bpo-36867: Create the resource_tracker before launching SharedMemoryManagers

[pierreglaser]

TLDR; The resource_tracker needs to be created before SharedMemoryManager processes are launched (which is not the case for managers created using fork)

Take this example below:

frommultiprocessing.managersimportSharedMemoryManagerfrommultiprocessingimportresource_trackersmm=SharedMemoryManager()
smm.start()
sl=smm.ShareableList(range(10))
smm.shutdown()

This simple and legitimate python scripts results in a very noisy output:

/home/pierreglaser/repos/cpython/Lib/multiprocessing/resource_tracker.py:198: UserWarning: resource_tracker: Thereappeartobe1leakedshared_memoryobjectstocleanupatshutdownwarnings.warn('resource_tracker: There appear to be %d '/home/pierreglaser/repos/cpython/Lib/multiprocessing/resource_tracker.py:211: UserWarning: resource_tracker: '/psm_3bfa242d': [Errno2] Nosuchfileordirectory: '/psm_3bfa242d'warnings.warn('resource_tracker: %r: %s'% (name, e))

Here is why it happens:

1. smm.start() launches a new manager process. At this point in the main process, no resource_tracker is created if the manager is launched using fork.

2. sl = smm.ShareableList(range(10)) does two things:

   • First, it creates a new shared_memory object from the parent process. At creation, a resource_tracker.register call is done -> resource_tracker.ensure_running is done -> a new resource_tracker process is created. In addition, this resource_tracker now tracks sl. However, since the manager process was created before, it does not know that a new resource_tracker process was just created in the parent process..
   • Once sl is created, the SharedMemoryManager asks its SharedMemoryTracker ( which lives in the manager process) to track it.

3. smm.shutdown shuts down the manager. It thus asks the SharedMemoryTracker to destroy sl, which it tracks. To do so,

   • it first opens it (using a SharedMemory(name)) call, which itself triggers a resource_tracker.register call -> resource_tracker.ensure_running call in the manager process, where no resource_tracker exists yet! Thus, another ResourceTracker instance /process is created and starts tracking sl
   • immediatly after, it unlinkssl. This sends an unregister call to the newly created resource_tracker.

When the script ends however, nothing has been done to tell the original resource_tracker created in the parent process that sl was properly unlinked. The resource_tracker thus thinks a leak happened, and tells the user (first line of my output)). However, there was no leak, as sl was properly unlinked by the manager. Therefore, the cleanup attempt of the resource_tracker fails, resulting the second error.

The solution to this problem is simply to create the right before the manager process is created.

https://bugs.python.org/issue36867

[auvipy]

no test?

[pierreglaser]

@pitrou how does this look to you?

[pitrou]

I have a more general question: what if the SharedMemoryManager is created in an unrelated process, and then connected to using address and authkey?

[pierreglaser]

@pitrou then most likely the process connecting to the SharedMemoryManager and the ResourceTracker will not have the same resource_tracker processes. Similar problems than the one this PR tries to solve would occur.

[pitrou]

I see. Is there the same problem with semaphores?

[pierreglaser]

A quick skim through SyncManager, it seems that no, semaphores will not show such problems, because the SyncManager API returns a proxy objcet and not a Semaphore object to the parent process. As no actual semaphore object is created in the parent process, the parent process will not ask its resource_tracker to the object.

[pitrou]

I see. I guess we'll have to live with the limitation.

[pitrou]

By the way, I noticed that the multiprocessing docs still mention the "semaphore tracker process". This should be fixed.

[pierreglaser]

OK I'll make a docfix PR.

[pitrou]

Ok, I'll merge this now.

[pitrou]

Hmm, it seems there are still resource tracker warnings on CI (see e.g. Travis builds). Are they expected?

Edit: the warning is from test_shared_memory_cleaned_after_process_termination. So it's expected (and I don't think there's an easy way to silence it).

[pitrou]

There's another ResourceWarning:

test_shared_memory_SharedMemoryManager_reuses_resource_tracker (test.test_multiprocessing_fork.WithProcessesTestSharedMemory) ... /home/antoine/cpython/default/Lib/unittest/case.py:683: ResourceWarning: unclosed file <_io.BufferedReader name=3>
testMethod()

Edit: will fix it myself.

[pierreglaser]

@pitrou I could not reproduce the second ResourceWarning locally...

[pitrou]

@pierreglaser You probably needed -Wd.

[pierreglaser]

Looking good @pitrou :)

[pitrou]

Looks like we can move on :-)