On 04/05/2013 04:06 PM, Johannes Weiner wrote: > On Fri, Apr 05, 2013 at 02:01:11PM +0400, Glauber Costa wrote: >> Right now, allocation for page_cgroup is a bit complicated, dependent on >> a variety of system conditions: >> >> For flat memory, we are likely to need quite big pages, so the page >> allocator won't cut. We are forced to init flatmem mappings very early, >> because if we run after the page allocator is in place those allocations >> will be denied. Flatmem mappings thus resort to the bootmem allocator. >> >> We can fix this by using vmalloc for flatmem mappings. However, we now >> have the situation in which flatmem mapping allocate using vmalloc, but >> sparsemem may or may not allocate with vmalloc. It will try the >> page_allocator first, and retry vmalloc if it fails. > > Vmalloc space is a precious resource on 32-bit systems and harder on > the TLB than the identity mapping. > > It's a last resort thing for when you need an unusually large chunk of > contiguously addressable memory during runtime, like loading a module, > buffers shared with userspace etc.. But here we know, during boot > time, the exact amount of memory we need for the page_cgroup array. > > Code cleanup is not a good reason to use vmalloc in this case, IMO. > This is indeed a code cleanup, but a code cleanup with a side goal: freeing us from the need to register page_cgroup mandatorily at init time. This is done because page_cgroup_init_flatmem will use the bootmem allocator, to avoid the page allocator limitations. What I can try to do, and would happily do, is to try a normal page allocation and then resort to vmalloc if it is too big. Would that be okay to you ? -- To unsubscribe from this list: send the line "unsubscribe cgroups" in the body of a message to majordomo@xxxxxxxxxxxxxxx More majordomo info at http://vger.kernel.org/majordomo-info.html
