Messages in this thread | | | Date | Sun, 16 Dec 2012 20:16:21 +0000 | From | Al Viro <> | Subject | Re: [PATCH] mm: Downgrade mmap_sem before locking or populating on mmap |
| |
On Sun, Dec 16, 2012 at 05:04:03PM +0000, Al Viro wrote:
> That's just from a couple of days of RTFS. The locking in there is far too > convoluted as it is; worse, it's not localized code-wise, so rechecking > correctness is going to remain a big time-sink ;-/ > > Making it *more* complex doesn't look like a good idea, TBH...
... and another fun place: kvm_setup_async_pf() grabs a _passive_ reference to current->mm (->mm_count, not ->mm_users), sticks it into work->mm and schedules execution of async_pf_execute(). Which does use_mm() (still no active refs acquired), grabs work->mm->mmap_sem shared and proceeds to call get_user_pages(). What's going to happen if somebody does kill -9 to the process that had started that?
get_user_pages() in parallel with exit_mmap() is a Bad Thing(tm) and I don't see anything on the exit path that would've waited for that work to finish. I might've missed something here, but... Note that aio (another place playing with use_mm(), also without an active ref) has an explicit hook for mmput() to call before proceeding to exit_mmap(); I don't see anything similar here.
Not that aio.c approach had been all that safe - get_task_mm() will refuse to pick use_mm'ed one, but there are places open-coding it without the check for PF_KTHREAD. Few of them, fortunately, but...
| |