Messages in this thread | | | Date | Thu, 22 Feb 2018 07:59:43 +0100 | From | Michal Hocko <> | Subject | Re: Use higher-order pages in vmalloc |
| |
On Wed 21-02-18 09:01:29, Matthew Wilcox wrote: > On Wed, Feb 21, 2018 at 08:16:22AM -0800, Dave Hansen wrote: > > On 02/21/2018 07:42 AM, Matthew Wilcox wrote: > > > This prompted me to write a patch I've been meaning to do for a while, > > > allocating large pages if they're available to satisfy vmalloc. I thought > > > it would save on touching multiple struct pages, but it turns out that > > > the checking code we currently have in the free_pages path requires you > > > to have initialised all of the tail pages (maybe we can make that code > > > conditional ...) > > > > What the concept here? If we can use high-order pages for vmalloc() at > > the moment, we *should* use them? > > Right. It helps with fragmentation if we can keep higher-order > allocations together.
Hmm, wouldn't it help if we made vmalloc pages migrateable instead? That would help the compaction and get us to a lower fragmentation longterm without playing tricks in the allocation path.
> > One of the coolest things about vmalloc() is that it can do large > > allocations without consuming large (high-order) pages, so it has very > > few side-effects compared to doing a bunch of order-0 allocations. This > > patch seems to propose removing that cool thing. Even trying the > > high-order allocation could kick off a bunch of reclaim and compaction > > that was not there previously. > > Yes, that's one of the debatable things. It'd be nice to have a GFP > flag that stopped after calling get_page_from_freelist() and didn't try > to do compaction or reclaim.
GFP_NOWAIT, you mean?
> > If you could take this an only _opportunistically_ allocate large pages, > > it could be a more universal win. You could try to make sure that no > > compaction or reclaim is done for the large allocation. Or, maybe you > > only try it if there are *only* high-order pages in the allocator that > > would have been broken down into order-0 *anyway*. > > > > I'm not sure it's worth it, though. I don't see a lot of folks > > complaining about vmalloc()'s speed or TLB impact. > > No, I'm not sure it's worth it either, although Konstantin's mail > suggesting improvements in fork speed were possible by avoiding vmalloc > reminded me that I'd been meaning to give this a try.
Maybe we should consider kvmalloc for the kernel stack? -- Michal Hocko SUSE Labs
| |