On Wed, Jun 10, 2015 at 12:38:32PM -0700, Andy Lutomirski wrote: > On Wed, Jun 10, 2015 at 12:36 PM, Josh Poimboeuf <jpoimboe@xxxxxxxxxx> wrote: > > On Wed, Jun 10, 2015 at 11:40:06AM -0700, Andi Kleen wrote: > >> Josh Poimboeuf <jpoimboe@xxxxxxxxxx> writes: > >> > >> > On Wed, Jun 10, 2015 at 05:04:12PM +0200, Andi Kleen wrote: > >> >> > > > - duplicate the destination code inside the function > >> >> > > > - convert the jump to a call > >> >> > > > >> >> > > That all won't work for a lot of cases. > >> >> > > >> >> > Hm, could you give an example? > >> >> > >> >> Just a standard *_user exception handler. > >> > > >> > I'm afraid I don't follow. Exception handlers don't work via jump > >> > instructions, but rather via CPU exceptions. > >> > > >> > Or are you talking about something else? > >> > >> Let's take an example: > >> > >> 102: > >> .section .fixup,"ax" > >> 103: addl %ecx,%edx /* ecx is zerorest also */ > >> jmp copy_user_handle_tail > >> .previous > >> > >> _ASM_EXTABLE(100b,103b) > >> _ASM_EXTABLE(101b,103b) > >> > >> The exception handling code is part of the function, but it's out of line. > > > > The jump instruction is in the .fixup section, not in the callable > > function itself. So it doesn't violate the asmvalidate rules. > > It still won't unwind correctly unless .pushsection somehow magically > propagates CFI state. (Does it?) I don't think it does. We'll probably need some intelligence in the CFI generation tooling to deal properly with the extable stuff. > >> > Are you suggesting that we implement this gcc optimization in kernel asm > >> > code? > >> > >> It was how Linux traditionally implemented locking code for example. > >> Have the hot path handle the uncontended fast path, and the slow path > >> call. > >> > >> I don't know if there is much left of it (a lot of it was removed because > >> it was hard to describe in dwarf3, needs dwarf4). But it seems bad > >> to completely disallow it. > >> > >> But yes eventually gcc generated code should use it again, because it's > >> great for icache usage if you measure it correctly at run time > >> (not the broken "size" approach that is unfortunately far too common) > > > > This patch set has no relationship to gcc generated code whatsoever. So > > it doesn't disallow anything there. > > > > For kernel asm code, AFAIK, such a mechanism for hot/cold path > > separation in separate sections doesn't exist today. So it's not > > "disallowed" there either. It's just apparently not currently done. > > > > If somebody were to create such a mechanism, I think we could > > standardize it in such a way that it could be compatible with > > asmvalidate. > > Hopefully true. The entry code is full of tail calls, though. Well, I wasn't talking specifically about tail calls here. But either way, as long as they're not in a callable function (which is the case for most of the entry code), asmvalidate doesn't care. -- Josh -- To unsubscribe from this list: send the line "unsubscribe live-patching" in the body of a message to majordomo@xxxxxxxxxxxxxxx More majordomo info at http://vger.kernel.org/majordomo-info.html
