Hi -
Forwarded from ananth, whose emails are not hitting utrace-devel for
some reason:
Date: Wed, 22 Jul 2009 13:21:43 +0530
From: Ananth N Mavinakayanahalli <ana...@in.ibm.com>
To: utrace-devel@redhat.com
Subject: Q: attaching engines to all threads
Roland, Oleg,
What do you think is the best model to attach engines to all threads of
a process safely using utrace? Uprobes implements it in a slightly
convoluted fashion, while I wrote (assuming 'task' is a valid
task_struct, corresponding to the tgid of the multi-threaded process):
+ rcu_read_lock();
+ do {
+ get_task_struct(t);
+ rcu_read_unlock();
+ ret = <attach_your_engine_and_set_events>(t);
+ if (ret && (ret != EINPROGRESS))
+ goto err;
+ rcu_read_lock();
+ put_task_struct(t);
+ ret = 0;
+ t = next_thread(t);
+ } while (t != task);
+
+ rcu_read_unlock();
But this is still not 'safe' as there are still races/possibility of t
dying within the unlock/lock in the loop. And its not possible to hold
the rcu_read_lock all the way since the routine that attaches the engine
will need to do a kzalloc and the like for housekeeping.
What is your preferred approach for this?
Ananth
PS: I believe getting a feature/callback implemented in utrace so not every
user of the infrastructure needs to reinvent the wheel is a good idea.
In fact, it may be a better thing to have simply to avoid buggy
implementations.
