Ok, so according to the manpage, mremap generates EFAULT when "the range old_address to old_address+old_size is an invalid virtual memory address for this process". This is what the kernel does for the stack guard. However, the mappings in setup_arg_pages() will only ever provoke an ENOMEM, because there is no artifical way to turn a page into an invalid address. So as long as target bits >= host bits, this works as expected and EFAULT is generated, because then mremap is basically passed through and the kernel responds directly. But when reserved_va is set, this needs to be special-cased to fake kernel behavior.
I'm open to other suggestions. I also understand that the code duplication in elfload.c and mmap.c to handle this is undesirable, but the most viable alternative seems to be introducing more globals. On 6/15/20 11:28 PM, Tobias Koch wrote: > Hm, I see I need to have another look at this :) > > On 6/15/20 10:17 AM, Tobias Koch wrote: >> Hi Laurent, >> >> the code in musl libc probing the stack is in >> >> https://git.musl-libc.org/cgit/musl/plain/src/thread/pthread_getattr_np.c >> >> The setup in elfload.c does work, but only when reserved_va is not set. In >> that case, any stack guard violation is >> handled by the host kernel and thus results in the expected EFAULT. >> >> However, in case of e.g. a 32bit target being emulated on a 64bit host, >> reserved_va is set and the current code in >> mmap.c will only produce a more generic ENOMEM, deviating from the kernel's >> behavior. >> >> >> On 5/7/20 5:35 PM, Laurent Vivier wrote: >>> Le 05/03/2020 à 22:05, Tobias Koch a écrit : >>>> If the address range starting at old_address overlaps with the stack guard >>>> it >>>> is invalid and mremap must fail with EFAULT. The musl c library relies on >>>> this >>>> behavior to detect the stack size, which it does by doing consecutive >>>> mremaps >>>> until it hits the stack guard. Without this patch, software (such as the >>>> Ruby >>>> interpreter) that calls pthread_getattr_np under musl will crash on 32 bit >>>> targets emulated on a 64 bit host. >>> Could you share some pointers to the code that is doing this? >>> >>> We have already this kind of code in linux-user/elfload.c, >>> setup_arg_pages(): could you check why it doesn't work?