Re: [PATCH] proc: Fix proc_tid_readdir so it handles the weird cases.

[ebiederm@xmission.com (Eric W. Biederman)]

Louis Rilling <Louis.Rilling@kerlabs.com> writes:

> On 18/03/09  6:39 -0700, Eric W. Biederman wrote:
>> 
>> Fix the logic in proc_task_readdir so that we provide the
>> guarantee that in the sequence:
>> opendir
>> readdir
>> readdir
>> ....
>> readdir
>> closedir
>> If a thread exists from the beginning until the end we are
>> guaranteed to see it, even if some threads are created or
>> worse are destroyed during the readdir.
>> 
>> This guarantee is provided by reusing the same logic used
>> by proc_pid_readdir for the root directory of /proc.  The
>> pid is encoded into the offset, and we walk the pid bitmap
>> of all pids in the system, and test each of them to see if
>> it belongs to the specified thread group.
>> 
>> If we seek to a bad location or if the task we say last
>> exits it is ok because we can numerically find the next
>> task we would have returned.
>> 
>> This is slower that the previous version, but it should
>> not be too slow as this techinique is already use for
>> the root directory of /proc without problems.
>
> This makes 'ps -T x' an O(n^2) thing, compared to O(n) before the patch, right?
> It would be good to, at least, check for thread_group_empty().

Sort of.  It is more like it adds a constant to the traversing each thread
directory.  With everything in cache I don't expect it to be a big constant,
as I have reports that when I changed the pid directory it sped things up.

We definitely don't hold any locks so only the process doing the traversal will pay
the price.

Since this is actually a correctness issue I'm willing to pay a little more to get
the right result.

Eric