On 2019-10-14, Tejun Heo tj@kernel.org wrote:
Hello, Aleksa.
On Tue, Oct 15, 2019 at 02:59:31AM +1100, Aleksa Sarai wrote:
On 2019-10-14, Tejun Heo tj@kernel.org wrote:
On Sat, Oct 12, 2019 at 12:05:39PM +1100, Aleksa Sarai wrote:
Because pids->limit can be changed concurrently (but we don't want to take a lock because it would be needlessly expensive), use the appropriate memory barriers.
I can't quite tell what problem it's fixing. Can you elaborate a scenario where the current code would break that your patch fixes?
As far as I can tell, not using *_ONCE() here means that if you had a process changing pids->limit from A to B, a process might be able to temporarily exceed pids->limit -- because pids->limit accesses are not protected by mutexes and the C compiler can produce confusing intermediate values for pids->limit[1].
But this is more of a correctness fix than one fixing an actually exploitable bug -- given the kernel memory model work, it seems like a good idea to just use READ_ONCE() and WRITE_ONCE() for shared memory access.
READ/WRITE_ONCE provides protection against compiler generating multiple accesses for a single operation. It won't prevent split writes / reads of 64bit variables on 32bit machines. For that, you'd have to switch them to atomic64_t's.
Maybe I'm misunderstanding Documentation/atomic_t.txt, but it looks to me like it's explicitly saying that I shouldn't use atomic64_t if I'm just using it for fetching and assignment.
The non-RMW ops are (typically) regular LOADs and STOREs and are canonically implemented using READ_ONCE(), WRITE_ONCE(), smp_load_acquire() and smp_store_release() respectively. Therefore, if you find yourself only using the Non-RMW operations of atomic_t, you do not in fact need atomic_t at all and are doing it wrong.
As for 64-bit on 32-bit machines -- that is a separate issue, but from [1] it seems to me like there are more problems that *_ONCE() fixes than just split reads and writes.
[1]: https://lwn.net/Articles/793253/