 specifies RRCODE 2 for this usage, which is now more
widely known as SERVFAIL (RFC 1035, Section 4.1.1. "Header section format"):

    RCODE           Response code - this 4 bit field is set as part of
                    responses.  The values have the following
                    interpretation:
                    (...)

                    2               Server failure - The name server was
                                    unable to process this query due to a
                                    problem with the name server.

For the DNSSEC-related usage of SERVFAIL, here is RFC 4033
Section 5. "Scope of the DNSSEC Document Set and Last Hop Issues":

    A validating resolver can determine the following 4 states:
    (...)

    Insecure: The validating resolver has a trust anchor, a chain of
       trust, and, at some delegation point, signed proof of the
       non-existence of a DS record.  This indicates that subsequent
       branches in the tree are provably insecure.  A validating resolver
       may have a local policy to mark parts of the domain space as
       insecure.

    Bogus: The validating resolver has a trust anchor and a secure
       delegation indicating that subsidiary data is signed, but the
       response fails to validate for some reason: missing signatures,
       expired signatures, signatures with unsupported algorithms, data
       missing that the relevant NSEC RR says should be present, and so
       forth.
    (...)

    This specification only defines how security-aware name servers can
    signal non-validating stub resolvers that data was found to be bogus
    (using RCODE=2, "Server Failure"; see [RFC4035]).

Notice the difference between a definite negative answer ("Insecure"
state), and an indefinite error condition ("Bogus" state).  The second
type of error may be specific to a recursive resolver, for instance
because its system clock has been incorrectly set, or because it does not
implement newer cryptographic primitives.  Another recursive resolver may
succeed for the same query.

There are other similar situations in which the specified behaviour is
similar to the one implemented by this commit.

For instance, RFC 2136 specifies the behaviour of a "requestor" that wants
to update a zone using the DNS UPDATE mechanism.  The requestor tries to
contact all authoritative name servers for the zone, with the following
behaviour specified in RFC 2136, Section 4:

    4.6. If a response is received whose RCODE is SERVFAIL or NOTIMP, or
    if no response is received within an implementation dependent timeout
    period, or if an ICMP error is received indicating that the server's
    port is unreachable, then the requestor will delete the unusable
    server from its internal name server list and try the next one,
    repeating until the name server list is empty.  If the requestor runs
    out of servers to try, an appropriate error will be returned to the
    requestor's caller.