Re: ãQuestionãWhether it's legal toenable same physical DMA memory mapped for different NIC device?

From: James Bottomley
Date: Fri Jan 06 2012 - 11:00:44 EST


On Fri, 2012-01-06 at 10:05 +0800, Yanfei Wang wrote:
> On Fri, Jan 6, 2012 at 12:20 AM, James Bottomley
> <James.Bottomley@xxxxxxxxxxxxxxxxxxxxx> wrote:
> > On Thu, 2012-01-05 at 20:40 +0800, Yanfei Wang wrote:
> >> On Wed, Jan 4, 2012 at 11:59 PM, James Bottomley
> >> <James.Bottomley@xxxxxxxxxxxxxxxxxxxxx> wrote:
> >> > On Wed, 2012-01-04 at 10:44 +0800, Yanfei Wang wrote:
> >> >> On Wed, Jan 4, 2012 at 4:33 AM, Konrad Rzeszutek Wilk
> >> >> <konrad.wilk@xxxxxxxxxx> wrote:
> >> >> > On Wed, Dec 07, 2011 at 10:16:40PM +0800, ustc.mail wrote:
> >> >> >> Dear all,
> >> >> >>
> >> >> >> In NIC driver, to eliminate the overhead of dma_map_single() for DMA
> >> >> >> packet data, we have statically allocated huge DMA memory buffer ring
> >> >> >> at once instead of calling dma_map_single() per packet. Considering
> >> >> >> to further reduce the copy overhead between different NIC(port) ring
> >> >> >> while forwarding, one packet from a input NIC(port) will be
> >> >> >> transferred to output NIC(port) with no any copy action.
> >> >> >>
> >> >> >> To satisfy this requirement, the packet memory should be mapped into
> >> >> >> input port and unmapped when leaving input port, then mapped into
> >> >> >> output port and unmapped later.
> >> >> >>
> >> >> >> Whether it's legal to map the same DMA memory into input and output
> >> >> >> port simultaneously? If it's not, then the zero-copy for packet
> >> >> >> forwarding is not feasible?
> >> >> >>
> >> >> >
> >> >> > Did you ever a get a response about this?
> >> >> No.
> >> >
> >> > This is probably because no-one really understands what you're asking.
> >> > As far as mapping memory to PCI devices goes, it's the job of the bridge
> >> > (or the iommu which may or may not be part of the bridge). A standard
> >> > iommu tends not to care about devices and functions, so a range once
> >> > mapped is available to everything behind the bridge. A more secure
> >> > virtualisation based iommu (like the on in VT-D) does, and tends to map
> >> > ranges per device. I know of none that map per device and function, but
> >> > maybe there are.
> >> >
> >> > Your question reads like you have a range of memory mapped to a PCI
> >> > device that you want to use for two different purposes, can you do this?
> >> > to which the answer is that a standard PCI bridge really doesn't care
> >> > and it all depends on the mechanics of the actual device. The only
> >> > wrinkle might be if the two different purposes are on two separate PCI
> >> > functions of the device and the iommu does care.
> >> >
> >> >> >
> >> >> > Is the output/input port on a seperate device function? Or is it
> >> >> > just a specific MMIO BAR in your PCI device?
> >> >> >
> >> >> Platform: x86, intel nehalem 8Core NUMA, linux 2.6.39, 10G
> >> >> 82599NIC(two ports per NIC card);
> >> >> Function: Forwarding packets between different ports.
> >> >> Targets: Forwarding packets with Zero-Overhead, despite other obstacles.
> >> Besides hardware and OS presented above, more detailed descriptions as follows,
> >>
> >> When IXGBE driver do initialization, DMA Descriptors Ring Buffers are
> >> allocated statically and mapped as cache coherent. Instead of
> >> dynamically allocating skb buffers for packet data, to reduce the huge
> >> overhead from skb memory allocation, huge Packet data buffers are
> >> pre-allocated and mapped when driver is loaded. The same strategy is
> >> done for RX end and TX end.
> >> For simple packet forwarding application, one packet from RX should be
> >> replicated from kernel space to userspace, then copied TX end. Here,
> >> One packet at least, should be copied twice to accomplish forwarding.
> >> When doing high performance network application, the copy action want
> >> to be reduced. If Zero-copy can be done, that's better. (May be you
> >> will find that Zero-copy will bring other obstacles, such as memory
> >> management overhead with high performance. We do not care about it
> >> temporally.)
> >> To achieve this goal, a alternative approach is that, unmapping the
> >> packets buffer after receiving it from A device, then mapping this
> >> packet buffer to B device. We hope to reduce the two mapping
> >> operation, so one packet DMA buffer should be mapped to A device(NIC
> >> port) as well as B device simultaneously.
> >> Q: Can this come to ture? Is it legal for mmaping operation at this platform?
> >
> > But you still haven't answered the question upon which all this depends.
> > Let me make it simple. In PCI terms are A and B
> >
> > I. Same Device, Same Function
> > II. Same Device, Different Function
> > III. Different Devices
> $lspci
> 09:00.0 Ethernet controller: Intel Corporation 82599EB 10-Gigabit
> Network Connection (rev 01)
> 09:00.1 Ethernet controller: Intel Corporation 82599EB 10-Gigabit
> Network Connection (rev 01)
> Here, only one physical 10G NIC device, which contains two 10G ports.
> So, the information list above shows that one port correspond one
> function.
> If forwarding packet between two ports, then A,B are different functions.
> Otherwise, A,B are different devices if forwarding packets between two
> physical NIC device. Is it right?
> The OS is built on bare-metal hardware, no kvm, no xen, no VT.

OK, so we already gave you your answer: for same device different
function there's unlikely to be a platform problem (even with IOVA
devices, it would be the device causing the problem, not the platform).
For different devices, whether it works is platform dependent. The
problem platforms will mostly be those designed for virtualisation which
use iommus for isolation.

James



--
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/