Am Montag, dem 23.06.2025 um 14:05 +0200 schrieb Gert Wollny: > On Fri, 2025-06-20 at 22:22 +0200, Lucas Stach wrote: > > > > > @@ -1807,6 +1808,11 @@ static int etnaviv_gpu_bind(struct device > > > *dev, struct device *master, > > > ret = -ENXIO; > > > goto out_sched; > > > } > > > + > > > + if (etnaviv_flop_reset_ppu_require(&gpu->identity) && > > > + !priv->flop_reset_data_ppu) > > > + etnaviv_flop_reset_ppu_init(priv); > > > + > > I don't see why you would need to do this in the bind callback. You > > should be able to move this to etnaviv_gpu_init(), so you have the > > needed identification data. gpu_init is also executed serially over > > all GPUs in the device, so there is no problem with potential races > > there. > > I moved this here because you wrote before: > > > But then you should allocate this buffer > > in etnaviv_gpu_bind to avoid races between GPUs starting up and > > trying to allocate this global memory. > Yea, sorry about this. I hadn't noticed the dependency on the HW identification when I wrote this.
> If etnaviv_gpu_init() is fine, I'll move it there. I'm not saying that I may not again miss some implicit dependency, but as far as I can see right now moving it there should be fine. gpu_init() fulfills the same properties with regard to init ordering between the GPUs as gpu_bind(). Regards, Lucas
