Messages in this thread | | | Date | Mon, 17 Mar 2008 20:22:28 +0100 | From | Eric Dumazet <> | Subject | Re: per cpun+ spin locks coexistence? |
| |
Peter Teoh a écrit : > Thanks for the explanation, much apologies for this newbie discussion. > But I still find it inexplicable: > > On Mon, Mar 17, 2008 at 4:20 AM, Johannes Weiner <hannes@saeurebad.de> wrote: > >> A per-cpu variable is basically an array the size of the number of >> possible CPUs in the system. get_cpu_var() checks what current CPU we >> are running on and gets the array-element corresponding to this CPU. >> >> So, really oversimplified, get_cpu_var(foo) translates to something like >> foo[smp_processor_id()]. >> >> > > Ok, so calling get_cpu_var() always return the array-element for the > current CPU, and since by design, only the current CPU can > modify/write to this array element (this is my assumption - correct?), > and the other CPU will just read it (using the per_cpu construct). > So far correct? So why do u still need to spin_lock() to lock other > CPU from accessing - the other CPU will always just READ it, so just > go ahead and let them read it. Seemed like it defeats the purpose of > get_cpu_var()'s design? > > But supposed u really want to put a spin_lock(), just to be sure > nobody is even reading it, or modifying it, so then what is the > original purpose of get_cpu_var() - is it not to implement something > that can be parallelized among different CPU, without affecting each > other, and using no locks? > > The dual use of spin_lock+get_cpu_var() confuses me here :-). (not > the per_cpu(), which I agree is supposed to be callabe from all the > different CPU, for purpose of enumeration or data collection). > > You are right Peter, that fs/file.c contains some leftover from previous implementation of defer queue, that was using a timer.
So we can probably provide a patch that :
- Use spin_lock() & spin_unlock() instead of spin_lock_bh() & spin_unlock_bh() in free_fdtable_work() since we dont anymore use a softirq (timer) to reschedule the workqueue.
( this timer was deleted by the following patch : http://readlist.com/lists/vger.kernel.org/linux-kernel/50/251040.html
But, you cannot avoid use of spin_lock()/spin_unlock() because schedule_work() makes no garantee that the work will be done by this cpu.
(free_fdtable_work() can be called to flush the fd defer queue of CPU X on behalf CPU Y, with X != Y . You then can have a corruption because CPU X is inside free_fdtable_rcu() and CPU Y is inside free_fdtable_work() )
So both spin_lock() and get_cpu_var() are necessary :
- One to get the precpu data for optimal performance on SMP (but not mandatory) - One to protect the data from corruption on SMP.
-- To unsubscribe from this list: send the line "unsubscribe linux-kernel" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html Please read the FAQ at http://www.tux.org/lkml/
| |