On Mon, Aug 13, 2018 at 9:37 PM, Ravi Bangoria <ravi.bangoria@xxxxxxxxxxxxx> wrote: > Hi Song, > > On 08/13/2018 10:42 PM, Song Liu wrote: >> On Mon, Aug 13, 2018 at 6:17 AM, Oleg Nesterov <oleg@xxxxxxxxxx> wrote: >>> On 08/13, Ravi Bangoria wrote: >>>> >>>>> But damn, process creation (exec) is trivial. We could add a new uprobe_exec() >>>>> hook and avoid delayed_uprobe_install() in uprobe_mmap(). >>>> >>>> I'm sorry. I didn't get this. >>> >>> Sorry for confusion... >>> >>> I meant, if only exec*( could race with _register(), we could add another uprobe >>> hook which updates all (delayed) counters before return to user-mode. >>> >>>>> Afaics, the really problematic case is dlopen() which can race with _register() >>>>> too, right? >>>> >>>> dlopen() should internally use mmap() right? So what is the problem here? Can >>>> you please elaborate. >>> >>> What I tried to say is that we can't avoid uprobe_mmap()->delayed_uprobe_install() >>> because dlopen() can race with _register() too, just like exec. >>> >>> Oleg. >>> >> >> How about we do delayed_uprobe_install() per file? Say we keep a list >> of delayed_uprobe >> in load_elf_binary(). Then we can install delayed_uprobe after loading >> all sections of the >> file. > > I'm not sure if I totally understood the idea. But how this approach can > solve dlopen() race with _register()? > > Rather, making delayed_uprobe_list an mm field seems simple and effective > idea to me. The only overhead will be list_empty(mm->delayed_list) check. > > Please let me know if I misunderstood you. > > Thanks, > Ravi I misunderstood the problem here. I guess mm->delayed_list is the easiest solution of the race condition. Thanks, Song
