Hi Christoph, On Mon, Mar 05, 2012 at 08:48:33AM -0600, Christoph Lameter wrote: > On Sun, 4 Mar 2012, Minchan Kim wrote: > > > I read this thread and I feel the we don't reach right point. > > I think it's not a compound page problem. > > We can face above problem where we allocates big order page without __GFP_COMP > > and free middle page of it. > > Yes we can do that and doing such a thing seems to be more legitimate > since one could argue that the user did not request an atomic allocation > unit from the page allocator and therefore the freeing of individual > pages in that group is permissible. If memory serves me right we do that > sometimes. To be leitimate, user have to handle subpages's ref counter well. But I think it's not desirable. If user want it, he should use split_page instead of modifying ref counter directly. > > However if compound pages are requested then such an atomic allocation > unit *was* requested and the page allocator should not allow to free > individual pages. Yes. In fact, I am not sure this problem is related to compound page. If it is compound page, tail page's ref count should be zero. When user calls __free_pages in tail page by mistake, it should not pass into __free_pages_ok but reference count would be underflow. Later, when head page is freed, we could catch it in free_pages_check. So I had a question to Namhyung that he can see bad page message by PG_slab when he uses SLUB with his patch. If the problem still happens, something seems to modify tail page's ref count directly without get_page. It's apparently BUG. -- To unsubscribe, send a message with 'unsubscribe linux-mm' in the body to majordomo@xxxxxxxxx. For more info on Linux MM, see: http://www.linux-mm.org/ . Fight unfair telecom internet charges in Canada: sign http://stopthemeter.ca/ Don't email: <a href=mailto:"dont@xxxxxxxxx";> email@xxxxxxxxx </a>