On Wed, Jan 27, 2016 at 02:03:36PM +0100, Vlastimil Babka wrote: > On 01/27/2016 01:47 PM, Christian Borntraeger wrote: > > On 01/27/2016 01:41 PM, Vlastimil Babka wrote: > >> On 01/27/2016 01:59 AM, Joonsoo Kim wrote: > >> > >> I think it might be worth also to convert debug_pagealloc_enabled() to be > >> based > >> on static key, like I did for page_owner [1]. That should help make it > >> possible > >> to have virtually no overhead when compiling kernel with > >> CONFIG_DEBUG_PAGEALLOC > >> without enabling it boot-time. I assume it's one of the goals here? > > > > We could do something like that but dump_stack and setup of the initial > > identity > > mapping of the kernel as well as the initial page protection are not hot > > path > > as far as I can tell. Any other places? > > Well, mostly kernel_map_pages() which is used in page allocation hotpaths.
We cannot just convert it because setup_arch() is called before jump_label_init(). We can do it by introducing _early variant but I'm not sure it's worth doing. I think that just make it unlikely works well. Recently, I tested a micro benchmark on slab alloc/free on CONFIG_SLUB with applying static branch for debug option which currently is using unlikely branch and can't find any noticeble performance difference. Thanks.
