On Fri, Jul 21, 2023 at 05:05:08PM +0800, Yafang Shao wrote: > On Fri, Jul 21, 2023 at 4:55 PM Oleg Nesterov <oleg@xxxxxxxxxx> wrote: > > > > Sorry for noise, apparently I haven't woken up yet, > > > > On 07/21, Oleg Nesterov wrote: > > > > > > Or. We can simply kill uprobe_multi_link_filter(). In this case uprobe_register() > > > can touch CLONE_VM'ed mm's we are not interested in, but everything should work > > > correctly. > > > > please ignore CLONE_VM'ed above, I have no idea why did I say this... > > > > Yes, everything should work "correctly", but if you add a probe into (say) libc.so, > > every task will be penalized, so I don't think this is a good option. > > > > Thanks for your explanation. > Currently we can filter out the tgid in bpf prog, but if we could > filter out it earlier, the prog won't be triggered, that would take > less overhead. > However, it is fine by me if tgid won't be supported. right, I'll check on this.. I think we can add this later and have it enabled by new bit in attr->link_create.uprobe_multi.flags jirka
