Re: [PATCH 2.6.34] ehci-hcd: add option to enable 64-bit DMA support

[Robert Hancock]

On Thu, Feb 18, 2010 at 6:47 PM, Greg KH <greg@xxxxxxxxx> wrote:
>> > So you did not measure it?
>> >
>> > Hm, I guess this change must not be necessary :)
>>
>> I'll try and run some tests and see what I can quantify. However, I
>> only have 4GB of RAM on my machine (with a 1GB memory hole) and so a
>> random memory allocation only has a 25% chance of ending up in the
>> area where it would make a difference, so it may take a bit of doing.
>
> Without any good justification, including real tests being run, I can't
> take this patch, the risk is just too high.

Again, this particular patch has essentially zero risk for anyone that
doesn't choose to experiment with the option. One can hardly say it
presents much of a long-term maintenance burden either..

>
> And really, for USB 2.0 speeds, I doubt you are going to even notice
> this kind of overhead, it's in the noise.  Especially given that almost
> always the limiting factor is the device itself, not the host.

Well, I do have some results. This is from running this "dd
if=/dev/sdg of=/dev/null bs=3800M iflag=direct" against an OCZ Rally2
USB flash drive, which gets about 30 MB/sec on read, with CPU-burning
tasks on all cores in the background. (The huge block size and
iflag=direct is to try to force more of the IO to happen to memory
above the 4GB mark.) With that workload, swiotlb_bounce shows up as
between 1.5 to 4% of the CPU time spent in the kernel according to
oprofile. Obviously with the 64-bit DMA enabled, that disappears. Of
course, the overall kernel time is only around 2% of the total time,
so that's a pretty small overall percentage.

I'll try some tests later with a faster SATA-to-IDE device that should
stress things a bit more, but a huge difference doesn't seem likely.
One thing that's uncertain is just how much of the IO is needing to be
bounced - an even distribution of the buffer across all of physical
RAM would suggest 25% in this case, but I don't know an easy way to
verify that.

Aside from speed considerations though, I should point out another
factor: IOMMU/SWIOTLB space is in many cases a limited resource for
all IO in flight at a particular time (SWIOTLB is typically 64MB). The
number of hits when Googling for "Out of IOMMU space" indicates it is
a problem that people do hit from time to time. From that perspective,
anything that prevents unnecessary use of bounce buffers is a good
thing.
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/