Re: [PATCH v3 00/13] Virtually mapped stacks with guard pages (x86, core)

[Andy Lutomirski]

On Sat, Jun 25, 2016 at 4:19 PM, Andy Lutomirski <luto@xxxxxxxxxxxxxx> wrote:
> I rebased my series onto your tree and then rebased this thing onto my
> series, tweaked it some, and split it up a bit.  My version works a
> bit differently (thread_info has a single element if the new option is
> set) but is otherwise more or less your code.  It seems to work.
>
> https://git.kernel.org/cgit/linux/kernel/git/luto/linux.git/commit/?h=x86/vmap_stack&id=01ac3242f314405c1bac0f820ec3575a850509d3
> https://git.kernel.org/cgit/linux/kernel/git/luto/linux.git/commit/?h=x86/vmap_stack&id=87194cac139aebecec89a68eff719ab44f0469a2
>
> On top of *that*, I taught the kernel to free stacks in release_task
> and to cache stacks if vmalloced.  That still blows up: when
> cryptomgr_test calls do_exit during boot, do_exit calls exit_notify,
> which observes that the task state is TASK_DEAD and thus calls
> release_task on itself and goes boom.
>
> Linus, Oleg, help?  How am I supposed to quickly free the stack if the
> task goes through this code path?  In fact, why does this work on
> current kernels?  After all, can't schedule() indicates an RCU grace
> period?  In principle, what prevents delayed_put_task_struct from
> deleting the running stack before scheduling?
>
> My kludged up patch that only early-releases the stack if release_task
> is called from a different task is here:
>
> https://git.kernel.org/cgit/linux/kernel/git/luto/linux.git/commit/?h=x86/vmap_stack&id=2327ca2ab3d634d120ea185e737fef2e4e9cf012

Maybe I'm misunderstanding the role of release_task.  It looks like
there's this path in the scheduler I can borrow:

    if (unlikely(prev_state == TASK_DEAD)) {

With a kludge in place to free the stack in there and release_task and
__put_task_struct, whichever is first, I get a nice speedup.
Benchmarks coming later on.  Can I rely on that code path always being
called?

--Andy