Re: [dpdk-dev] [PATCH 6/6] virtio: arm/arm64: memory mapped IO support in pmd driver

[Stephen Hemminger <stephen@networkplumber.org>]

On Thu, 10 Dec 2015 00:49:08 +0530
Santosh Shukla <sshukla@mvista.com> wrote:

> On Thu, Dec 10, 2015 at 12:34 AM, Stephen Hemminger
> <stephen@networkplumber.org> wrote:
> > On Thu, 10 Dec 2015 00:29:30 +0530
> > Santosh Shukla <sshukla@mvista.com> wrote:
> >
> >> On Tue, Dec 8, 2015 at 6:23 PM, Santosh Shukla <sshukla@mvista.com> wrote:
> >> >
> >> >
> >> > On Mon, Dec 7, 2015 at 10:38 PM, Stephen Hemminger
> >> > <stephen@networkplumber.org> wrote:
> >> >>
> >> >> On Fri,  4 Dec 2015 23:05:19 +0530
> >> >> Santosh Shukla <sshukla@mvista.com> wrote:
> >> >>
> >> >> >
> >> >> > +#ifdef RTE_EXEC_ENV_LINUXAPP
> >> >> > +/* start address of first pci_iobar slot (user-space virtual-addres) */
> >> >> > +void *ioport_map;
> >> >> > +#if defined(RTE_ARCH_ARM) || defined(RTE_ARCH_ARM64)
> >> >> > +
> >> >> > +#include <sys/mman.h>
> >> >> > +#define DEV_NAME             "/dev/igb_ioport"
> >> >> > +
> >> >> > +/* Keeping pci_ioport_size = 4k.
> >> >> > + * So maximum mmaped pci_iobar supported =
> >> >> > + * (ioport_size/pci_dev->mem_resource[0].len)
> >> >> > + *
> >> >> > + * note: kernel could allow maximum 32 virtio-net-pci interface, that
> >> >> > mean
> >> >> > + * maximum 32 PCI_IOBAR(s) where each PCI_IOBAR_LEN=0x20, so
> >> >> > virtio_map_ioport()
> >> >> > + * func by theory gonna support 4k/0x20 ==> 128 PCI_IOBAR(s), more than
> >> >> > + * max-virtio-net-pci interface.
> >> >> > + */
> >> >> > +#define PAGE_SIZE            4096
> >> >> > +#define PCI_IOPORT_SIZE              PAGE_SIZE
> >> >> > +#define PCI_IOPORT_MAX               128 /* 4k / 0x20 */
> >> >> > +
> >> >> > +int ioport_map_cnt;
> >> >> > +#endif /* ARM, ARM64 */
> >> >> > +#endif /* RTE_EXEC_ENV_LINUXAPP */
> >> >>
> >> >> These variables should be static.
> >> >>
> >> >
> >> > (Sorry for delayed follow, Was travelling..)
> >> > Right,
> >> >
> >> >>
> >> >> Also, it is should be possible to extract the I/O bar stuff in a generic
> >> >> way through sysfs
> >> >> and not depend on a character device. The long term goal for DPDK
> >> >> acceptance is to
> >> >> eliminate (or at least reduce to a minumum) any requirement for special
> >> >> kernel drivers.
> >> >
> >> >
> >> > I agree. Existing implementation does read pci_iobar for start address and
> >> > size, But for non-x86 arch, we need someway to map pci_iobar and thats-why
> >> > thought of adding device file for a purpose, as archs like arm lack iopl()
> >> > privileged io syscall support, However iopl() too quite native driver design
> >> > assumption.
> >> >
> >> > I have few idea in my mind such that - Right now I am updating ioport_mapped
> >> > addr {kernel-virtual-addr-io-memory} to /sys/bus/pci/pci_bus_xxxx/xxx/map
> >> > field, instead of mapping their, I'll try to map to uio's pci interface and
> >> > then use existing pci_map_resource() api to mmap kernel-virtual-io-address
> >> > to user-space-virtual-ioaddr. We'll come back on this.
> >> >
> >>
> >>
> >> Spent sometime digging dpdk's uio/pci source code, Intent was to map
> >> pci ioport region via uio-way. In order to achieve that I tried to
> >> hack the virtio-net-pci pmd driver. Right now in virtio-net-pci case,
> >> It creates two sysfs entry for pci bars: resource0 /1.
> >>
> >> Resource0; is ioport region
> >> Resource1; is iomem region.
> >>
> >> By appending a RTE_PCI_DRV_NEED_MAPPING flag to drv_flag and passing
> >> hw->io_base = resource1 type pci.mem_resource[slot].addr; where slot
> >> =1. Resource1 is IORESOURCE_MEM type so uio/pci driver able to mmap.
> >> That way I could get the valid user-space virtual address. However
> >> this hack did not worked for me because at qemu side: virtio-pxe.rom
> >> has virtio_headers located at ioport pci region and guest driver
> >> writing at iomem region, that's why driver init failed. Note that
> >> default driver doesn't use resource1 memory.
> >>
> >> This made me think that either I had to add dependent code in kernel
> >> or something similar proposed in this patch.
> >> It is because:
> >> - uio driver and dependent user-space pci api's in dpdk mmaps
> >> IORESOURCE_MEM types address only {refer igbuio_setup_bars() and in
> >> particular function pci_parse_sysfs_resource()}.
> >> - That mmap in userspace backed by arch specific api
> >> pci_mmap_page_range() in kernel.
> >> - pci_mmap_page_range() does not support mmaping to IO_RESOURCE_IO type memory.
> >> - Having said that, we need routine or a way to to map pci_iobar
> >> region from kernel virtual-address to user-space virtual address.
> >
> > There a couple of gotcha's with this. It turns out the iomem region
> > is not mappable on some platforms. I think GCE was one of them.
> 
> afaik, In linux kernel if arch supports pci_mmap_page_range then iomem
> region should map, right? I am confused by reading your reply, also I
> am not aware of GCE? which platform is GCE, please suggest.

I think it was Google Compute Environment that reported an memory region
which was huge and not accessible, they have there own vhost.