On Mon, 30 Sep 2002, Peter Wemm wrote: > Daniel Eischen wrote: > > At the end is a potentially longer term fix for the ABI > > breakage that was introduced when the i386 mcontext_t > > was changed/enlarged. > > > - ret = set_fpcontext(td, &ucp->uc_mcontext); > > - if (ret != 0) > > - return (ret); > > + /* > > + * Intentionally ignore the error to keep binary > > + * compatibility with applications that fiddle with > > + * the FPU save area in the context. The kernel > > + * now saves the FPU state in the context, but it > > + * gets corrupted by those applications that try > > + * to work around the kernel NOT saving it. > > + */ > > + (void)set_fpcontext(td, &ucp->uc_mcontext); > > Maybe we could have something like this instead? > > ret = set_fpcontext(td, &ucp->uc_mcontext); > #if !defined(COMPAT_FREEBSD4) && !defined(COMPAT_43) > if (ret != 0) > return (ret); > #endif > > ie: ignore the error only if we have to be compatable.
Sure that's totally doable. It might not be enough to just call set_fpcontext() and ignore the error. Thinking a bit more about it, the mc_len, mc_fpformat, and mc_ownedfp fields now occupy the first couple of slots where fpregs[] used to be. The format of an fnsave() stores the control, status and tag words in these slots. There are 32-bits of storage allocated for each of these, but the fnsave (according to what I see in npx.h), only uses the lower 16 bits. It might be possible to save a control word or status word that turn out to be valid for mc_fpformat or mc_ownedfp (0, 1, or 2). In this case we'd think the FP context was valid, and try to restore it (it would be trashed). I think if we put some magic in the upper 16 bits of mc_ownedfp, mc_fpformat, then we could prevent this. > Longer term, I was thining that we could/should do what sparc64 does, ie: > libc provides the trampoline and it can then call the correct sigreturn > syscall. That means we add a new sigreturn syscall each time we > significantly break the sigreturn ABI (as in this case) and applications > will be able to use the correct one. Paired with a new sigaction syscall > which would specify the "new" context format we can then be future proof. Sounds good. If we added a new sigaction and sigreturn now, we can still do the same thing, without having the trampoline in libc. I thought the point of having the trampoline in libc would prevent having to create new syscalls... -- Dan Eischen To Unsubscribe: send mail to [EMAIL PROTECTED] with "unsubscribe freebsd-current" in the body of the message
