Am Freitag, den 18.03.2016, 02:18 +0900 schrieb Joonsoo Kim: > 2016-03-17 1:49 GMT+09:00 Lucas Stach <l.stach@xxxxxxxxxxxxxx>: > > This fixes an annoying race in the CMA code leading to lots of "PFNs busy" > > messages when CMA is used concurrently. This is harmless normally as CMA > > will just retry the allocation at a different place, but it might lead to > > increased fragmentation of the CMA area as well as failing allocations > > when CMA is under memory pressure. > > > > The issue is that test_pages_isolated checks if the range is free by > > checking that all pages in the range are buddy pages. For this to work > > the start pfn needs to be aligned to the higher order buddy page > > including the start pfn if there is any. > > > > This is not a problem for the memory hotplug code, as it always offlines > > whole pageblocks, but CMA may want to isolate a smaller range. So for > > the check to work correctly it down-aligns the start pfn to the higher > > order buddy page. As the zone is not yet locked at that point a > > concurrent page free might coalesce the pages to be checked into an > > even bigger buddy page, causing the check to fail, while all pages are > > in fact buddy pages. > > > > By moving the zone locking to the caller of the test function, it's > > possible to do it before CMA tries to find the proper start page and stop > > any concurrent page coalescing to happen until the check is finished. > > I think that this patch cannot prevent the same race on > isolate_freepages_range(). If buddy merging happens after we > passed test_pages_isolated(), isolate_freepages_range() cannot see > buddy page and will fail. > Your analysis seems correct. I'll fix this patch to hold the zone lock across isolate_freepages_range. Thanks, Lucas -- 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/ . Don't email: <a href=mailto:"dont@xxxxxxxxx";> email@xxxxxxxxx </a>