On Thu, May 02, 2019 at 11:02:40AM -0700, Linus Torvalds wrote:
On Thu, May 2, 2019 at 9:21 AM Peter Zijlstra peterz@infradead.org wrote:
TL;DR, on x86_32 kernel->kernel IRET frames are only 3 entries and do not include ESP/SS, so not only wasn't regs->sp setup, if you changed it it wouldn't be effective and corrupt random stack state.
Indeed, the 32-bit case for same-RPL exceptions/iret is entirely different, and I'd forgotten about that.
And honestly, this makes the 32-bit case much worse. Now the entry stack modifications of int3 suddenly affect not just the entry, but every exit too.
This is _exactly_ the kind of subtle kernel entry/exit code I wanted us to avoid.
I actually love this patch (absent the bugs). This is already something that has been sorely needed for years.
The "struct pt_regs is incomplete on x86-32" thing is a monstrosity which has long been a source of confusion and bugs. Sure, this patch adds some complexity to the entry code, but on the other hand it actually makes it possible to use pt_regs sanely: regs->sp is no longer uninitialized. So a class of (very non-obvious) bugs is eliminated.
I don't think it would make sense to make this change for int3 only, because the benefits are global.