multiprocessing.shared_memory: failed attach unlinks another process's block

Bug report

Bug description

SharedMemory.__init__ calls self.unlink() from its OSError cleanup handler:

try:
    if create and size:
        os.ftruncate(self._fd, size)
    stats = os.fstat(self._fd)
    size = stats.st_size
    self._mmap = mmap.mmap(self._fd, size)
except OSError:
    self.unlink()
    raise

This runs unconditionally, including on the attach path (create=False).
unlink() calls shm_unlink(), which permanently destroys the named block.
So if mmap() fails while attaching to a block created by another process
(e.g. ENOMEM under memory pressure), the failing attach destroys another
owner's live shared memory block.

Reproduction (simulating an mmap failure during attach):

from unittest import mock
from multiprocessing import shared_memory

owner = shared_memory.SharedMemory(create=True, size=1024)
try:
    with mock.patch("multiprocessing.shared_memory.mmap.mmap",
                    side_effect=OSError("ENOMEM")):
        shared_memory.SharedMemory(owner.name)   # attach
except OSError:
    pass

shared_memory.SharedMemory(owner.name)   # FileNotFoundError: block destroyed

Secondary issue: resource_tracker.register() runs after this try/except, so
on the create=True path the block is not yet registered. unlink() still
calls resource_tracker.unregister(), so the resource_tracker does
cache[rtype].remove(name) on an unknown name → KeyError printed as a
traceback, with the tracker's exit code set to 3.

Expected behavior

A failed attach must not unlink a block it does not own. On error the fd should
just be closed; the block should be unlinked only when it was created in this
call, and without a spurious unregister for a never-registered block.

Your environment

• CPython main (3.16.0a0); also affects earlier versions
• Linux / POSIX shared memory (_USE_POSIX)

Linked PRs

• gh-153280

Commencez par multiprocessing.shared_memory.SharedMemory.init et reproduisez l’échec de attach avec le mock montré dans l’issue. Vérifiez le comportement du cleanup et du resource-tracker pour les chemins d’attach et de create. C’est terminé lorsqu’un attach échoué laisse le bloc existant utilisable et ne produit pas d’erreur parasite du tracker ; l’issue inclut la PR liée gh-153280.