On Friday 29 January 2016 19:14:58 Richard W.M. Jones wrote:
On Fri, Jan 29, 2016 at 07:25:30PM +0100, Pino Toscano wrote:
The API for UNIX sockets limits the path to a static size (usually 104 or 108 characters, NULL included), which is internally represented as UNIX_PATH_MAX. If the temporary directory set is long enough (e.g. when running tools as uninstalled, using "run") then these socket paths get trucated, if not even cause failures when binding the sockets.
Introduce a new internal API to create paths for sockets, and a new helper path where store them in the above case. This new path is created only when needed, on a temporary directory under /tmp (hardcoded), and cleaned when the handle is destroyed.
I think reading between the lines, you want to hard code /tmp so that UNIX_PATH_MAX can never be exceeded, even if the user sets $TMPDIR to some stupid long path?
That's the idea of the initial approach, yes. Of course that could be changed to use $TMPDIR (or /tmp) as fallback, still failing if even this path would be too long.
Are there machines where /tmp is unusable?
/tmp is a mandatory part of FHS [1], so we could assume it is usable (we basically do that already, see guestfs_impl_get_tmpdir). [1] http://refspecs.linuxfoundation.org/FHS_3.0/fhs-3.0.html
Should we provide an API to read the sockname, analogous to guestfs_get_tmpdir and guestfs_get_cachedir?
This makes me think that we could actually move all the sockets (regardless of the length of tmpdir) to a new sockdir, and expose that as read-only API.
Should we use /run or /dev/shm instead? I would say, not /run/user because systemd doesn't reliably create it, unfortunately, but maybe somewhere else in /run would be acceptable.
Theoretically /run/user/$UID is what systemd sets as $XDG_RUNTIME_DIR, which is described in the XDG basedir spec [2]. We could try to use it if set, creating a new guestfs-sockXXXXXX temporary directory. [2] http://standards.freedesktop.org/basedir-spec/latest/ar01s03.html
How does libvirt handle socket paths? Does it put them in /run, and can we emulate its behaviour?
It uses $XDG_RUNTIME_DIR if set, falling back to ~/.cache ($XDG_CACHE_HOME actually), creating the directories (mkdir -p, basically).
Are there implications for testing? [When running tests, the ./run script sets TMPDIR to point to /path/to/libguestfs/tmp.
Minot nitpick: "run" does not set $TMPDIR, but $LIBGUESTFS_TMPDIR, so temporary stuff of anything non-libguestfs are still outside of libguestfs' tmpdir. Aside the above: the appliance will still be in the same possible places (including where "run" points tmpdir to), we need to find a better place for sockets only (whose paths have a limitation on the length). -- Pino Toscano