lkml.org 
[lkml]   [2014]   [Feb]   [24]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
Date
From
SubjectRe: [PATCH 4/7] pid: modify task_tgid_nr to work without task->tgid.
On 02/21, Richard Guy Briggs wrote:
>
> On 14/02/20, Oleg Nesterov wrote:
> > On 01/23, Richard Guy Briggs wrote:
> > >
> > > task->tgid is an error prone construct and results in duplicate maintenance.
> > > Start it's demise by modifying task_tgid_nr to not use it.
> >
> > Well, I disagree.
> >
> > Yes I agree that ->tgid should probably die. But this change itself doesn't
> > help, it only makes task_tgid_nr() slower. We need to convert other users
> > first, then consider this change along with ->tgid removal.
>
> I thought I recently saw a statistic that there were only 7 instances of
> ->tgid, but a quick search comes up with >50 outside of audit!

Yes, so I don't think it makes sense to start from task_tgid_nr().

> > > - return tsk->tgid;
> > > + return pid_nr(task_tgid(tsk));
> > > }
> >
> > And what protect task_tgid? This is racy.
> >
> > The race is very unlikely, pid_nr() will likely hit pid == NULL if tsk
> > exits. But still it can use the freed/unmapped/reused memory.
>
> So at the very least, I'd need
>
> static inline pid_t task_tgid_nr(struct task_struct *tsk)
> {
> return pid_nr(is_alive(tsk) ? task_tgid(tsk) : NULL);
> }
>
> And it sounds like I might even need this since the status of is_alive
> could change between the time I call it and the time I call task_tgid:
>
> static inline pid_t task_tgid_nr(struct task_struct *tsk)
> {
> pid_t pid;
>
> task_lock(&tsk);
> pid = pid_nr(is_alive(tsk) ? task_tgid(tsk) : NULL);
> task_unlock(&tsk);
> return pid;
> }
>
> Is task_lock() sufficient, or do I need the heavier
> read_lock(&tasklist_lock)?

You need rcu_read_lock(), I think. In any case task_lock() has nothing to
do with pids.

> > And even if we add rcu_read_lock() the patch will add the semantics change,
> > task_tgid_nr() can return 0 if tsk has already exited. At least this should
> > be documented, but you also need to audit the users.
>
> Basically check for an inline error return of 0 signalling a failure
> rather than the idle task.

Perhaps...

Oleg.



\
 
 \ /
  Last update: 2014-02-24 20:21    [W:0.097 / U:0.320 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site