On Mon, Aug 19, 2013 at 11:14 PM, Greg Kroah-Hartman
<gre...@linuxfoundation.org> wrote:
> On Mon, Aug 19, 2013 at 11:06:31PM +0800, Ming Lei wrote:
>> On Mon, Aug 19, 2013 at 10:00 PM, Greg Kroah-Hartman
>> <gre...@linuxfoundation.org> wrote:
>> > On Mon, Aug 19, 2013 at 07:04:18PM +0800, Ming Lei wrote:
>> >> Because usb_hcd_submit_urb is in the hotest path of usb core,
>> >> so use percpu counter to count URB instead of using atomic variable
>> >> because atomic operations are much slower than percpu operations.
>> >>
>> >> Cc: Oliver Neukum <oli...@neukum.org>
>> >> Cc: Alan Stern <st...@rowland.harvard.edu>
>> >> Signed-off-by: Ming Lei <ming....@canonical.com>
>> >> ---
>> >> drivers/usb/core/hcd.c | 4 ++--
>> >> drivers/usb/core/sysfs.c | 7 ++++++-
>> >> drivers/usb/core/usb.c | 9 ++++++++-
>> >> drivers/usb/core/usb.h | 1 +
>> >> include/linux/usb.h | 2 +-
>> >> 5 files changed, 18 insertions(+), 5 deletions(-)
>> >
>> > And this really speeds things up? Exactly what does it?
>> >
>> > And it's not that atomic operations are "slower", it's just that the
>>
>> For SMP, atomic_inc/atomic_dec are much slower than percpu
>> variable inc/dec, see 4.1(Why Isn’t Concurrent Count-ing Trivial?)
>> of [1].
>>
>> However, it is slower: on a Intel Core Duo laptop, it is about six
>> times slower than non-atomic increment when a single thread
>> is incrementing, and more than ten times slower if two threads
>> are incrementing.
>>
>> Considered that most of desktop & laptop are SMP now, and with
>> USB3.0, the submitted URBs per second may reach tens of thousand
>> or more, and we can remove the atomic inc/dec operations in the hot
>> path, so why don't do it?
>
> Because you really didn't do it, there are lots of other atomic
> operations on that same path.
Not lots in the path of usbcore.
>
> And, thens of thousands of urbs should be trivial, did you measure this
> to see if it changed anything? I'm not taking patches like this that
> are not quantifiable, sorry.
The number may be too trivial to measure, but I will try to test
with perf.
>
> The gating problem in USB right now is the hardware, it's the slowest
> thing, not the kernel, from everything I have ever tested, or seen.
The problem may not speed up usb performance, but might decrease
CPU utilization a bit, or cache miss.
>
> Well, bad host controller silicon is also a problem (i.e. raspberry pi),
> but there's not much we can do about braindead problems like that...
>
>> > barriers involved can be slower, depending on what else is happening.
>> > If you look, you are already hitting atomic variables in the same path,
>> > so how can this change speed anything up?
>>
>> No, no barriers are involved in atomic_inc/atomic_dec at all.
>
> None? Hm, you might want to rethink that statement :)
Please see Documentation/memory-barriers.txt:
The following also do _not_ imply memory barriers, and so may require explicit
memory barriers under some circumstances (smp_mb__before_atomic_dec() for
instance):
atomic_add();
atomic_sub();
atomic_inc();
atomic_dec();
Thanks,
--
Ming Lei
--
To unsubscribe from this list: send the line "unsubscribe linux-usb" in
the body of a message to majord...@vger.kernel.org
More majordomo info at http://vger.kernel.org/majordomo-info.html
