On Tue, Apr 18, 2023 at 08:03:53AM -0700, Sean Christopherson wrote: > On Mon, Apr 17, 2023, Aaron Lewis wrote: > > On Thu, Mar 23, 2023, Sean Christopherson wrote: > > > > +static char *number(char *str, long num, int base, int size, int precision, > > > > + int type) > > > > > > Do we actually need a custom number()? I.e. can we sub in a libc equivalent? > > > That would reduce the craziness of this file by more than a few degrees. > > > > Yeah, I think we need it. One of the biggest problems I'm trying to avoid > > here is the use of LIBC in a guest. Using it always seems to end poorly > > because guests generally don't set up AVX-512 or a TLS segmet, nor should > > they have to. Calling into LIBC seems to require both of them too often, > > so it seems like it's better to just avoid it. > > True, we'd probably end up in a world of hurt. > > I was going to suggest copy+pasting from a different source, e.g. musl, in the > hopes of reducing the crazy by a degree, but after looking at the musl source, > that's a terrible idea :-) > > And copying from the kernel has the advantage of keeping any bugs/quirks that > users may be expecting and/or relying on. What about trying to use tools/include/nolibc/? Maybe we could provide our own tools/include/nolibc/arch-*.h files where the my_syscall* macros get implemented with ucalls, and then the ucalls would implement the syscalls, possibly just forwarding the parameters to real syscalls. We can implement copy_from/to_guest() functions to deal with pointer parameters. Thanks, drew
