On Wed, May 19, 2010 at 12:27:43PM +0200, Oleg Nesterov wrote: > On 05/19, Peter Zijlstra wrote: > > > > On Tue, 2010-05-18 at 23:22 +0200, Oleg Nesterov wrote: > > > > > And, looking at this patch I think that schedule() can be simplified > > > a little bit. > > > > > > "sched: Reassign prev and switch_count when reacquire_kernel_lock() fail" > > > commit 6d558c3ac9b6508d26fd5cadccce51fc9d726b1c says: > > > > > > Assume A->B schedule is processing, > > > ... > > > Then on B's context, > > > ... > > > prev and switch_count are related to A > > > > > > How so? switch_count - yes, we should change it. But prev must be > > > equal to B, and it must be equal to current. When we return from > > > switch_to() registers/stack should be restored correctly, so we > > > can do > > > > What if schedule() got called at a different stack depth than we are > > now? > > > > I don't think we can assume anything about the stack context we just > > switched to. > > Not sure I understand... > > OK. Firstly, we shouldn't worry about the freshly forked tasks, they > never "return" from switch_to() but call ret_from_fork()->schedule_tail(), > right? > > Now suppose that A calls schedule() and we switch to B. When switch_to() > returns on B's context, this context (register/stack) matches the previous > context which was used by B when it in turn called schedule(), correct? > > IOW. B calls schedule, prev == B. schedule() picks another task, prev > is saved on B's stck after switch_to(). A calls schedule(), prev == A > before context_switch(A, B), but after that switch_to() switches to > B's stack and prev == B. > > No? I think you are right. > > > I am looking into the git history now... and I guess I understand why > reacquire_kernel_lock() uses current. Because schedule() did something > like > > prev = context_switch(prev, next); // prev == last > > finish_task_switch(prev); > > reacquire_kernel_lock(current); // prev != current This is what I think when I wrote that patch. Now the task switch is entirely finished in context_switch(). So commit log in 6d558c3a has some flaw in it. The "prev" is also a churn. Thanks, Yong -- To unsubscribe from this list: send the line "unsubscribe linux-arch" in the body of a message to majordomo@xxxxxxxxxxxxxxx More majordomo info at http://vger.kernel.org/majordomo-info.html