Linus Torvalds <torvalds@xxxxxxxxxxxxxxxxxxxx> writes: > On Tue, Jun 19, 2018 at 8:24 PM Eric W. Biederman <ebiederm@xxxxxxxxxxxx> wrote: >> >> I stared at this code for quite a while and I finally concluded that the >> best course forward is to simply things and remove the internal kernel >> mount of proc. The internal mount of proc is directly responsible for >> this regression and it has been the source of pain over the years. > > This is not the kind of patch that I'm willing to take outside the > merge window. This is *way* too subtle, and making sysctl do a > kern_mount()/kern_umount() seems odd. I understand the feedback about breaking up the patch and the concern about the race with pid->count. I don't understand the feedback about only accepting something like this during the merge window. The entire point of my change was to remove subtlety. The code was very straight forward to test. This is a silent regression of a security feature so it is possible some people have upgraded their kernel and not noticed the regression but are affected by the information leak not honoring hidepid introduces. That seems to me to be a candidate for stable and thus an rc kernel. Would you prefer a patch that does less towards fixing the root cause for now and to be backported to stable? > The pid->count test also looks potentially racy to me. The function proc_flush_task is already racy, it is just an optimization that needs to work the vast majority of the time or we get lots of stale useless cached dentries in proc. So I don't think a little race between testing pid->count and someone accessing a proc inode matters. They could always perform the access after proc_flush_task is done and before unhash_process runs, and achieve the same effect. Though in retrospect my testing showed processes acessing proc self from libc or something so the pid->count optimization never really hit. So it is probably better just to remove it. The kern_mount/kern_umount are definitely odd and not my favorite. But the code does work. It is my intention and hope that they can both the uml and the sysctl(2) code can both be removed. I need to double check but I don't think there are even any enterprise kernels that enable sysctl(2) support in the kernel any more. Eric
