On Fri, May 17, 2019 at 7:11 PM Michal Hocko <mhocko@xxxxxxxxxx> wrote: > > On Fri 17-05-19 09:27:54, Kees Cook wrote: > > On Fri, May 17, 2019 at 04:01:08PM +0200, Michal Hocko wrote: > > > On Fri 17-05-19 15:37:14, Alexander Potapenko wrote: > > > > > > > Freeing a memory is an opt-in feature and the slab allocator can already > > > > > > > tell many (with constructor or GFP_ZERO) do not need it. > > > > > > Sorry, I didn't understand this piece. Could you please elaborate? > > > > > > > > > > The allocator can assume that caches with a constructor will initialize > > > > > the object so additional zeroying is not needed. GFP_ZERO should be self > > > > > explanatory. > > > > Ah, I see. We already do that, see the want_init_on_alloc() > > > > implementation here: https://patchwork.kernel.org/patch/10943087/ > > > > > > > So can we go without this gfp thing and see whether somebody actually > > > > > > > finds a performance problem with the feature enabled and think about > > > > > > > what can we do about it rather than add this maint. nightmare from the > > > > > > > very beginning? > > > > > > > > > > > > There were two reasons to introduce this flag initially. > > > > > > The first was double initialization of pages allocated for SLUB. > > > > > > > > > > Could you elaborate please? > > > > When the kernel allocates an object from SLUB, and SLUB happens to be > > > > short on free pages, it requests some from the page allocator. > > > > Those pages are initialized by the page allocator > > > > > > ... when the feature is enabled ... > > > > > > > and split into objects. Finally SLUB initializes one of the available > > > > objects and returns it back to the kernel. > > > > Therefore the object is initialized twice for the first time (when it > > > > comes directly from the page allocator). > > > > This cost is however amortized by SLUB reusing the object after it's been freed. > > > > > > OK, I see what you mean now. Is there any way to special case the page > > > allocation for this feature? E.g. your implementation tries to make this > > > zeroying special but why cannot you simply do this > > > > > > > > > struct page * > > > ____alloc_pages_nodemask(gfp_t gfp_mask, unsigned int order, int preferred_nid, > > > nodemask_t *nodemask) > > > { > > > //current implementation > > > } > > > > > > struct page * > > > __alloc_pages_nodemask(gfp_t gfp_mask, unsigned int order, int preferred_nid, > > > nodemask_t *nodemask) > > > { > > > if (your_feature_enabled) > > > gfp_mask |= __GFP_ZERO; > > > return ____alloc_pages_nodemask(gfp_mask, order, preferred_nid, > > > nodemask); > > > } > > > > > > and use ____alloc_pages_nodemask from the slab or other internal > > > allocators? Given that calling alloc_pages() with __GFP_NO_AUTOINIT doesn't visibly improve the chosen benchmarks, and the next patch in the series ("net: apply __GFP_NO_AUTOINIT to AF_UNIX sk_buff allocations") only improves hackbench, shall we maybe drop both patches altogether? > > If an additional allocator function is preferred over a new GFP flag, then > > I don't see any reason not to do this. (Though adding more "__"s seems > > a bit unfriendly to code-documentation.) What might be better naming? > > The naminig is the last thing I would be worried about. Let's focus on > the most simplistic implementation first. And means, can we really make > it as simple as above? At least on the page allocator level. > > > This would mean that the skb changes later in the series would use the > > "no auto init" version of the allocator too, then. > > No, this would be an internal function to MM. I would really like to > optimize once there are numbers from _real_ workloads to base those > optimizations. > -- > Michal Hocko > SUSE Labs -- Alexander Potapenko Software Engineer Google Germany GmbH Erika-Mann-Straße, 33 80636 München Geschäftsführer: Paul Manicle, Halimah DeLaine Prado Registergericht und -nummer: Hamburg, HRB 86891 Sitz der Gesellschaft: Hamburg
