If the queues are full, we drop the present packet, which is better for
network traffic flow. Also, we try to fix up the memory leaks with not
putting buffers from our shared pool.
Fixed an issue in CreateBind for Linux:
If ipv6 was not supported the error code would be
correctly identified as EAFNOSUPPORT and ipv4 binding attempted.
However the port would be set to 0,
which results in the subsequent create4 call requesting
a random port rather than the one provided to CreateBind.
This issue was identified by:
Kent Friis <leeloored@gmx.com>
While we don't want people to ever use old protocols, people will
complain if the API "changes", so explicitly make the unset protocol
mean the latest, and add a dummy mechanism of specifying the protocol on
a per-peer basis, which we hope nobody actually ever uses.
Interestingly, ksh(1) on OpenBSD does not export PWD by default, and it
also has a notion of the "logical cwd" vs the "physical cwd", with the
latter being passed to chdir, but the former being stored in the
non-exported PWD and displayed to the user. This means that if you `cd`
into a directory that's comprised of symlinks, exec'd processes will see
the physical path. Observe:
# ksh
# mkdir a
# ln -s a b
# cd b
# pwd
/root/b
# ksh -c pwd
/root/a
The fact of separating physical and logical paths is not too uncommon
for shells (bash does it too), but not exporting PWD is very odd.
Since this is common behavior for many shells, libraries that return the
working directory will do something strange: they `stat(".")` and then
`stat(getenv("PWD"))`, and if these point to the same inode, they roll
with the value of `getenv("PWD")`, or otherwise fallback to asking the
kernel for the cwd.
Since PWD was not exported by ksh(1), Go's dep utility did not understand
it was operating inside of our faked GOPATH and became upset.
This patch works around the whole situation by simply exporting PWD
before executing dep.
