On Tue, 28 Sep 2010 16:52:01 -0700 (PDT) David Rientjes <rientjes@xxxxxxxxxx> wrote: > On Tue, 28 Sep 2010, Andrew Morton wrote: > > > > > > So we can definitely remove __GFP_REPEAT for any order-0 allocation and > > > > > it's based on its implementation -- poorly defined as it may be -- and the > > > > > inherit design of any sane page allocator that retries such an allocation > > > > > if it's going to use reclaim in the first place. > > > > > > > > Why was __GFP_REPEAT used in those callsites? What were people trying > > > > to achieve? > > > > > > > > > > I can't predict what they were trying to achieve > > > > Using my super powers it took me all of three minutes. > > > > git://git.kernel.org/pub/scm/linux/kernel/git/torvalds/old-2.6-bkcvs.git > > > > Do `git log > foo', and search foo for GFP_REPEAT. > > > > A couple of interesting ones are: > > > > Ok, so __GFP_REPEAT was used to replace the open-coding of retry loops by > putting that logic into the page allocator, and that logic was > subsequently changed to tie the bit to how many pages were reclaimed and > retry iff we haven't reclaimed the number of pages needed (in my patch, > that would be a single page). It also shows that there was at least one version of the page allocator (written by a very skilled and experienced Linux MM developer) which would simply return NULL when there was no memory. And here you are assuring us that there will never ever be such a version again. > It also shows that the page allocator has infinitely looped for > allocations under PAGE_ALLOC_COSTLY_ORDER since your patch from over seven > years ago. Really don't give a shit what the allocator-of-the-moment happens to be doing. What I care about is that very smart, experienced and hard-working Linux developers decided, a long time ago, that page-table allocations are special, and need special treatment. This is information! What I also care about is lazy MM developers who just rip stuff out without understanding it and without even bothering to make an *attempt* to understand it. > So, given the fact that the PAGE_ALLOC_COSTLY_ORDER logic has existed > since the same time, the semantics of __GFP_REPEAT have changed and are > often misrepresented, and we don't even invoke the __GFP_REPEAT logic for > any of the allocations in my patch since they are oom killable, Probably this is because lazy ignorant MM developers broke earlier intentions without even knowing that they were doing so. > I think my patch should be merged. Well before destroying the information, you tell me: why did MM developers decide that page-table allocations needed special treatment? What experience led them to decide to implement that? Has the problem which they solved been fixed by other means? Are those means sufficiently permanently embedded (see "-aa" above) that we can afford to destroy this piece of information? If we break stuff and destroy this sort of historical knowledge then we get to make the same mistakes every decade or so and the damn thing *never* converges. -- To unsubscribe from this list: send the line "unsubscribe linux-arch" in the body of a message to majordomo@xxxxxxxxxxxxxxx More majordomo info at http://vger.kernel.org/majordomo-info.html