On 06/20/2014 01:21 PM, Alex Williamson wrote:
> On Thu, 2014-06-19 at 13:48 +1000, Alexey Kardashevskiy wrote:
>> On 06/19/2014 11:50 AM, Alexey Kardashevskiy wrote:
>>> On 06/19/2014 10:50 AM, Alexey Kardashevskiy wrote:
>>>> On 06/19/2014 04:35 AM, Alex Williamson wrote:
>>>>> On Wed, 2014-06-18 at 21:36 +1000, Alexey Kardashevskiy wrote:
>>>>>> VFIO exposes BARs to user space as a byte stream so userspace can
>>>>>> read it using pread()/pwrite(). Since this is a byte stream, VFIO should
>>>>>> not do byte swapping and simply return values as it gets them from
>>>>>> PCI device.
>>>>>>
>>>>>> Instead, the existing code assumes that byte stream in read/write is
>>>>>> little-endian and it fixes endianness for values which it passes to
>>>>>> ioreadXX/iowriteXX helpers. This works for little-endian as PCI is
>>>>>> little endian and le32_to_cpu/... are stubs.
>>>>>
>>>>> vfio read32:
>>>>>
>>>>> val = cpu_to_le32(ioread32(io + off));
>>>>>
>>>>> Where the typical x86 case, ioread32 is:
>>>>>
>>>>> #define ioread32(addr) readl(addr)
>>>>>
>>>>> and readl is:
>>>>>
>>>>> __le32_to_cpu(__raw_readl(addr));
>>>>>
>>>>> So we do canceling byte swaps, which are both nops on x86, and end up
>>>>> returning device endian, which we assume is little endian.
>>>>>
>>>>> vfio write32 is similar:
>>>>>
>>>>> iowrite32(le32_to_cpu(val), io + off);
>>>>>
>>>>> The implicit cpu_to_le32 of iowrite32() and our explicit swap cancel
>>>>> out, so input data is device endian, which is assumed little.
>>>>>
>>>>>> This also works for big endian but rather by an accident: it reads 4 bytes
>>>>>> from the stream (@val is big endian), converts to CPU format (which should
>>>>>> be big endian) as it was little endian (@val becomes actually little
>>>>>> endian) and calls iowrite32() which does not do swapping on big endian
>>>>>> system.
>>>>>
>>>>> Really?
>>>>>
>>>>> In arch/powerpc/kernel/iomap.c iowrite32() is just a wrapper around
>>>>> writel(), which seems to use the generic implementation, which does
>>>>> include a cpu_to_le32.
>>>>
>>>>
>>>> Ouch, wrong comment. iowrite32() does swapping. My bad.
>>>>
>>>>
>>>>>
>>>>> I also see other big endian archs like parisc doing cpu_to_le32 on
>>>>> iowrite32, so I don't think this statement is true. I imagine it's
>>>>> probably working for you because the swap cancel.
>>>>>
>>>>>> This removes byte swapping and makes use ioread32be/iowrite32be
>>>>>> (and 16bit versions) on big-endian systems. The "be" helpers take
>>>>>> native endian values and do swapping at the moment of writing to a PCI
>>>>>> register using one of "store byte-reversed" instructions.
>>>>>
>>>>> So now you want iowrite32() on little endian and iowrite32be() on big
>>>>> endian, the former does a cpu_to_le32 (which is a nop on little endian)
>>>>> and the latter does a cpu_to_be32 (which is a nop on big endian)...
>>>>> should we just be using __raw_writel() on both?
>>>>
>>>>
>>>> We can do that too. The beauty of iowrite32be on ppc64 is that it does not
>>>> swap and write separately, it is implemented via the "Store Word
>>>> Byte-Reverse Indexed X-form" single instruction.
>>>>
>>>> And some archs (do not know which ones) may add memory barriers in their
>>>> implementations of ioread/iowrite. __raw_writel is too raw :)
>>>>
>>>>> There doesn't actually
>>>>> seem to be any change in behavior here, it just eliminates back-to-back
>>>>> byte swaps, which are a nop on x86, but not power, right?
>>>>
>>>> Exactly. No dependency for QEMU.
>>>
>>> How about that:
>>> ===
>>>
>>> VFIO exposes BARs to user space as a byte stream so userspace can
>>> read it using pread()/pwrite(). Since this is a byte stream, VFIO should
>>> not do byte swapping and simply return values as it gets them from
>>> PCI device.
>>>
>>> Instead, the existing code assumes that byte stream in read/write is
>>> little-endian and it fixes endianness for values which it passes to
>>> ioreadXX/iowriteXX helpers in native format. The IO helpers do swapping
>>> again. Since both byte swaps are nops on little-endian host, this works.
>>>
>>> This also works for big endian but rather by an accident: it reads 4 bytes
>>> from the stream (@val is big endian), converts to CPU format (which should
>>> be big endian) as it was little endian (and @val becomes actually little
>>> endian) and calls iowrite32() which does swapping on big endian
>>> system again. So byte swap gets cancelled, __raw_writel() receives
>>> a native value and then
>>> *(volatile unsigned int __force *)PCI_FIX_ADDR(addr) = v;
>>> just does the right thing.
>>
>> I am wrong here, sorry. This is what happens when you watch soccer between
>> 2am and 4am :)
>>
>>
>>>
>>> This removes byte swaps and makes use of ioread32be/iowrite32be
>>> (and 16bit versions) which do explicit byte swapping at the moment
>>> of write to a PCI register. PPC64 uses a special "Store Word
>>> Byte-Reverse Indexed X-form" instruction which does swap and store.
