490369515c
Signed-off-by: Yoan Blanc <yoan@dosimple.ch> |
||
|---|---|---|
| .. | ||
| README.md | ||
| ns_linux.go | ||
README.md
Namespaces, Threads, and Go
On Linux each OS thread can have a different network namespace. Go's thread scheduling model switches goroutines between OS threads based on OS thread load and whether the goroutine would block other goroutines. This can result in a goroutine switching network namespaces without notice and lead to errors in your code.
Namespace Switching
Switching namespaces with the ns.Set() method is not recommended without additional strategies to prevent unexpected namespace changes when your goroutines switch OS threads.
Go provides the runtime.LockOSThread() function to ensure a specific goroutine executes on its current OS thread and prevents any other goroutine from running in that thread until the locked one exits. Careful usage of LockOSThread() and goroutines can provide good control over which network namespace a given goroutine executes in.
For example, you cannot rely on the ns.Set() namespace being the current namespace after the Set() call unless you do two things. First, the goroutine calling Set() must have previously called LockOSThread(). Second, you must ensure runtime.UnlockOSThread() is not called somewhere in-between. You also cannot rely on the initial network namespace remaining the current network namespace if any other code in your program switches namespaces, unless you have already called LockOSThread() in that goroutine. Note that LockOSThread() prevents the Go scheduler from optimally scheduling goroutines for best performance, so LockOSThread() should only be used in small, isolated goroutines that release the lock quickly.
Do() The Recommended Thing
The ns.Do() method provides partial control over network namespaces for you by implementing these strategies. All code dependent on a particular network namespace (including the root namespace) should be wrapped in the ns.Do() method to ensure the correct namespace is selected for the duration of your code. For example:
err = targetNs.Do(func(hostNs ns.NetNS) error {
dummy := &netlink.Dummy{
LinkAttrs: netlink.LinkAttrs{
Name: "dummy0",
},
}
return netlink.LinkAdd(dummy)
})
Note this requirement to wrap every network call is very onerous - any libraries you call might call out to network services such as DNS, and all such calls need to be protected after you call ns.Do(). All goroutines spawned from within the ns.Do will not inherit the new namespace. The CNI plugins all exit very soon after calling ns.Do() which helps to minimize the problem.
When a new thread is spawned in Linux, it inherits the namespace of its parent. In versions of go prior to 1.10, if the runtime spawns a new OS thread, it picks the parent randomly. If the chosen parent thread has been moved to a new namespace (even temporarily), the new OS thread will be permanently "stuck in the wrong namespace", and goroutines will non-deterministically switch namespaces as they are rescheduled.
In short, there was no safe way to change network namespaces, even temporarily, from within a long-lived, multithreaded Go process. If you wish to do this, you must use go 1.10 or greater.
Creating network namespaces
Earlier versions of this library managed namespace creation, but as CNI does not actually utilize this feature (and it was essentially unmaintained), it was removed. If you're writing a container runtime, you should implement namespace management yourself. However, there are some gotchas when doing so, especially around handling /var/run/netns. A reasonably correct reference implementation, borrowed from rkt, can be found in pkg/testutils/netns_linux.go if you're in need of a source of inspiration.