Re: [Qemu-devel] [RFC PATCH] vfio/pci: Add ioeventfd support

[Alex Williamson]

On Tue, 13 Mar 2018 13:38:00 +0100
Auger Eric <address@hidden> wrote:
> On 08/02/18 02:22, Alexey Kardashevskiy wrote:
> > On 08/02/18 01:12, Alex Williamson wrote:  
> >> On Wed, 7 Feb 2018 15:48:26 +1100
> >> Alexey Kardashevskiy <address@hidden> wrote:
> >>> On 07/02/18 15:25, Alex Williamson wrote:  
> >>>> On Wed, 7 Feb 2018 15:09:22 +1100
> >>>> Alexey Kardashevskiy <address@hidden> wrote:    
> >>>>> On 07/02/18 11:08, Alex Williamson wrote:    
> >>>>>> diff --git a/include/uapi/linux/vfio.h b/include/uapi/linux/vfio.h
> >>>>>> index e3301dbd27d4..07966a5f0832 100644
> >>>>>> --- a/include/uapi/linux/vfio.h
> >>>>>> +++ b/include/uapi/linux/vfio.h
> >>>>>> @@ -503,6 +503,30 @@ struct vfio_pci_hot_reset {
> >>>>>>  
> >>>>>>  #define VFIO_DEVICE_PCI_HOT_RESET     _IO(VFIO_TYPE, VFIO_BASE + 13)
> >>>>>>  
> >>>>>> +/**
> >>>>>> + * VFIO_DEVICE_IOEVENTFD - _IOW(VFIO_TYPE, VFIO_BASE + 14,
> >>>>>> + *                              struct vfio_device_ioeventfd)
> >>>>>> + *
> >>>>>> + * Perform a write to the device at the specified device fd offset, 
> >>>>>> with
> >>>>>> + * the specified data and width when the provided eventfd is 
> >>>>>> triggered.
> >>>>>> + *
> >>>>>> + * Return: 0 on success, -errno on failure.
> >>>>>> + */
> >>>>>> +struct vfio_device_ioeventfd {
> >>>>>> +      __u32   argsz;
> >>>>>> +      __u32   flags;
> >>>>>> +#define VFIO_DEVICE_IOEVENTFD_8               (1 << 0) /* 1-byte 
> >>>>>> write */
> >>>>>> +#define VFIO_DEVICE_IOEVENTFD_16      (1 << 1) /* 2-byte write */
> >>>>>> +#define VFIO_DEVICE_IOEVENTFD_32      (1 << 2) /* 4-byte write */
> >>>>>> +#define VFIO_DEVICE_IOEVENTFD_64      (1 << 3) /* 8-byte write */
> >>>>>> +#define VFIO_DEVICE_IOEVENTFD_SIZE_MASK       (0xf)
> >>>>>> +      __u64   offset;                 /* device fd offset of write */
> >>>>>> +      __u64   data;                   /* data to be written */
> >>>>>> +      __s32   fd;                     /* -1 for de-assignment */
> >>>>>> +};
> >>>>>> +
> >>>>>> +#define VFIO_DEVICE_IOEVENTFD         _IO(VFIO_TYPE, VFIO_BASE + 14)  
> >>>>>>     
> >>>>>
> >>>>>
> >>>>> Is this a first ioctl with endianness fixed to little-endian? I'd 
> >>>>> suggest
> >>>>> to comment on that as things like vfio_info_cap_header do use the host
> >>>>> endianness.    
> >>>>
> >>>> Look at our current read and write interface, we call leXX_to_cpu
> >>>> before calling iowriteXX there and I think a user would logically
> >>>> expect to use the same data format here as they would there.    
> >>>
> >>> If the data is "char data[8]" (i.e. bytestream), then it can be expected 
> >>> to
> >>> be device/bus endian (i.e. PCI == little endian), but if it is u64 - then 
> >>> I
> >>> am not so sure really, and this made me look around. It could be "__le64
> >>> data" too.
> >>>  
> >>>> Also note
> >>>> that iowriteXX does a cpu_to_leXX, so are we really defining the
> >>>> interface as little-endian or are we just trying to make ourselves
> >>>> endian neutral and counter that implicit conversion?  Thanks,    
> >>>
> >>> Defining it LE is fine, I just find it a bit confusing when
> >>> vfio_info_cap_header is host endian but vfio_device_ioeventfd is not.  
> >>
> >> But I don't think we are defining the interface as little-endian.
> >> iowriteXX does a cpu_to_leXX byteswap.  Therefore in order to maintain
> >> endian neutrality, if the data does a cpu->le swap on the way out, I
> >> need to do a le->cpu swap on the way in, right?  Please defend the
> >> assertion that we're creating a little-endian interface.  Thanks,  
> > 
> > 
> > vfio_pci_ioctl() passes "endian-neutral" ioeventfd.data to
> > vfio_pci_ioeventfd() which immediately does the leXX_to_cpu() conversion
> > (and uses the result later on in iowriteXX(), which is not VFIO API) so I
> > read it as the ioctl really expects LE.
> > 
> > The QEMU part - vfio_nvidia_mirror_quirk MR - does not swap bytes but the
> > MR itself it declared DEVICE_LITTLE_ENDIAN which means
> > vfio_nvidia_quirk_mirror_write() receives byteswapped @data in the host
> > endian == bigendian on a big endian host. So the ioctl() handler will
> > receive a BE value, do byteswap #1 in leXX_to_cpu(), and then do byteswap
> > #2 in iowriteXX() so after all a BE will be written to a device. So I'd say
> > we rather do not need leXX_to_cpu() in vfio_pci_ioeventfd(). Correct me
> > where I am wrong. Thanks,  
> 
> It is not crystal clear to me what is the outcome of this discussion.
> Please can you clarify?
> 
> At the beginning I understood we had a chain of lexx_to_cpu and
> cpu_to_lexx (in iowritexx) so it was neutral. Now I am lost about what
> we want.

I've tried to address this with patch 2/3 in the newer series, adding
helpers such that the implicit endian-ness of the io{read,write}
functions is hidden and no extraneous swapping is done.  Therefore the
leXX_to_cpu() is gone, as Alexey wanted.  Unless there's a new
objection, this is what I intend to go with.  Thanks,

Alex