>>
>> No swapping is done here if we use ioread32be as it calls in_be32 and that
>> animal does "lwz" which is simple load from memory.
>>
>> So @val (16/32 bit variable on stack) will have different values on LE and
>> BE but since we do not handle it the host and just memcpy it to the buffer,
>> nothing breaks here.
>>
>>
>> So it should be like this:
>> ===
>> VFIO exposes BARs to user space as a byte stream so userspace can
>> read it using pread()/pwrite(). Since this is a byte stream, VFIO should
>> not do byte swapping and simply return values as it gets them from
>> PCI device and copy_to_user will save bytes in the correct
>> same true for writes.
>>
>> Instead, the existing code assumes that byte stream in read/write is
>> little-endian and it fixes endianness for values which it passes to
>> ioreadXX/iowriteXX helpers in native format. The IO helpers do swapping
>> again. Since both byte swaps are nops on little-endian host, this works.
>>
>> This also works for big endian but rather by an accident: it reads 4 bytes
>> from the stream (@val is big endian), converts to CPU format (which should
>> be big endian) as it was little endian (and @val becomes actually little
>> endian) and calls iowrite32() which does swapping on big endian
>> system again. So byte swap in the host gets cancelled and __raw_writel()
>> writes the value which was swapped originally by the guest.
>>
>> This removes byte swaps and makes use of ioread32be/iowrite32be
>> (and 16bit versions) which do not do byte swap on BE hosts.
>> For LE hosts, ioread32/iowrite32 are still used.
>>
>> ===
>
> Working on big endian being an accident may be a matter of perspective.
> The comment remains that this patch doesn't actually fix anything except
> the overhead on big endian systems doing redundant byte swapping and
> maybe the philosophy that vfio regions are little endian.
They are little-endian only between ioread32() and copy_to_user() calls,
besides that it is a bytes stream which does not have endianness so I do
not understand the comment about philosophy...
> I'm still not a fan of iowrite vs iowritebe, there must be something we
> can use that doesn't have an implicit swap. Calling it iowrite*_native
> is also an abuse of the namespace. Next thing we know some common code
> will legitimately use that name. If we do need to define an alias
> (which I'd like to avoid) it should be something like vfio_iowrite32.
We can still use __raw_writel&co, would that be ok?
> Thanks,
>
> Alex
>
>>> ===
>>>
>>> any better?
>>>
>>>
>>>
>>>
>>>>>> Suggested-by: Benjamin Herrenschmidt <benh@xxxxxxxxxxxxxxxxxxx>
>>>>>> Signed-off-by: Alexey Kardashevskiy <aik@xxxxxxxxx>
>>>>>> ---
>>>>>> drivers/vfio/pci/vfio_pci_rdwr.c | 20 ++++++++++++++++----
>>>>>> 1 file changed, 16 insertions(+), 4 deletions(-)
>>>>>>
>>>>>> diff --git a/drivers/vfio/pci/vfio_pci_rdwr.c b/drivers/vfio/pci/vfio_pci_rdwr.c
>>>>>> index 210db24..f363b5a 100644
>>>>>> --- a/drivers/vfio/pci/vfio_pci_rdwr.c
>>>>>> +++ b/drivers/vfio/pci/vfio_pci_rdwr.c
>>>>>> @@ -21,6 +21,18 @@
>>>>>>
>>>>>> #include "vfio_pci_private.h"
>>>>>>
>>>>>> +#ifdef __BIG_ENDIAN__
>>>>>> +#define ioread16_native ioread16be
>>>>>> +#define ioread32_native ioread32be
>>>>>> +#define iowrite16_native iowrite16be
>>>>>> +#define iowrite32_native iowrite32be
>>>>>> +#else
>>>>>> +#define ioread16_native ioread16
>>>>>> +#define ioread32_native ioread32
>>>>>> +#define iowrite16_native iowrite16
>>>>>> +#define iowrite32_native iowrite32
>>>>>> +#endif
>>>>>> +
>>>>>> /*
>>>>>> * Read or write from an __iomem region (MMIO or I/O port) with an excluded
>>>>>> * range which is inaccessible. The excluded range drops writes and fills
>>>>>> @@ -50,9 +62,9 @@ static ssize_t do_io_rw(void __iomem *io, char __user *buf,
>>>>>> if (copy_from_user(&val, buf, 4))
>>>>>> return -EFAULT;
>>>>>>
>>>>>> - iowrite32(le32_to_cpu(val), io + off);
>>>>>> + iowrite32_native(val, io + off);
>>>>>> } else {
>>>>>> - val = cpu_to_le32(ioread32(io + off));
>>>>>> + val = ioread32_native(io + off);
>>>>>>
>>>>>> if (copy_to_user(buf, &val, 4))
>>>>>> return -EFAULT;
>>>>>> @@ -66,9 +78,9 @@ static ssize_t do_io_rw(void __iomem *io, char __user *buf,
>>>>>> if (copy_from_user(&val, buf, 2))
>>>>>> return -EFAULT;
>>>>>>
>>>>>> - iowrite16(le16_to_cpu(val), io + off);
>>>>>> + iowrite16_native(val, io + off);
>>>>>> } else {
>>>>>> - val = cpu_to_le16(ioread16(io + off));
>>>>>> + val = ioread16_native(io + off);
>>>>>>
>>>>>> if (copy_to_user(buf, &val, 2))
>>>>>> return -EFAULT;
>>>>>
>>>>>
>>>>>
>>>>
>>>>
>>>
>>>
>>
>>
>
>
>
--
Alexey
--
To unsubscribe from this list: send the line "unsubscribe kvm" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
