Re: [PATCH] drivers/char/mem.c: Add /dev/ioports, supporting 16-bit and 32-bit ports

[Alex Williamson]

On Thu, 2015-12-31 at 15:03 +0530, Santosh Shukla wrote:
> On Tue, Dec 29, 2015 at 11:01 PM, Alex Williamson
> <alex.williamson@xxxxxxxxxx> wrote:
> > On Tue, 2015-12-29 at 22:00 +0530, Santosh Shukla wrote:
> > > On Tue, Dec 29, 2015 at 9:50 PM, Arnd Bergmann <arnd@xxxxxxxx>
> > > wrote:
> > > > On Tuesday 29 December 2015 21:25:15 Santosh Shukla wrote:
> > > > > mistakenly added wrong email-id of alex, looping his correct
> > > > > one.
> > > > > 
> > > > > On 29 December 2015 at 21:23, Santosh Shukla <santosh.shukla@
> > > > > lina
> > > > > ro.org> wrote:
> > > > > > On 29 December 2015 at 18:58, Arnd Bergmann <arnd@xxxxxxxx>
> > > > > > wrote:
> > > > > > > On Wednesday 23 December 2015 17:04:40 Santosh Shukla
> > > > > > > wrote:
> > > > > > > > On 23 December 2015 at 03:26, Arnd Bergmann <arnd@arndb
> > > > > > > > .de>
> > > > > > > > wrote:
> > > > > > > > > On Tuesday 22 December 2015, Santosh Shukla wrote:
> > > > > > > > > > }
> > > > > > > > > > 
> > > > > > > > > > So I care for /dev/ioport types interface who could
> > > > > > > > > > do
> > > > > > > > > > more than byte
> > > > > > > > > > data copy to/from user-space. I tested this patch
> > > > > > > > > > with
> > > > > > > > > > little
> > > > > > > > > > modification and could able to run pmd driver for
> > > > > > > > > > arm/arm64 case.
> > > > > > > > > > 
> > > > > > > > > > Like to know how to address pci_io region mapping
> > > > > > > > > > problem for
> > > > > > > > > > arm/arm64, in-case /dev/ioports approach is not
> > > > > > > > > > acceptable or else I
> > > > > > > > > > can spent time on restructuring the patch?
> > > > > > > > > > 
> > > > > > > > > 
> > > > > > > > > For the use case you describe, can't you use the vfio
> > > > > > > > > framework to
> > > > > > > > > access the PCI BARs?
> > > > > > > > > 
> > > > > > > > 
> > > > > > > > I looked at file: drivers/vfio/pci/vfio_pci.c, func
> > > > > > > > vfio_pci_map() and
> > > > > > > > it look to me that it only maps ioresource_mem pci
> > > > > > > > region,
> > > > > > > > pasting
> > > > > > > > code snap:
> > > > > > > > 
> > > > > > > > if (!(pci_resource_flags(pdev, index) &
> > > > > > > > IORESOURCE_MEM))
> > > > > > > > return -EINVAL;
> > > > > > > > ....
> > > > > > > > 
> > > > > > > > and I want to map ioresource_io pci region for arm
> > > > > > > > platform
> > > > > > > > in my
> > > > > > > > use-case. Not sure vfio maps pci_iobar region?
> > > > > > > 
> > > > > > > Mapping I/O BARs is not portable, notably it doesn't work
> > > > > > > on
> > > > > > > x86.
> > > > > > > 
> > > > > > > You should be able access them using the read/write
> > > > > > > interface
> > > > > > > on
> > > > > > > the vfio device.
> > > > > > > 
> > > > > > Right, x86 doesn't care as iopl() could give userspace
> > > > > > application
> > > > > > direct access to ioports.
> > > > > > 
> > > > > > Also, Alex in other dpdk thread [1] suggested someone to
> > > > > > propose io
> > > > > > bar mapping in vfio-pci, I guess in particular to non-x86
> > > > > > arch
> > > > > > so I
> > > > > > started working on it.
> > > > > > 
> > > > > 
> > > > 
> > > > So what's wrong with just using the existing read/write API on
> > > > all
> > > > architectures?
> > > > 
> > > 
> > > nothing wrong, infact read/write api will still be used so to
> > > access
> > > mmaped io pci bar at userspace. But right now vfio_pci_map()
> > > doesn't
> > 
> > vfio_pci_mmap(), the read/write accessors fully support i/o port.
> > 
> 
> (Sorry for delayed response!)
> Right.
> > > map io pci bar in particular (i.e.. ioresource_io) so I guess
> > > need to
> > > add that bar mapping in vfio. pl. correct me if i misunderstood
> > > anything.
