2008-07-28 22:59:57

by David Howells

[permalink] [raw]
Subject: Request for documentation of PID handling functions


Hi Pavel,

Can you please provide documentation for the PID handling functions you
supplied in commit:

7af5729474b5b8ad385adadab78d6e723e7655a3

In particular I'd like to know what locks are required when. I'm trying to
determine if in this piece of code:

read_lock(&tasklist_lock);
ret = 0;
if (pid != task_pid_vnr(current))
ret = -EPERM;
read_unlock(&tasklist_lock);

the lock of tasklist_lock is necessary.

David


2008-07-30 02:04:25

by Pavel Emelyanov

[permalink] [raw]
Subject: Re: Request for documentation of PID handling functions

David Howells wrote:
> Hi Pavel,
>
> Can you please provide documentation for the PID handling functions you
> supplied in commit:
>
> 7af5729474b5b8ad385adadab78d6e723e7655a3
>
> In particular I'd like to know what locks are required when. I'm trying to
> determine if in this piece of code:
>
> read_lock(&tasklist_lock);
> ret = 0;
> if (pid != task_pid_vnr(current))
> ret = -EPERM;
> read_unlock(&tasklist_lock);
>
> the lock of tasklist_lock is necessary.

The RCU protection is enough to get pids from tasks. However, the pids
are detached under the tasklist lock write-locked, so provided you're
*already* sitting under it the RCU is excessive.

However, Oleg pointed me out some time ago, that in some cases RCU-only
protection was not enough in some cases, so I put him in Cc.

> David
>

2008-07-30 09:52:28

by Oleg Nesterov

[permalink] [raw]
Subject: Re: Request for documentation of PID handling functions

On 07/30, Pavel Emelyanov wrote:
>
> David Howells wrote:
> > Hi Pavel,
> >
> > Can you please provide documentation for the PID handling functions you
> > supplied in commit:
> >
> > 7af5729474b5b8ad385adadab78d6e723e7655a3
> >
> > In particular I'd like to know what locks are required when. I'm trying to
> > determine if in this piece of code:
> >
> > read_lock(&tasklist_lock);
> > ret = 0;
> > if (pid != task_pid_vnr(current))
> > ret = -EPERM;
> > read_unlock(&tasklist_lock);
> >
> > the lock of tasklist_lock is necessary.
>
> The RCU protection is enough to get pids from tasks. However, the pids
> are detached under the tasklist lock write-locked, so provided you're
> *already* sitting under it the RCU is excessive.

Yes, tasklist OR rcu is enough.

But, since the task == current, the code above doesn't need any locks at all.

Unless this code runs after exit_notify() of course, in that case we need
rcu/tasklist and task_pid_vnr(current) can return 0.

Oleg.