memory leak caused by PinnedArray

[doronf-cortica]

Reporting a bug

• I have tried using the latest released version of Numba (most recent is
  visible in the release notes
  (https://numba.readthedocs.io/en/stable/release-notes-overview.html).
• I have included a self contained code sample to reproduce the problem.
  i.e. it's possible to run as 'python bug.py'.

class PinnedArray has no __del__ function, so when memhostalloc allocates memory and passes it the pointer, that pointer will never be freed when the PinnedArray object is deleted.

see: https://github.com/numba/numba/blob/main/numba/cuda/cudadrv/driver.py#L2077

example:

import numba.cuda
import psutil
import os
import numba

proc = psutil.Process()
for i in range(10):
    print('memory before alloc: ', proc.memory_info().rss)
    f = numba.cuda.pinned_array(200000000)
    print('memory before del: ', proc.memory_info().rss)
    del f
    print('memory after del: ', proc.memory_info().rss)

[doronf-cortica]

i see now that numba.cuda.current_context().reset() fixes it. this isn't documented, though

[Graham Markall]

You shouldn't need to reset the context just to get your memory back though! :-)

The implementation of pinned_array() indeed looks wrong - I don't think it's even using PinnedArray, but if it did then the PinnedArray instance would (or should have) been constructed with a finalizer to free the memory when it's been garbage collected.

[doronf-cortica]

yeah, i noticed the finalizer after i had posted this issue and didn't look further into it...

[Siu Kwan Lam]

Isn't this behavior expected because how the cuda Context defer deallocations?

[Graham Markall]

Isn't this behavior expected because how the cuda Context defer deallocations?

The problem is that we're not even registering a finalizer when calling pinned_array() - so it's expected that a deallocation occurs eventually, but it's just never occurring in this case.