On Wed, 25 Mar 2015, Hugh Dickins wrote: > When you say "tracking back to 3.2.x", I think you mean you've tried as > far back as 3.2.x and found the same behaviour, but not tried further? > > From the source, it looks like this is unchanged since MAP_HUGETLB was > introduced in 2.6.32. And is the same behaviour as you've been given > with hugetlbfs since it arrived in 2.5.46. Went back checking the application logs, an I had to patch the code (the application one - to align size on munmap()) on May 2014. But it might be we started noticing it at that time, because before the allocation pattern was simply monotonic, so it could be it was always there. The bug test app is ten lines of code, to verify that. > The patch looks to me as if it will do what you want, and I agree > that it's displeasing that you can mmap a size, and then fail to > munmap that same size. > > But I tend to think that's simply typical of the clunkiness we offer > you with hugetlbfs and MAP_HUGETLB: that it would have been better to > make a different choice all those years ago, but wrong to change the > user interface now. > > Perhaps others will disagree. And if I'm wrong, and the behaviour > got somehow changed in 3.2, then that's a different story and we > should fix it back. This is not an interface change, in the sense old clients will continue to work. This is simply respecting the mmap(2) POSIX specification, for a feature which has been exposed via mmap(2). - Davide -- 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>