Skip to contentSkip to conversation

proc: task->signal can't be NULL

PATCH In review Oleg Nesterov 2 participants 5 messages Mar 22, 2010 – Mar 24, 2010

lore
Share & tools
lore
#0 lore raw

cover proc: task->signal can't be NULL

With the recent changes in -mm it is always safe to dereference task->signal. It can't be NULL and it is pinned to task_struct.

fs/proc becomes the only valid user of signal->count which should either die or become "int nr_threads".

Alexey, Eric.

Can't we kill this counter? Afaics, get_nr_threads() doesn't need to be "precise", we probably can estimate the number of threads using signal->live (yes sure, we can't use ->live as nr_threads).

Except: first_tid() uses get_nr_threads() for optimization. Is this optimization really important? Afaics, it only helps in the unlikely case, probably in that case the extra lockless while_each_thread() doesn't hurt.

IOW, how about

	--- a/fs/proc/base.c
	+++ b/fs/proc/base.c
	@@ -3071,11 +3071,6 @@ static struct task_struct *first_tid(str
				goto found;
		}
	 
	-	/* If nr exceeds the number of threads there is nothing todo */
	-	pos = NULL;
	-	if (nr && nr >= get_nr_threads(leader))
	-		goto out;
	-
		/* If we haven't found our starting place yet start
		 * with the leader and walk nr threads forward.
		 */

?

Not that I think it is terribly important to kill this counter, and probably signal->nr_threads can make sense anyway, so far I am just curious.

Oleg.

#1 lore raw
Quoted 35 lines · Oleg Nesterov

With the recent changes in -mm it is always safe to dereference task->signal. It can't be NULL and it is pinned to task_struct.

fs/proc becomes the only valid user of signal->count which should either die or become "int nr_threads".

Alexey, Eric.

Can't we kill this counter? Afaics, get_nr_threads() doesn't need to be "precise", we probably can estimate the number of threads using signal->live (yes sure, we can't use ->live as nr_threads).

Except: first_tid() uses get_nr_threads() for optimization. Is this optimization really important? Afaics, it only helps in the unlikely case, probably in that case the extra lockless while_each_thread() doesn't hurt.

IOW, how about

	--- a/fs/proc/base.c
	+++ b/fs/proc/base.c
	@@ -3071,11 +3071,6 @@ static struct task_struct *first_tid(str
				goto found;
		}
	 
	-	/* If nr exceeds the number of threads there is nothing todo */
	-	pos = NULL;
	-	if (nr && nr >= get_nr_threads(leader))
	-		goto out;
	-
		/* If we haven't found our starting place yet start
		 * with the leader and walk nr threads forward.
		 */
Quoted 5 lines

?

Not that I think it is terribly important to kill this counter, and probably signal->nr_threads can make sense anyway, so far I am just curious.

I think that was just a sanity check since it was easy. I want to say it prevents a DOS attack with user space passing unreasonably large file position but that DOS attack is handled by ensuring we don't walk through the list if threads more than once.

However:
proc_task_getattr uses get_nr_threads to get it's nlink count correct.

Not walking the thread list to get the number of threads seems like an important cpu time saving measure.

Eric

#2 lore raw
Quoted 31 lines · Eric W. Biederman
Quoted 28 lines · Oleg Nesterov

Can't we kill this counter? Afaics, get_nr_threads() doesn't need to be "precise", we probably can estimate the number of threads using signal->live (yes sure, we can't use ->live as nr_threads).

Except: first_tid() uses get_nr_threads() for optimization. Is this optimization really important? Afaics, it only helps in the unlikely case, probably in that case the extra lockless while_each_thread() doesn't hurt.

IOW, how about

	--- a/fs/proc/base.c
	+++ b/fs/proc/base.c
	@@ -3071,11 +3071,6 @@ static struct task_struct *first_tid(str
				goto found;
		}

	-	/* If nr exceeds the number of threads there is nothing todo */
	-	pos = NULL;
	-	if (nr && nr >= get_nr_threads(leader))
	-		goto out;
	-
		/* If we haven't found our starting place yet start
		 * with the leader and walk nr threads forward.
		 */

?

Quoted 8 lines
Quoted 3 lines

Not that I think it is terribly important to kill this counter, and probably signal->nr_threads can make sense anyway, so far I am just curious.

I think that was just a sanity check since it was easy. I want to say it prevents a DOS attack with user space passing unreasonably large file position but that DOS attack is handled by ensuring we don't walk through the list if threads more than once.

If a bad user passes the large f_pos > nr_threads then this check eliminates the unneeded while_each_thread() loop, yes. But it can use f_pos == nr_threads and provoke the same loop?

Or. just do rewinddir() + readdir(big_count). Now we walk through the list and call proc_task_fill_cache() for each entry.

IOW, I don't understand how this check can help from the DOS pov.

Quoted 2 lines

However:
proc_task_getattr uses get_nr_threads to get it's nlink count correct.

Yes. But we don't need the exactly precise number here if we are racing with fork/exit ?

Quoted 2 lines

Not walking the thread list to get the number of threads seems like an important cpu time saving measure.

Not sure I understand... Also, first_tid() could use sig->sigcnt (the reference counter) instead of sig->count. This is not the same, but I think in practice this is fine.

OK. Let's keep this counter as "int nr_thread".

Besides, when I tried to re-implement get_nr_threads() using signal->live I got the really ugly result ;)

Thanks.

Oleg.

#3 lore raw
Quoted 4 lines · Oleg Nesterov
Quoted 2 lines · Eric W. Biederman

Oleg Nesterov <[email protected]> writes:

Quoted 8 lines

If a bad user passes the large f_pos > nr_threads then this check eliminates the unneeded while_each_thread() loop, yes. But it can use f_pos == nr_threads and provoke the same loop?

Or. just do rewinddir() + readdir(big_count). Now we walk through the list and call proc_task_fill_cache() for each entry.

IOW, I don't understand how this check can help from the DOS pov.

It can't. I just want it to be able to ;)

Quoted 6 lines
Quoted 2 lines

However:
proc_task_getattr uses get_nr_threads to get it's nlink count correct.

Yes. But we don't need the exactly precise number here if we are racing with fork/exit ?

Quoted 6 lines
Quoted 2 lines

Not walking the thread list to get the number of threads seems like an important cpu time saving measure.

Not sure I understand... Also, first_tid() could use sig->sigcnt (the reference counter) instead of sig->count. This is not the same, but I think in practice this is fine.

We need a value that can be computed in constant time, and is not correct except when the number of threads is actively changing.

Quoted 4 lines

OK. Let's keep this counter as "int nr_thread".

Besides, when I tried to re-implement get_nr_threads() using signal->live I got the really ugly result ;)

Sounds good.

Eric

#4 lore raw
Quoted 3 lines · Eric W. Biederman

We need a value that can be computed in constant time, and is not correct except when the number of threads is actively changing.

Sure. I was thinking of something like

	int get_nr_threads(struct task_struct *tsk)
	{
		int nr = atomic_read(&task->signal->live);
		int reasonable_min = 1;

		rcu_read_lock();
		if (!thread_group_leader(tsk) && pid_alive(tsk) &&
		    tsk->group_leader->exit_state)
			reasonable_min = 2;
		rcu_read_unlock();

		return max(nr, reasonable_min);
	}

but as I said this doesn't look nice at all.

Quoted 6 lines
Quoted 4 lines

OK. Let's keep this counter as "int nr_thread".

Besides, when I tried to re-implement get_nr_threads() using signal->live I got the really ugly result ;)

Sounds good.

OK, please see the "final" patch I am going to send...

Oleg.