This changeset builds on previous 2-3 commits and represents the main
libqb's answer to the original question behind pacemaker's security
defect known as CVE-2016-7035.
Beside the helper partly unifying handling of qb_rb_force_close and
qb_rb_close, it provides the former with ability to use file truncating
as a fallback for when unlinking fails, e.g., because client (note that
mentioned is currently only relevant for the client side as normally
server is responsible for the lifecycle of the materialized files,
unless it crashes and only client is left to do its best) is not the
owner while they are placed at a directory with restricted deletion,
which enforces this very ownership condition.
In practice, this means that, at worst, just the zero-size files are
left behind, so not that much space exhaustion (usually "ramdisk"
like tmpfs is what backs default storage directory /dev/shm, so it
boils down to physical memory exhaustion, even if it can be just
for page cache and related overhead) can happen even on repeated
crashes as the memory mappings are cleared as much as possible.
Also openat/unlinkat functions (sported in qb_sys_unlink_or_truncate_at
as of the previous commit) are, when applicable, used so as to limit
possible race conditions between/during individual path traversals
(both files being got rid of presumably share the same directory).
Few words on which actions are attempted in which order for the
equivalent of qb_rb_force_close now:
There are subtle interactions between what's externally visible
(files) and what's not (memory mappings associated with such files),
and perhaps between memory pages management from the perspective of
the former (usually "ramdisk"/tmpfs) and the latter (mmap + munmap).
If the associated file is no longer publicly exposed by the means of
unlink (even if the object survives internally as refcounting is in
the game, with mmap holding a reference), memory mapping is not
affected. On the other hand, if it's just limited by truncation
to zero size, memory mapping is aware and generates SIGBUS in response
to accessing respective addresses. Similarly, accessing munmap'd
(no refcounting here) memory generates SIGSEGV. For delicacy,
the inputs for all of unlink, truncate, and munmap are stored
at the mmap'd location we are about to drop, but that's just a matter
of making copies ahead of time.
At Ken's suggestion, the scheme is: (unlink or truncate) then munmap,
which has a benefit that externally visible (and program's life span
otherwise surviving!) part is eliminated first, with memory mappings
(disposed at program termination automatically at latest) to follow.
(There was originally a paranoid expectation on my side that truncate
on tmpfs actually does silent munmap, so that our munmap could in fact
tear down the mapping added in the interim by the libraries, signal
handler or due to requirements of another thread, also because of
munmap on the range without any current mappings will not fail, and
thus there's likely no portable way to non-intrusively check the
status, but also due to documented SIGBUS vs. SIGSEGV differences
the whole assumption appears bogus on the second thought.)
Relevant unit tests that exercise client-side unlinking:
- check_ipc: test_ipc_server_fail_shm, test_ipc_exit_shm
- new test in a subsequent commit
This ringbuffer is usable across processes.
the point is to use this for IPC to provide async
connections from client to server, but with inherient
flow control.
This still needs a bit of clean up, but committing now
for feedback and as it is quite functional.
Signed-off-by: Angus Salkeld <asalkeld@redhat.com>
