Re: [PATCH] kernel/sys: Optimize do_prlimit lock scope to reduce contention

Re: [PATCH] kernel/sys: Optimize do_prlimit lock scope to reduce contention

[Andrew Morton]

On Wed, 20 Nov 2024 21:21:56 +0800 Zhen Ni <zhen.ni@easystack.cn> wrote:

> Refines the lock scope in the do_prlimit function to reduce
> contention on task_lock(tsk->group_leader). The lock now protects only
> sections that access or modify shared resources (rlim). Permission
> checks (capable) and security validations (security_task_setrlimit)
> are placed outside the lock, as they do not modify rlim and are
> independent of shared data protection.

Let's cc linux-security-module@vger.kernel.org, as we're proposing
altering their locking environment!

> The security_task_setrlimit function is a Linux Security Module (LSM)
> hook that evaluates resource limit changes based on security policies.
> It does not alter the rlim data structure, as confirmed by existing
> LSM implementations (e.g., SELinux and AppArmor). Thus, this function
> does not require locking, ensuring correctness while improving
> concurrency.

Seems sane.

Does any code call do_prlimit() frequently enough for this to matter?

> --- a/kernel/sys.c
> +++ b/kernel/sys.c
> @@ -1481,18 +1481,20 @@ static int do_prlimit(struct task_struct *tsk, unsigned int resource,
>  
>  	/* Holding a refcount on tsk protects tsk->signal from disappearing. */
>  	rlim = tsk->signal->rlim + resource;
> -	task_lock(tsk->group_leader);
>  	if (new_rlim) {
>  		/*
>  		 * Keep the capable check against init_user_ns until cgroups can
>  		 * contain all limits.
>  		 */
>  		if (new_rlim->rlim_max > rlim->rlim_max &&
> -				!capable(CAP_SYS_RESOURCE))
> -			retval = -EPERM;
> -		if (!retval)
> -			retval = security_task_setrlimit(tsk, resource, new_rlim);
> +		    !capable(CAP_SYS_RESOURCE))
> +			return -EPERM;
> +		retval = security_task_setrlimit(tsk, resource, new_rlim);
> +		if (retval)
> +			return retval;
>  	}
> +
> +	task_lock(tsk->group_leader);
>  	if (!retval) {
>  		if (old_rlim)
>  			*old_rlim = *rlim;

Re: [PATCH] kernel/sys: Optimize do_prlimit lock scope to reduce contention

[Oleg Nesterov]

On 11/27, Andrew Morton wrote:
>
> On Wed, 20 Nov 2024 21:21:56 +0800 Zhen Ni <zhen.ni@easystack.cn> wrote:
>
> > The security_task_setrlimit function is a Linux Security Module (LSM)
> > hook that evaluates resource limit changes based on security policies.
> > It does not alter the rlim data structure, as confirmed by existing
> > LSM implementations (e.g., SELinux and AppArmor). Thus, this function
> > does not require locking, ensuring correctness while improving
> > concurrency.
>
> Seems sane.
>
> Does any code call do_prlimit() frequently enough for this to matter?

I have the same question...

> > -	task_lock(tsk->group_leader);
> >  	if (new_rlim) {
> >  		/*
> >  		 * Keep the capable check against init_user_ns until cgroups can
> >  		 * contain all limits.
> >  		 */
> >  		if (new_rlim->rlim_max > rlim->rlim_max &&
> > -				!capable(CAP_SYS_RESOURCE))
> > -			retval = -EPERM;
> > -		if (!retval)
> > -			retval = security_task_setrlimit(tsk, resource, new_rlim);
> > +		    !capable(CAP_SYS_RESOURCE))
> > +			return -EPERM;
> > +		retval = security_task_setrlimit(tsk, resource, new_rlim);
> > +		if (retval)
> > +			return retval;
> >  	}
> > +
> > +	task_lock(tsk->group_leader);

The problem is that task_lock(tsk->group_leader) doesn't look right with or
without this patch. I'll try to make a fix on weekend.

If the caller is sys_prlimit64() and tsk != current, then ->group_leader is
not stable, do_prlimit() can race with mt exec and take the wrong lock.

Oleg.

Re: [PATCH] kernel/sys: Optimize do_prlimit lock scope to reduce contention

[Oleg Nesterov]

On 11/28, Oleg Nesterov wrote:
>
> The problem is that task_lock(tsk->group_leader) doesn't look right with or
> without this patch. I'll try to make a fix on weekend.
>
> If the caller is sys_prlimit64() and tsk != current, then ->group_leader is
> not stable, do_prlimit() can race with mt exec and take the wrong lock.

... and task_unlock(tsk->group_leader) is simply unsafe.

perhaps something like below, but it doesn't look nice, I'll try to think
more. And grep, may be there are more lockless users of tsk->group_leader
when tsk != current.

Oleg.

--- a/kernel/sys.c
+++ b/kernel/sys.c
@@ -1464,6 +1464,7 @@ SYSCALL_DEFINE2(setdomainname, char __user *, name, int, len)
 static int do_prlimit(struct task_struct *tsk, unsigned int resource,
 		      struct rlimit *new_rlim, struct rlimit *old_rlim)
 {
+	struct task_struct *leader;
 	struct rlimit *rlim;
 	int retval = 0;
 
@@ -1481,7 +1482,14 @@ static int do_prlimit(struct task_struct *tsk, unsigned int resource,
 
 	/* Holding a refcount on tsk protects tsk->signal from disappearing. */
 	rlim = tsk->signal->rlim + resource;
-	task_lock(tsk->group_leader);
+
+	if (tsk != current)
+		read_lock(&tasklist_lock);
+	leader = READ_ONCE(tsk->group_leader);
+	task_lock(leader);
+	if (tsk != current)
+		read_unlock(&tasklist_lock);
+
 	if (new_rlim) {
 		/*
 		 * Keep the capable check against init_user_ns until cgroups can
@@ -1499,7 +1507,7 @@ static int do_prlimit(struct task_struct *tsk, unsigned int resource,
 		if (new_rlim)
 			*rlim = *new_rlim;
 	}
-	task_unlock(tsk->group_leader);
+	task_unlock(leader);
 
 	/*
 	 * RLIMIT_CPU handling. Arm the posix CPU timer if the limit is not

Re: [PATCH] kernel/sys: Optimize do_prlimit lock scope to reduce contention

[Oleg Nesterov]

On 11/28, Oleg Nesterov wrote:
>
> On 11/28, Oleg Nesterov wrote:
> >
> > The problem is that task_lock(tsk->group_leader) doesn't look right with or
> > without this patch. I'll try to make a fix on weekend.
> >
> > If the caller is sys_prlimit64() and tsk != current, then ->group_leader is
> > not stable, do_prlimit() can race with mt exec and take the wrong lock.
>
> ... and task_unlock(tsk->group_leader) is simply unsafe.
>
> perhaps something like below,

No, this is wrong too,

> I'll try to think more.

Yes...

Oleg.