lkml.org 
[lkml]   [2006]   [Nov]   [20]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
Date
From
SubjectRe: [patch] cpufreq: mark cpufreq_tsc() as core_initcall_sync
On Mon, Nov 20, 2006 at 06:55:54PM +0100, Jens Axboe wrote:
> On Mon, Nov 20 2006, Paul E. McKenney wrote:
> > On Mon, Nov 20, 2006 at 08:15:14AM +0100, Jens Axboe wrote:
> > > On Sun, Nov 19 2006, Paul E. McKenney wrote:
> > > > On Sat, Nov 18, 2006 at 09:46:24PM +0300, Oleg Nesterov wrote:
> > > > > On 11/17, Paul E. McKenney wrote:
> > > > > >
> > > > > > Oleg, any thoughts about Jens's optimization? He would code something
> > > > > > like:
> > > > > >
> > > > > > if (srcu_readers_active(&my_srcu))
> > > > > > synchronize_srcu();
> > > > > > else
> > > > > > smp_mb();
> > > > >
> > > > > Well, this is clearly racy, no? I am not sure, but may be we can do
> > > > >
> > > > > smp_mb();
> > > > > if (srcu_readers_active(&my_srcu))
> > > > > synchronize_srcu();
> > > > >
> > > > > in this case we also need to add 'smp_mb()' into srcu_read_lock() after
> > > > > 'atomic_inc(&sp->hardluckref)'.
> > > > >
> > > > > > However, he is doing ordered I/O requests rather than protecting data
> > > > > > structures.
> > > > >
> > > > > Probably this makes a difference, but I don't understand this.
> > > >
> > > > OK, one hypothesis here...
> > > >
> > > > The I/Os must be somehow explicitly ordered to qualify
> > > > for I/O-barrier separation. If two independent processes
> > > > issue I/Os concurrently with a third process doing an
> > > > I/O barrier, the I/O barrier is free to separate the
> > > > two concurrent I/Os or not, on its whim.
> > > >
> > > > Jens, is the above correct? If so, what would the two processes
> > >
> > > That's completely correct, hence my somewhat relaxed approach with SRCU.
> >
> > OK, less scary in that case. ;-)
>
> Yep, it's really not scary in any ordering sense!
>
> > > > need to do in order to ensure that their I/O was considered to be
> > > > ordered with respect to the I/O barrier? Here are some possibilities:
> > >
> > > If we consider the barrier a barrier in a certain stream of requests,
> > > it is the responsibility of the issuer of that barrier to ensure that
> > > the queueing is ordered. So if two "unrelated" streams of requests with
> > > barriers hit __make_request() at the same time, we don't go to great
> > > lengths to ensure who gets there firt.
> >
> > So the "preceding" requests have to have completed their I/O system
> > calls? If this is the case, does this include normal (non-direct/raw)
> > writes and asynchronous reads? My guess is that it would include
> > asynchronous I/O, but not buffered writes.
>
> They need not have completed, but they must have been queued at the
> block layer level. IOW, the io scheduler must know about them. Since
> it's a block layer device property, we really don't care about system
> calls since any of them could amount to 1 or lots more individual io
> requests.
>
> But now we have taken a detour from the original problem. As I wrote
> above, the io scheduler must know about the requests. When the plug list
> ends up in the private process context, the io scheduler doesn't know
> about it yet. When a barrier is queued, the block layer does not care
> about io that hasn't been issued yet (dirty data in the page cache
> perhaps), since if it hasn't been seen, it's by definition not
> interesting. But if some of the requests reside in a different process
> private request list, then that is a violation of this rule since it
> should technically belong to the block layer / io scheduler at that
> point. This is where I wanted to use SRCU.

OK. Beyond a certain point, I would need to see the code using SRCU.

> > > > 1. I/O barriers apply only to preceding and following I/Os from
> > > > the process issuing the I/O barrier.
> > > >
> > > > 2. As for #1 above, but restricted to task rather than process.
> > > >
> > > > 3. I/O system calls that have completed are ordered by the
> > > > barrier to precede I/O system calls that have not yet
> > > > started, but I/O system calls still in flight could legally
> > > > land on either side of the concurrently executing I/O
> > > > barrier.
> > > >
> > > > 4. Something else entirely?
> > > >
> > > > Given some restriction like one of the above, it is entirely possible
> > > > that we don't even need the memory barrier...
> > >
> > > 3 is the closest. The request queue doesn't really know the scope of the
> > > barrier, it has to rely on the issuer getting it right. If you have two
> > > competing processes issuing io and process A relies on process B issuing
> > > a barrier, they have to synchronize that between them. Normally that is
> > > not a problem, since that's how the file systems always did io before
> > > barriers on items that need to be on disk (it was a serialization point
> > > anyway, it's just a stronger one now).
> >
> > So something like a user-level mutex or atomic instructions must be used
> > by the tasks doing the pre-barrier I/Os to announce that these I/Os have
> > been started in the kernel.
>
> We don't do barriers from user space, it's purely a feature available to
> file systems to ensure ordering of writes even at the disk platter
> level.

Got it. The question then becomes whether the barriers involved in
locking/whatever for the queues enter into the picture.

> > > That said, I think the
> > >
> > > smp_mb();
> > > if (srcu_readers_active(sp))
> > > synchronize_srcu();
> > >
> > > makes the most sense.
> >
> > If the user-level tasks/threads/processes must explicitly synchronize,
> > and if the pre-barrier I/O-initation syscalls have to have completed,
> > then I am not sure that the smp_mb() is needed. Seems like the queuing
> > mechanisms in the syscall and the user-level synchronization would have
> > supplied the needed memory barriers. Or are you using some extremely
> > lightweight user-level synchronization?
>
> Once a process holds a queue plug, any write issued to that plug list
> will do an srcu_read_lock(). So as far as I can tell, the smp_mb() is
> needed to ensure that an immediately following synchronize_srcu() from a
> barrier write queued on a different CPU will see that srcu_read_lock().

Is the srcu_read_lock() invoked before actually queueing the I/O?
Is there any interaction with the queues befory calling synchronize_srcu()?
If yes to both, it might be possible to remove the memory barrier.

That said, the overhead of the smp_mb() is so small compared to that of
the I/O path that it might not be worth it.

> There are no syscall or user-level synchronization.

Got it, thank you for the explanation!

Thanx, Paul
-
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/

\
 
 \ /
  Last update: 2006-11-20 21:11    [from the cache]
©2003-2011 Jasper Spaans. Advertise on this site