On Tue, Jan 07, 2020 at 08:32:04AM -0800, Tejun Heo wrote: > On Mon, Dec 23, 2019 at 07:15:03AM +0100, Christian Brauner wrote: > > +static struct cgroup *cgroup_get_from_file(struct file *f) > > +{ > > + struct cgroup_subsys_state *css; > > + struct cgroup *cgrp; > > + > > + css = css_tryget_online_from_dir(f->f_path.dentry, NULL); > > + if (IS_ERR(css)) > > + return ERR_CAST(css); > > + > > + cgrp = css->cgroup; > > + if (!cgroup_on_dfl(cgrp)) { > > + cgroup_put(cgrp); > > + return ERR_PTR(-EBADF); > > + } > > + > > + return cgrp; > > +} > > It's minor but can you put this refactoring into a separate patch? Yep, will do. > > ... > > +static int cgroup_css_set_fork(struct task_struct *parent, > > + struct kernel_clone_args *kargs) > > + __acquires(&cgroup_mutex) __acquires(&cgroup_threadgroup_rwsem) > > +{ > > + int ret; > > + struct cgroup *dst_cgrp = NULL, *src_cgrp; > > + struct css_set *cset; > > + struct super_block *sb; > > + struct file *f; > > + > > + if (kargs->flags & CLONE_INTO_CGROUP) { > > + ret = mutex_lock_killable(&cgroup_mutex); > > + if (ret) > > + return ret; > > + } > > I don't think this is necessary. cgroup_mutex should always only be > held for a finite enough time; otherwise, processes would get stuck on > random cgroupfs accesses or even /proc/self/cgroup. Ok, so a simple mutex_lock() should suffice then. > > ... > > + spin_lock_irq(&css_set_lock); > > + src_cgrp = task_cgroup_from_root(parent, &cgrp_dfl_root); > > + spin_unlock_irq(&css_set_lock); > > You can simply do cset->dfl_root here, which is consistent with other > code paths which know that they want the dfl cgroup. Ah, great! > > > + ret = cgroup_attach_permissions(src_cgrp, dst_cgrp, sb, > > + !!(kargs->flags & CLONE_THREAD)); > > + if (ret) > > + goto err; > > So, the existing perm check depends on the fact that for the write > operation to have started, it already should have passed write perm > check on the destination cgroup.procs file. We're missing that here, > so we prolly need to check that explicitly. I need to look into this before I can say yay or nay. :) > > > @@ -214,13 +215,21 @@ static void pids_cancel_attach(struct cgroup_taskset *tset) > > +static int pids_can_fork(struct task_struct *parent, struct task_struct *child, > > + struct kernel_clone_args *args) > > { > > + struct css_set *new_cset = NULL; > > struct cgroup_subsys_state *css; > > struct pids_cgroup *pids; > > int err; > > > > - css = task_css_check(current, pids_cgrp_id, true); > > + if (args) > > + new_cset = args->cset; > > + > > + if (!new_cset) > > + css = task_css_check(current, pids_cgrp_id, true); > > + else > > + css = new_cset->subsys[pids_cgrp_id]; > > Heh, this kinda sucks. Would it be better to pass in the new css into > the callbacks rather than clone args? Hm, maybe. My reasoning was that the can_fork callbacks are really only ever used when - well - fork()ing/clone{3}()ing. Additionally, I was trying to make sure that struct css_set doesn't show up in too many places outside of cgroup core. But I'm fine with changing this to just take the css_set directly. Let's try that... > > > diff --git a/kernel/fork.c b/kernel/fork.c > > index 2508a4f238a3..1604552f7cd3 100644 > > --- a/kernel/fork.c > > +++ b/kernel/fork.c > > @@ -2165,16 +2165,15 @@ static __latent_entropy struct task_struct *copy_process( > > INIT_LIST_HEAD(&p->thread_group); > > p->task_works = NULL; > > > > - cgroup_threadgroup_change_begin(current); > > /* > > * Ensure that the cgroup subsystem policies allow the new process to be > > * forked. It should be noted the the new process's css_set can be changed > > * between here and cgroup_post_fork() if an organisation operation is in > > * progress. > > */ > > - retval = cgroup_can_fork(p); > > + retval = cgroup_can_fork(current, p, args); > > if (retval) > > - goto bad_fork_cgroup_threadgroup_change_end; > > + goto bad_fork_put_pidfd; > > > > /* > > * From this point on we must avoid any synchronous user-space > > Maybe we can move these changes into a prep patch together with the > get_from_file change so that this patch only contains the actual > feature implementation? Should be doable! > > Other than that, looks good to me. Once the above review points are > addressed and Oleg is okay with it, I'll be happy to route this > through the cgroup tree. > > Thanks so much for working on this. This is really cool. Thanks and I agree! :) Christian