On Wed, May 29, 2019 at 02:23:42PM +0100, Catalin Marinas wrote: > On Wed, May 29, 2019 at 01:42:25PM +0100, Dave P Martin wrote: > > On Tue, May 28, 2019 at 05:34:00PM +0100, Catalin Marinas wrote: > > > On Tue, May 28, 2019 at 04:56:45PM +0100, Dave P Martin wrote: > > > > On Tue, May 28, 2019 at 04:40:58PM +0100, Catalin Marinas wrote: > > > > > > > > [...] > > > > > > > > > My thoughts on allowing tags (quick look): > > > > > > > > > > brk - no > > > > > > > > [...] > > > > > > > > > mlock, mlock2, munlock - yes > > > > > mmap - no (we may change this with MTE but not for TBI) > > > > > > > > [...] > > > > > > > > > mprotect - yes > > > > > > > > I haven't following this discussion closely... what's the rationale for > > > > the inconsistencies here (feel free to refer me back to the discussion > > > > if it's elsewhere). > > > > > > _My_ rationale (feel free to disagree) is that mmap() by default would > > > not return a tagged address (ignoring MTE for now). If it gets passed a > > > tagged address or a "tagged NULL" (for lack of a better name) we don't > > > have clear semantics of whether the returned address should be tagged in > > > this ABI relaxation. I'd rather reserve this specific behaviour if we > > > overload the non-zero tag meaning of mmap() for MTE. Similar reasoning > > > for mremap(), at least on the new_address argument (not entirely sure > > > about old_address). > > > > > > munmap() should probably follow the mmap() rules. > > > > > > As for brk(), I don't see why the user would need to pass a tagged > > > address, we can't associate any meaning to this tag. > > > > > > For the rest, since it's likely such addresses would have been tagged by > > > malloc() in user space, we should allow tagged pointers. > > > > Those arguments seem reasonable. We should try to capture this > > somewhere when documenting the ABI. > > > > To be clear, I'm not sure that we should guarantee anywhere that a > > tagged pointer is rejected: rather the behaviour should probably be > > left unspecified. Then we can tidy it up incrementally. > > > > (The behaviour is unspecified today, in any case.) > > What is specified (or rather de-facto ABI) today is that passing a user > address above TASK_SIZE (e.g. non-zero top byte) would fail in most > cases. If we relax this with the TBI we may end up with some de-facto I may be being too picky, but "would fail in most cases" sounds like "unspecified" ? > ABI before we actually get MTE hardware. Tightening it afterwards may be > slightly more problematic, although MTE needs to be an explicit opt-in. > > IOW, I wouldn't want to unnecessarily relax the ABI if we don't need to. So long we don't block foreseeable future developments unnecessarily either -- I agree there's a balance to be struck. I guess this can be reviewed when we have nailed down the details a bit further. Cheers ---Dave _______________________________________________ amd-gfx mailing list amd-gfx@xxxxxxxxxxxxxxxxxxxxx https://lists.freedesktop.org/mailman/listinfo/amd-gfx
