Original Patch From Brad Spencer:
https://c-ares.haxx.se/mail/c-ares-archive-2016-04/0000.shtml
My modifications include:
* Dynamically find GetBestRoute2 since it is a Windows Vista+ symbol, and will fall back to prior behavior when not available.
* Prefer get_DNS_AdaptersAddresses as the modifications should alleviate the concerns which caused us to prefer get_DNS_NetworkParams
* Update AppVeyor to use MinGW-w64 instead of the legacy MinGW
* Fix compile error in test suite for Windows.
Original message from patch below:
From: Brad Spencer <bspencer@blackberry.com>
Date: Fri, 29 Apr 2016 14:26:23 -0300
On Windows, the c-ares DNS resolver tries first to get a full list of
DNS server addresses by enumerating the system's IPv4/v6 interfaces and
then getting the per-interface DNS server lists from those interfaces
and joining them together. The OS, at least in the way the c-ares
prefers to query them (which also may be the only or best way in some
environments), does not provide a unified list of DNS servers ordered
according to "current network conditions". Currently, c-ares will then
try to use them in whatever order the nested enumeration produces, which
may result in DNS requests being sent to servers on one interface
(hosting the current default route, for example) that are only intended
to be used via another interface (intended to be used when the first
interface is not available, for example). This, in turn, can lead to
spurious failures and timeouts simply because of the server address
order that resulted because of the enumeration process.
This patch makes the (safe?) assumption that there is no other better
rule to chose which interface's DNS server list should be prioritized.
After all, a DNS lookup isn't something "per network"; applications
don't look up "these DNS names on this interface and those DNS names on
that interface". There is a single resource pool of DNS servers and the
application should presume that any server will give it the "right"
answer. However, even if all DNS servers are assumed to give equally
useful responses, it is reasonable to expect that some DNS servers will
not accept requests on all interfaces. This patch avoids the problem by
sorting the DNS server addresses using the Windows IPv4/v6 routing tables.
For example, a request to DNS server C on interface 2 that is actually
sent over interface 1 (which may happen to have the default route) may
be rejected by or not delivered to DNS server C. So, better to use DNS
servers A and B associated with interface 1, at least as a first try.
By using the metric of the route to the DNS server itself as a proxy for
priority of the DNS server in the list, this patch is able to adapt
dynamically to changes in the interface list, the DNS server lists per
interface, which interfaces are active, the routing table, and so on,
while always picking a good "best" DNS server first.
In cases where any DNS server on any interface will do, this patch still
seems useful because it will prioritize a lower-metric route's (and thus
interface's) servers.
Add a new ares_library_init_mem() initialization function for the
library which allows the library user to specify their own malloc,
realloc & free equivalents for use library-wide.
Store these function pointers in library-wide global variables,
defaulting to libc's malloc(), realloc() and free().
Change all calls to malloc, realloc and free to use the function pointer
instead. Also ensure that ares_strdup() is always available
(even if the local environment includes strdup(3)), and change the
library code to always use it.
Convert calls to calloc() to use ares_malloc() + memset
Add comments for the benefit of the lcov tool, marking
lines that cannot be hit. Typically these are fall-back
protection arms that are already covered by earlier checks,
and so it's not worth taking out the unhittable code (in case
someone changes the code between the two places in future).
Previously, a single call to ares_library_cleanup() would deinitialise
the c-ares library, regardless of how many times ares_library_init() was
called. This behaviour may cause problems in programs linking two or
more libraries which, in turn, use c-ares. The present commit fixes this
problem, deinitializing the library only after a number of calls to
ares_library_cleanup() matching the number of calls to
ares_library_init().