Al Viro [off-list ref] writes:

On Mon, Jan 04, 2021 at 03:21:22PM -0800, Stephen Brennan wrote:

quoted

The pid_revalidate() function drops from RCU into REF lookup mode. When
many threads are resolving paths within /proc in parallel, this can
result in heavy spinlock contention on d_lockref as each thread tries to
grab a reference to the /proc dentry (and drop it shortly thereafter).

Investigation indicates that it is not necessary to drop RCU in
pid_revalidate(), as no RCU data is modified and the function never
sleeps. So, remove the LOOKUP_RCU check.

Umm...  I'm rather worried about the side effect you are removing here -
you are suddenly exposing a bunch of methods in there to RCU mode.
E.g. is proc_pid_permission() safe with MAY_NOT_BLOCK in the mask?
generic_permission() call in there is fine, but has_pid_permission()
doesn't even see the mask.  Is that thing safe in RCU mode?  AFAICS,
this
static int selinux_ptrace_access_check(struct task_struct *child,
                                     unsigned int mode)
{
        u32 sid = current_sid();
        u32 csid = task_sid(child);

        if (mode & PTRACE_MODE_READ)
                return avc_has_perm(&selinux_state,
                                    sid, csid, SECCLASS_FILE, FILE__READ, NULL);

        return avc_has_perm(&selinux_state,
                            sid, csid, SECCLASS_PROCESS, PROCESS__PTRACE, NULL);
}
is reachable and IIRC avc_has_perm() should *NOT* be called in RCU mode.
If nothing else, audit handling needs care...

And LSM-related stuff is only a part of possible issues here.  It does need
a careful code audit - you are taking a bunch of methods into the conditions
they'd never been tested in.  ->permission(), ->get_link(), ->d_revalidate(),
->d_hash() and ->d_compare() instances for objects that subtree.  The last
two are not there in case of anything in /proc/<pid>, but the first 3 very
much are.

You're right, this was a major oversight on my part. The main motivation
of this patch is to reduce contention on the /proc dentry, which occurs
directly after d_revalidate() returns -ECHILD the first time in
lookup_fast(). To drop into ref mode, we call unlazy_child(), while
nd->path still refers to /proc and dentry refers to /proc/PID. Grabbing
a reference to /proc is the heart of the contention issue.

But directly after a successful d_revalidate() in lookup_fast(), we
return and go to step_into(), which assigns the /proc/PID dentry to
nd->path. After this point, any unlazy operation will not try to grab
the /proc dentry, resulting in significantly less contention.

So it would already be a significant improvement if we kept this change
to pid_revalidate(), and simply added checks to bail out of each of the
other procfs methods if we're in LOOKUP_RCU. Would that be an acceptable
change for you?