| Date | Thu, 11 Jun 2015 16:13:53 +0200 | From | Ingo Molnar <> | Subject | Re: [RFC PATCH 00/12] x86/mm: Implement lockless pgd_alloc()/pgd_free() |
| |
[ I fat-fingered the linux-mm Cc:, so every reply will bounce on that, sorry about that :-/ Fixed it in this mail's Cc: list. ]
* Ingo Molnar <mingo@kernel.org> wrote:
> Waiman Long reported 'pgd_lock' contention on high CPU count systems and > proposed moving pgd_lock on a separate cacheline to eliminate false sharing and > to reduce some of the lock bouncing overhead.
So 'pgd_lock' is a global lock, used for every new task creation:
arch/x86/mm/fault.c:DEFINE_SPINLOCK(pgd_lock);
which with a sufficiently high CPU count starts to hurt.
Thanks,
Ingo
|