[uclibc-ng-devel] Re: [RFC 0/2] xtensa: add FDPIC ABI support

Hi Waldemar, On Mon, May 6, 2024 at 11:17 AM Max Filippov <jcmvbkbc(a)gmail.com> wrote: Most of the issues with the tests in the above list are caused by using direct ( -d ) mode to run them when targeting noMMU, but it seems to me that it doesn't have to be like that: even on noMMU the test skeleton can do vfork and exec itself in direct mode. I'll post a patch that implements this change, with that patch most of the above tests run to successful completion. There's still a few failures I observe with the xtensa FDPIC port even with that change: nptl/tst-attr3 -- fails because thread stack guard size in the test is the same as the default thread stack size. Building uclibc with bigger default stack size fixes that failure. nptl/tst-basic7 -- fails because it sets prlimit(RLIMIT_AS) to 10MB and tries to exhaust it, but mmap doesn't fail on noMMU until it runs out of memory. It looks like a kernel bug to me (there's no call to may_expand_vm() from the nommu do_mmap version). If it's supposed to be like that on the kernel side it means that this test is not applicable on a noMMU target. nptl/tst-cancel7, nptl/tst-cancelx7 -- fail because it expects that a shell started by system() execs the program without forking, so that killing it releases the file lock it might have taken. It seems to me that it relies on implementation details (not even the libc implementation details) and the test should be fixed. nptl/tst-mqueue6 -- fails because mq_notify() called with notification->sigev_notify set to SIGEV_THREAD fails with ENOSYS, because init_mq_netlink() calls pthread_atfork(), which is not implemented on noMMU. It seems to me that either pthread_atfork may be avoided on a noMMU target, or it can be implemented on a noMMU target. Or the test is not applicable on a noMMU target. unistd/errno -- fails because of the stupid mistake in the xtensa FDPICI port I introduced at a recent code rebase (missing ';' after FDPIC_LOAD_JUMPTARGET in the SYSCALL_ERROR_HANDLER variant for linuxthreads, that swallowed the following call to __errno_location()) and that surprisingly wasn't caught by the assembler. I fixed it by adding the ';' and also defining USE___THREAD in ASFLAGS in the xtensa NPTL port, so that the appropriate version of SYSCALL_ERROR_HANDLER is used for linuxthreads and for NPTL. -- Thanks. -- Max