On Tue, Jun 18, 2024 at 06:37:29PM +0200, Florian Weimer wrote:
| First source: | #define _GNU_SOURCE 1 | #include <sys/mman.h> | | | Second source: | #define _GNU_SOURCE 1 | #include <linux/mman.h> | | | Only in first source: MAP_ABOVE4G | Different values for MAP_HUGE_16GB: 2281701376 != -2013265920
The reason is that the kernel changed the definition of MAP_HUGE_16GB. Older UAPI headers use an undefined expression which is treated by compilers as a signed int, hence t he discrepancy.
Oops, sorry I never tested this against older kernel headers.
This was fixed on the kernel side in this commit:
commit 710bb68c2e3a24512e2d2bae470960d7488e97b1 Author: Matthias Goergens matthias.goergens@gmail.com Date: Mon Sep 5 11:19:04 2022 +0800
hugetlb_encode.h: fix undefined behaviour (34 << 26)Left-shifting past the size of your datatype is undefined behaviour in C. The literal 34 gets the type `int`, and that one is not big enough to be left shifted by 26 bits. An `unsigned` is long enough (on any machine that has at least 32 bits for their ints.) For uniformity, we mark all the literals as unsigned. But it's only really needed for HUGETLB_FLAG_ENCODE_16GB. Thanks to Randy Dunlap for an initial review and suggestion. Link: https://lkml.kernel.org/r/20220905031904.150925-1-matthias.goergens@gmail.co... Signed-off-by: Matthias Goergens matthias.goergens@gmail.com Acked-by: Randy Dunlap rdunlap@infradead.org Cc: Mike Kravetz mike.kravetz@oracle.com Cc: Muchun Song songmuchun@bytedance.com Signed-off-by: Andrew Morton akpm@linux-foundation.org
Sounds like this kernel commit should have been backported to stable but never was. The glibc test caught the issue. I'll go ahead and backport the fix to the linux stable branches.
Not sure what to do about this on the glibc side. Can we waive this in the CI, or should try to fix up the glibc test?
Hopefully this can be done until the fix propagates to the older kernel's UAPI.
Thanks Carlos Llamas