On Mon, Oct 24, 2016 at 7:08 PM, Andy Lutomirski <l...@amacapital.net> wrote: > On Mon, Oct 24, 2016 at 9:08 AM, Roman Pen > <roman.peny...@profitbricks.com> wrote: >> This patch avoids allocation of kthread structure on a stack, and simply >> uses kmalloc. Allocation on a stack became a huge problem (with memory >> corruption and all other not nice consequences) after the commit 2deb4be28 >> by Andy Lutomirski, which rewinds the stack on oops, thus ooopsed kthread >> steps on a garbage memory while completion of task->vfork_done structure >> on the following path: > > This is IMO a *huge* improvement. > > Shouldn't the patch also remove the try_get_task_stack() / > put_task_stack() hackery in kthread.c, though?
Do you mean that commit from Oleg https://patchwork.kernel.org/patch/9335295/ ? Hm, I missed that to_live_kthread() function completely. So, yes, we can remove put/get_task_stack(), but only replacing on get/put_kthread. So still need to be careful with the refs. Oleg, could you please take a look on the hunk below, just a quick thought. (get_kthread is simple atomic kthread->refs++). -- Roman --- a/kernel/kthread.c +++ b/kernel/kthread.c @@ -64,9 +64,20 @@ static inline struct kthread *to_kthread(struct task_struct *k) static struct kthread *to_live_kthread(struct task_struct *k) { struct completion *vfork = ACCESS_ONCE(k->vfork_done); - if (likely(vfork) && try_get_task_stack(k)) - return __to_kthread(vfork); - return NULL; + struct kthread *kthread = NULL; + + BUG_ON(!(k->flags & PF_KTHREAD)); + if (likely(vfork)) { + task_lock(k); + vfork = ACCESS_ONCE(k->vfork_done); + if (likely(vfork)) { + kthread = __to_kthread(vfork); + get_kthread(kthread); + } + task_unlock(k); + } + + return kthread; }
