On Tue, Jan 24, 2023 at 7:47 AM Yaniv Agman <yanivagman@xxxxxxxxx> wrote: > > בתאריך יום ג׳, 24 בינו׳ 2023 ב-14:30 מאת Alexei Starovoitov > <alexei.starovoitov@xxxxxxxxx>: > > > > On Mon, Jan 23, 2023 at 2:03 PM Yaniv Agman <yanivagman@xxxxxxxxx> wrote: > > > > > > בתאריך יום ב׳, 23 בינו׳ 2023 ב-23:25 מאת Jakub Sitnicki > > > <jakub@xxxxxxxxxxxxxx>: > > > > > > > > On Mon, Jan 23, 2023 at 11:01 PM +02, Yaniv Agman wrote: > > > > > בתאריך יום ב׳, 23 בינו׳ 2023 ב-22:06 מאת Martin KaFai Lau > > > > > <martin.lau@xxxxxxxxx>: > > > > >> > > > > >> On 1/23/23 9:32 AM, Yaniv Agman wrote: > > > > >> >>> interrupted the first one. But even then, I will need to find a way to > > > > >> >>> know if my program currently interrupts the run of another program - > > > > >> >>> is there a way to do that? > > > > >> May be a percpu atomic counter to see if the bpf prog has been re-entered on the > > > > >> same cpu. > > > > > > > > > > Not sure I understand how this will help. If I want to save local > > > > > program data on a percpu map and I see that the counter is bigger then > > > > > zero, should I ignore the event? > > > > > > > > map_update w/ BPF_F_LOCK disables preemption, if you're after updating > > > > an entry atomically. But it can't be used with PERCPU maps today. > > > > Perhaps that's needed now too. > > > > > > Yep. I think what is needed here is the ability to disable preemption > > > from the bpf program - maybe even adding a helper for that? > > > > I'm not sure what the issue is here. > > Old preempt_disable() doesn't mean that one bpf program won't ever > > be interrupted by another bpf prog. > > Like networking bpf prog in old preempt_disable can call into something > > where there is a kprobe and another tracing bpf prog will be called. > > The same can happen after we switched to migrate_disable. > > One difference here is that in what you describe the programmer can > know in advance which functions might call others and avoid that or > use other percpu maps, but if preemption can happen between functions > which are not related to one another (don't have a relation of caller > and callee), then the programmer can't have control over it Could you give a specific example?
