Paul Kocialkowski writes: > Hi, > > Le vendredi 29 mars 2019 à 16:25 +0100, Daniel Vetter a écrit : >> On Fri, Mar 29, 2019 at 04:02:23PM +0100, Paul Kocialkowski wrote: >> > Hi, >> > >> > On Fri, 2019-03-29 at 09:09 +0000, Daniel Stone wrote: >> > > Hi, >> > > >> > > On Thu, 28 Mar 2019 at 18:53, Daniel Vetter wrote: >> > > > On Thu, Mar 21, 2019 at 04:27:06PM +0100, Paul Kocialkowski wrote: >> > > > > I don't see other options either, and using firstclose/lastopen feels >> > > > > overall more readable in the driver code. >> > > > > >> > > > > I'm not sure there is such a big overhead associated with allocating >> > > > > the binner BO (it seems that the current implementation tries to alloc >> > > > > until the specific memory constraints for the buffer are met, so >> > > > > perhaps that can take time). But if there is, I suppose it's best to >> > > > > have that when starting up rather than delaying the first render >> > > > > operation. >> > > > >> > > > I'm not entirely buying the "we don't need this for fbcon only" argument - >> > > > there's plenty of dumb kms clients too (boot splash and whatever else >> > > > there might be). If you don't want to keep this around I think allocating >> > > > on first non-dumb bo allocation and dropping it when the last such fd >> > > > closes sounds like a much better idea. Needs a bit more state, you need to >> > > > track per drm_file whether you've already allocated a non-dumb bo, and a >> > > > drm_device refcount, but that's not much. Firstopen feels like the wrong >> > > > thing. >> > > > >> > > > Another option would be first_renderopen or something like that, except >> > > > you can also render on the legacy node and I'm not sure how much there's a >> > > > demand for this in other drivers. In the end you have open/close >> > > > callbacks, you can do all the driver specific things you want to do in >> > > > there. >> > > >> > > I'd like to avoid doing it in open where possible, since that hurts >> > > device enumeration from userspace. >> > >> > I've noticed the same issue with firstopen, where our buffer is >> > allocated/liberated a couple of times during enumeration, before the >> > final open that stays alive during use. >> > >> > I'm not sure what is preferable between that and allocating when the >> > GPU is first used. Slowing down the first GPU operation with the >> > allocation does not sound too great either and it feels like the buffer >> > should have been allocated earlier. >> > >> > To me, it feels I think it's better to have delays due to allocation at >> > enumeration / startup rather than later on, but I might be missing some >> > elements to have a clear idea. >> > >> > What do you think? >> >> We'll have the delay somewhere on driver load. Better to have it only once >> (when the driver starts using gem for real), than a bunch of time, at >> least once while enumerating and then once more while actually >> initializing. I think if you allocat this on first non-dumb gem_create, >> and on first command submission (just so evil userspace can't screw up the >> hw too badly), that should be perfectly fine. > > I'm not totally convinced that it's okay to have a delay outside of > init/enumeration, even if it's a smaller delay. You'll have non-dumb buffers created during GL context creation, so early in xserver or other KMS-and-GL-using application init anyway. Seems like a perfectly fine plan to me.