> > 
> > Maybe I misunderstood what you were asking for, it seemed like you
> > specifically wanted to be able to mmap i/o port space, which is
> > possible, just not something we can do on x86.ÂÂMaybe I should have
> > asked why.ÂÂThe vfio API already supports read/write access to i/o
> > port
> 
> Yes, I want to map io port pci space in vfio and reason for that is :
> I want to access virto-net-pci device at userspace using vfio and for
> that I am using vfio-noiommu latest linux-next patch. but I am not
> able to mmap io port pci space in vfio because of below condition -
> 
> 1)
> --- user space code snippet ----
> reg.index = i; // where i is {0..1} i.e.. {BAR0..BAR1} such that BAR0
> = io port pci space and BAR1 = pci config space
> 
> ret = ioctl(vfio_dev_fd, VFIO_DEVICE_GET_REGION_INFO, &reg);
> if ((reg.flags & VFIO_REGION_INFO_FLAG_MMAP) == 0) {
> ÂÂÂÂÂÂÂreturn err;
> }
> now consider i = 0 case where pci_rersource_flag set to IORESOURCE_IO
> 
> --- kernel / vfip-pci.c -------------
> so vfio_pci_ioctl() wont set info.flag to VFIO_REGION_INFO_FLAG_MMAP.
> And it won't set for two
> 1) pci_resource_flag & IORESOURCE_MEM
> 2) ioport size < PAZE_SIZE
> 
> The second one I addressed but first one is what I believe that need
> to add support in vfio.
> and Same applicable for vfio_pci_mmap() too..
> 
> This is why I am thinking to add IORESOURCE_IO space mapping support
> in vfio; in particular non-x86 archs.. pl. correct my understanding
> in
> case wrong.
> 
> > space, so if you intend to mmap it only to use read/write on top of
> > the
> > mmap, I suppose you might see some performance improvement, but not
> > really any new functionality.ÂÂYou'd also need to deal with page
> > size
> > issues since i/o port ranges are generally quite a bit smaller than
> > the
> > host page size and they'd need to be mapped such that each devices
> > does
> > not share a host page of i/o port space with other devices.ÂÂOn x86
> > i/o
> 
> Yes. I have taken care size < PAZE_SIZE condition.
> 
> > port space is mostly considered legacy and not a performance
> > critical
> > path for most modern devices; PCI SR-IOV specifically excludes i/o
> > port
> > space.ÂÂSo what performance gains do you expect to see in being
> > able to
> > mmap i/o port space and what hardware are you dealing with that
> > relies
> > on i/o port space rather than mmio for performance?ÂÂThanks,
> > 
> dpdk user space virtio-net pmd driver uses ioport space for driver
> initialization, as because virtio-net header resides in ioport area
> of
> virtio-pxe.rom file, also it is inlined to virtio spec (<= 0.95).
> Till
> now virtio-net dpdk pmd driver for x86 using iopl() to access those
> ioport for driver initialization but for non-x86 cases; we needed
> alternative i.e.. kernel to someway map ioport pci region either by
> architecture example powerpc does Or look in vfio for mapping. I hope
> I made my use-case clear.

Not really. ÂI still don't understand why you need to *mmap* ioport
space rather than access it via read/write. Âvfio already supports
assignment of numerous physical devices that rely on ioport space for
the device rom, device initialization, and even runtime operation in
QEMU using the accesses currently supported. ÂLegacy x86 ioport space
cannot be mmap'd on x86 hardware, it's only through sparse memory
mapping and emulation of ioport space provided on some architectures
that this is even possible, so you will not achieve
platform/architecture neutral support for mmap'ing ioport space, which
means that your userspace driver will not work universally if it
depends on this support.

If you were using iopl() and in*()/out*() before, simply drop the
iopl() and use pread()/pwrite() instead. ÂThanks,

AlexÂ
--
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/