On Thu 13-08-20 15:29:31, Uladzislau Rezki wrote:
[...]
> I was a bit out of focus and did not mention about one thing. Identifying the context
> type using preemtable() primitives looks a bit not suitable, IMHO. GFP_* flags are
> not supposed to identify a context type, because it is not possible for all cases.
Yes, GFP flags do not identify context and that is why my draft didn't
really consider gfp flags for anything but the retry logic which is
already gfp based already. The buddy allocator path simply always bails
out for those rt atomic paths whenever it gets close to zone->lock
> But that i
You meant to say more I guess
> Also, to bail out based on a context's type will not allow to get a page from atomic
> contexts, what we try to achieve :)
Yes lockdep will need some special treatment but I suspect we want to
address the underlying problem first and only then take care of the
lockdep.
> Whereas, i mean, we do have possibility to do lock-less per-cpu-list allocation without
> touching any zone lock.
>
> if (gfp_mask == 0) {
> check_pcp_lists();
> if (page)
> return page;
>
> bail out here without doing farther logic, like pre-fetching.
> return NULL;
> }
The existing code does that already. __rmqueue_pcplist will go
rmqueue_bulk only when pcp lists are empty. Or did I miss your point?
--
Michal Hocko
SUSE Labs
