On 2023-06-04 13:05:18+0200, Willy Tarreau wrote: > Hi Zhangjin, > > On Tue, May 30, 2023 at 06:47:38PM +0800, Zhangjin Wu wrote: > > Hi, Willy, Thomas > > > > This is not really for merge, but only let it work as a demo code to > > test whether it is possible to restore the next test when there is a bad > > pointer access in user-space [1]. > > > > Besides, a new 'run' command is added to 'NOLIBC_TEST' environment > > variable or arguments to control the running iterations, this may be > > used to test the reentrancy issues, but no failures found currently ;-) > > Since the tests we're running are essentially API tests, I'm having > a hard time seeing in which case it can be useful to repeat the tests. > I'm not necessarily against doing it, I'm used to repeating tests for > example in anything sensitive to timing or race conditions, it's just > that here I'm not seeing the benefit. And the fact you found no failure > is rather satisfying because the opposite would have surprised me. > > Regarding the efault handler, I don't think it's a good idea until we > have signal+longjmp support in nolibc. Because running different tests > with different libcs kind of defeats the purpose of the test in the > first place. The reason why I wanted nolibc-test to be portable to at > least one other libc is to help the developer figure if a failure is in > the nolibc syscall they're implementing or in the test itself. Here if > we start to say that some parts cannot be tested similarly, the benefit > disappears. > > I mentioned previously that I'm not particularly impatient to work on > signals and longjmp. But in parallel I understand how this can make the > life of some developers easier and even allow to widen the spectrum of > some tests. Thus, maybe in the end it could be beneficial to make progress > on this front and support these. We should make sure that this doesn't > inflate the code base however. I guess I'd be fine with ignoring libc- > based restarts on EINTR, alt stacks and so on and keeping this minimal > (i.e. catch a segfault/bus error/sigill in a test program, or a Ctrl-C > in a tiny shell). > > Just let us know if you think that's something you could be interested > in exploring. There might be differences between architectures, I have > not checked. If the goal is to handle hard errors like segfaults more gracefully, would it not be easier to run each testcase in a subprocess? Then we can just check if the child exited successfully. It should also be completely architecture agnostic. Thomas
