> On Dec 12, 2018, at 1:36 PM, Edward Cree <ec...@solarflare.com> wrote: > > On 12/12/18 21:15, Nadav Amit wrote: >>> On Dec 12, 2018, at 10:33 AM, Edward Cree <ec...@solarflare.com> wrote: >>> >>> AIUI the outline version uses a tail-call (i.e. jmpq *target) rather than an >>> additional call and ret. So I wouldn't expect it to be too expensive. >>> More to the point, it seems like it's easier to get right than the inline >>> version, and if we get the inline version working later we can introduce it >>> without any API change, much as Josh's existing patches have both versions >>> behind a Kconfig switch. >> I see. For my outlined blocks I used the opposite approach - a call followed >> by jmp > That's what Josh did too. I.e. caller calls the trampoline, which jmps to the > callee; later it rets, taking it back to the caller. Perhaps I wasn't clear. > The point is that there's still only one call and one ret.
Sorry for the misunderstanding. > >>> I was working on the assumption that it would be opt-in, wrapping a macro >>> around indirect calls that are known to have a fairly small number of hot >>> targets. There are plenty of indirect calls in the kernel that are only >>> called once in a blue moon, e.g. in control-plane operations like ethtool; >>> we don't really need to bulk up .text with trampolines for all of them. >> On the other hand, I’m not sure the static_call interface is so intuitive. >> And extending it into “dynamic_call” might be even worse. As I initially >> used an opt-in approach, I can tell you that it was very exhausting. > Well, if it's done with a gcc plugin after all, then it wouldn't be too hard > to make it opt-out. > One advantage of the explicit opt-in dynamic_call, though, which can be seen > in my patch is that multiple call sites can share the same learning-state, > if they're expected to call the same set of functions. An opt-out approach > would automatically give each indirect call statement its own individual BTB. > Either way, I think the question is orthogonal to what the trampolines > themselves look like (and even to the inline vs outline question). Not entirely. If the mechanism is opt-out and outlined, and especially if it also supports multiple targets, you may not want to allocate all the memory for them during build-time, and instead use module memory to allocate them dynamically (that’s what we did).
