Yak shaving is fun! not.

The build tries to ensure that we build from a clean git checkout.
That's nice, but it fails if git is not installed. While we could
conditionalize that on not being inside gitlab-ci, just installing git
is easier...

Since, apparently, apt doesn't have a stable interface.

Debian now has a CI infra on salsa; use that instead.

Tagging 3.18

(cherry picked from commit ccf099fa)

also, don't ignore it

(cherry picked from commit da78fd14)

Don't set unset (zero) timeout in nbd-client when configuring via netlink

When we are operating as a daemon we ignore unset (zero) timeout in
set_timeout() and don't set it in kernel.
However, when operating in the netlink mode netlink_configure() will add
NBD_ATTR_TIMEOUT attribute unconditionally.

This means that in the default case of timeout being unset it was being set
to zero in the kernel which resulted in it triggering almost immediately
upon any I/O request.

Fix this by not appending the timeout attribute in netlink_configure() when
the timeout is unset.

Signed-off-by: Maciej S. Szmigiero <mail@maciej.szmigiero.name>

The recent qemu 3.0 release added an additional supported context
namespace to its NBD_OPT_SET_META_CONTEXT implementation, named
"qemu:".  This serves as a registration of that third-party
namespace, while pointing to qemu documentation for the
specification of leaf-names within that namespace (currently just
a "dirty-bitmap:" sub-namespace that can select a named
dirty bitmap), and what they represent.  We don't have to track
the full details of the qemu implementation (especially since
qemu may add more sub-namespace or bit definitions over time),
but merely link to qemu documentation that should remain
up-to-date.

Signed-off-by: Eric Blake <eblake@redhat.com>
Reviewed-by: Vladimir Sementsov-Ogievskiy <vsementsov@virtuozzo.com>

When the NBD_CMD_BLOCK_STATUS extension was first discussed, the
idea of having the client's length be a hint was proposed, where
the server could reply beyond the client's request in order to
allow for fewer transactions when querying the entire disk. The
portion beyond the client's original request can only occur in
the final extent for a given context, and only if the additional
length matches the type given for the last byte actually requested
by the client.

In the meantime, qemu 2.12 was released as a first client
implementation of NBD_CMD_BLOCK_STATUS, which always sends the
NBD_CMD_FLAG_REQ_ONE flag, and which disconnects from the server
if the server's length exceeds the client request.  This was
relaxed for subsequent qemu, but it means that we have to be
explicit that a server should not send extra length except when
the client is not limiting its request to exactly one extent.

Furthermore, qemu 3.0 was recently released where the server
actually relies on this looser specification for its
qemu:dirty-bitmap:NAME context when the REQ_ONE flag is not used.

Signed-off-by: Eric Blake <eblake@redhat.com>

When nbd-server drops privileges, it was leaving supplementary
groups untouched. As nbd-server was normally dropping from root,
nbd-server kept membership to root supplementary groups.

Signed-off-by: Luiz Angelo Daros de Luca <luizluca@gmail.com>

The ability to request an NBD server to cache things locally,
without also spending time in the client reading what was just
cached, has proved useful to both the xNBD implementation
(public), and to a VM restore operation performed by Virtuozzo
atop qemu (not yet public).  Time to document this functionality,
rather than calling the command a failed experiment.

The documentation is rather fuzzy, in that it makes no
requirements on what constitutes a valid cache (a server can
implement this command as a no-op) and no guarantees that a
cache will benefit the client (although for a specific server
implementation, the client will indeed benefit).  But as the
command has a use in existing implementations, it's better
to document what we can to prevent future incompatible use
of the command.  In particular, we can envision the addition
of command flags to request specific caching behvaiors (perhaps
copying after posix_fadvise semantics), where a server must
reject a flag it does not know or cannot honor (in the extreme
case, a new flag to allow the client to drop a portion of the
cache must not be misinterpreted by the server that populates
the cache instead because it ignored an unrecognized flag).

Note that this proposal adds NBD_FLAG_SEND_CACHE which xNBD
did not use; meanwhile, the command NBD_CMD_CACHE is allowed
to fail with no visible side effects.  The documentation tries
to point out interoperability constraints resulting from
existing servers that understand the command but don't advertise
it, as well as existing clients that send the command even when
it is not advertised.

Signed-off-by: Vladimir Sementsov-Ogievskiy <vsementsov@virtuozzo.com>
Message-Id: <20180326161049.15981-1-vsementsov@virtuozzo.com>
[eblake: add commit message, add more documentation]
Signed-off-by: Eric Blake <eblake@redhat.com>

Thadeu Lima de Souza Cascardo authored May 18, 2018 and Wouter Verhelst committed May 20, 2018

The parent will receive the servename from the child to verify if it has
reached the max number of connections. When the servename is the empty name, it
will try to allocate a 0-sized buffer, which will return a NULL pointer, and
that segfaults when running strcmp.

Signed-off-by: Thadeu Lima de Souza Cascardo <cascardo@canonical.com>
Signed-off-by: Wouter Verhelst <w@uter.be>

Thadeu Lima de Souza Cascardo authored May 18, 2018 and Wouter Verhelst committed May 20, 2018

When fork fails, we return without closing the socket pair that was just
created. This might cause file descriptor leaks.

Signed-off-by: Thadeu Lima de Souza Cascardo <cascardo@canonical.com>
Signed-off-by: Wouter Verhelst <w@uter.be>

We used to have a value "modern" for every client which retained whether
or not the client had negotiated using the newstyle protocol. However,
since we don't support oldstyle anymore, this is now a dead assignment,
so drop it.

Not strictly necessary, but having consistent line widths makes it
easier to read the original markup rather than having to refer to
the rendered webpage. No content change.

Signed-off-by: Eric Blake <eblake@redhat.com>

Qemu 2.12 implemented NBD_CMD_BLOCK_STATUS, so it is time to promote
this extension to stable.

Don't make gnutls and libnl automagic

Introduce --without-gnutls and --without-libnl configure options so that
users can disable gnutls/libnl even if the packages are available on the
system. The default is unchanged from before this patch. If no
--with(out)-* option has been given on the command line, the macro looks
for presence and uses the packages if found.

Disabling tlshuge entirely means we don't test it on Linux either, where
it does succeed. Rather than doing that, mark it as an expected failure
on OSX.

We can fix it later, then.

We currently fail the test suite if syslog was not enabled at compile
time. While it can't work, it makes more sense to skip the test than to
fail it.

Sigh.

realpath is "portable", except not implemented on older versions of
coreutils, including the one in Ubuntu Trusty :-(

So, default to using realpath (which has the advantage of not requiring
options anywhere), and fall back to readlink -f if that doesn't exist in
our path.

If that also fails, then the test suite will fail, but have a nice
"readlink: -f: invalid option" or some such all over it, pointing to
what the likely culprit is.

We actually do need to pass non-symlinked stuff, just can't use readlink for
it. Use realpath, which is the portable alternative, instead.

autoconf has a facility for doing this, no need to create custom rules
in Makefile.am like we were doing. This also takes care of figuring out
that some systems don't support symlinks, and falling back to hardlinks
or copies, where necessary.