Re: [PATCH bpf] selftests/bpf: DENYLIST.aarch64: Remove fexit_sleep

[Puranjay Mohan]

Alexei Starovoitov <alexei.starovoitov@xxxxxxxxx> writes:

> On Mon, Jul 15, 2024 at 9:32 AM Puranjay Mohan <puranjay@xxxxxxxxxx> wrote:
>>
>>
>> Hi Daniel, Manu
>> I was able to reproduce this issue on KVM and found the root cause for
>> this hang! The other issue that we fixed is unrelated to this hang and
>> doesn't occur on self hosted github runners as they use 48-bit VAs.
>>
>> The userspace test code has:
>>
>>     #define STACK_SIZE (1024 * 1024)
>>     static char child_stack[STACK_SIZE];
>>
>>     cpid = clone(do_sleep, child_stack + STACK_SIZE, CLONE_FILES | SIGCHLD, fexit_skel);
>>
>> arm64 requires the stack pointer to be 16 byte aligned otherwise
>> SPAlignmentFault occurs, this appears as Bus error in the userspace.
>>
>> The stack provided to the clone system call is not guaranteed to be
>> aligned properly in this selftest.
>>
>> The test hangs on the following line:
>>     while (READ_ONCE(fexit_skel->bss->fentry_cnt) != 2);
>>
>> Because the child process is killed due to SPAlignmentFault, the
>> fentry_cnt remains at 0!
>>
>> Reading the man page of clone system call, the correct way to allocate
>> stack for this call is using mmap like this:
>>
>> stack = mmap(NULL, STACK_SIZE, PROT_READ | PROT_WRITE, MAP_PRIVATE | MAP_ANONYMOUS | MAP_STACK, -1, 0);
>>
>> This fixes the issue, I will send a patch to use this and once again
>> remove this test from DENYLIST and I hope this time it fixes it for good.
>
> Wow. Great find. Good to know.
> prog_tests/ns_current_pid_tgid.c has the same issue probably.

Yes, I checked that test as well using gdb and fortunately it gets a 16
byte aligned stack pointer, but this is just luck, so I will send a
patch to fix that test as well.

Thanks,
Puranjay

Attachment: signature.asc
Description: PGP signature