Hi Ilya, On 11/4/2024 6:07 PM, Ilya Leoshkevich wrote: > On Mon, 2024-11-04 at 09:34 +0800, Hou Tao wrote: >> Hi, >> >> On 11/3/2024 2:04 PM, Byeonguk Jeong wrote: >>> Hi, >>> >>> The selftest "verifier_bits_iter/bad words" has been failed with >>> retval 115, while I did not touched anything but a comment. >>> >>> Do you have any idea why it failed? I am not sure whether it >>> indicates >>> any bugs in the kernel. >>> >>> Best, >>> Byeonguk >> Sorry for the inconvenience. It seems the test case >> "verifier_bits_iter/bad words" is flaky. It may fail randomly, such >> as >> in [1]. I think calling bpf_probe_read_kernel_common() on 3GB addr >> under >> s390 host may succeed and the content of the memory address will >> decide >> whether the test case will succeed or not. Do not know the reason why >> reading 3GB address succeeds under s390. Hope to get some insight >> from >> Ilya. I think we could fix the failure first by using NULL as the >> address of bad words just like null_pointer test case does. Will >> merge >> the test in bad_words into the null_pointer case. > Hi, > > s390 kernel runs in a completely separate address space, there is no > user/kernel split at TASK_SIZE. The same address may be valid in both > the kernel and the user address spaces, there is no way to tell by > looking at it. The config option related to this property is > ARCH_HAS_NON_OVERLAPPING_ADDRESS_SPACE. > > Also, unfortunately, 0 is a valid address in the s390 kernel address > space. Thanks for the detailed explanation. It seems both arm and x86 have select ARCH_HAS_NON_OVERLAPPING_ADDRESS_SPACE. > > I wonder if we could use -4095 as an address that cannot be > dereferenced on all platforms? I have tested it in both arm64 and x86-64 that reading from -4095 by using copy_from_kernel_nofault() will return -EFAULT . For s390, I hope the bpf CI could help to test it. Will post a fix patch later. > > Best regards, > Ilya
