Messages in this thread | | | From | Linus Torvalds <> | Date | Mon, 28 Jun 2021 10:06:39 -0700 | Subject | Re: [GIT PULL] sigqueue cache fix |
| |
On Sun, Jun 27, 2021 at 10:14 PM Ingo Molnar <mingo@kernel.org> wrote: > > The most fundamental race we can have is this:
No. It's this (all on the same CPU):
sigqueue_cache_or_free():
if (!READ_ONCE(current->sigqueue_cache)) <-- Interrupt happens here WRITE_ONCE(current->sigqueue_cache, q);
and then the interrupt sends a SIGCONT, which ends up flushing previous process control signals, which ends up freeing them, which ends up in sigqueue_cache_or_free() again, at which point you have
if (!READ_ONCE(current->sigqueue_cache)) WRITE_ONCE(current->sigqueue_cache, q);
again.
And both the original and the interrupting one sees a NULL current->sigqueue_cache, so both of them will do that WRITE_ONCE(), and when the interrupt returns, the original case will overwrite the value that the interrupt free'd.
Boom - memory leak.
It does seem to be very small race window, and it's "only" a memory leak, but it's a very simple example of how this cache was broken even on UP.
Linus
| |