/test

Codecov Report

Merging #1789 into master will not change coverage.
The diff coverage is 100%.

@@          Coverage Diff           @@
##           master   #1789   +/-   ##
======================================
  Coverage      54%     54%           
======================================
  Files         106     106           
  Lines       13170   13170           
======================================
  Hits         7113    7113           
  Misses       5210    5210           
  Partials      847     847

A backport to stable branches is needed for this PR. @Ace-Tang can you please submit this against stable-1.6 and stable-1.7 branches?

@Ace-Tang How does this fix the problem? you mentioned that for the network namespace created, the tid is differerent from the pid.
But passing the tid to libnetwork would cause it to use the tid in entering the network namespace using /proc/%d/ns/net. How would that work? We are substituting a thread id here instead of process id.

I think with shimv1 we explicitly used LockOSThread() to make sure the netns is created on the main thread so that pid and tid are the same. @sboeuf had worked on this iirc.
We should probably look at why we see this beahviour in case of shimv2 where the tid ends up being different.

@amshinde - the thread is associated with the correct netns.

That path, using thread id, will resolve correctly to the /proc/pid/task..../ns/net

@egernst Thanks for the explanation,I guess the thread entries are hidden when you do a ls -l /proc

Now that this is merged, we should also have a test for this, I dont think we test shimv2 today with CNM.
@Ace-Tang Since you went through the process recently, can you list out the setup steps for shimv2 with CNM.

I think with shimv1 we explicitly used LockOSThread() to make sure the netns is created on the main thread so that pid and tid are the same. @sboeuf had worked on this iirc.
We should probably look at why we see this beahviour in case of shimv2 where the tid ends up being different.

this beahviour in case of shimv2 where the tid ends up being different , this is what I not figure out , I pause the v2 shim process after it enter netns, I saw that pid from getpid() still keep host netns, but only gettid() changed the netns.
And after this is figured out, maybe can have a better way to fix the problem.

@amshinde, as I posts in issue, I test this with pouch, since docker not support shim v2, but pouch does and pouch use libnetwork. So I simply run with pouch.

goroutine within LockOSThread()/UnlockOSThread() is executed in a different os thread. And grpc server uses goroutines to serve requests. Therefore we cannot rely on LockOSThread() to ensure all grpc request handlers run in the same kernel thread with shimv2. That is also the why tids ends up being different with shimv2.