Re: [PATCH v2 18/30] virtio_fs, dax: Set up virtio_fs dax_device

[Christian Borntraeger]

On 22.07.19 13:43, Cornelia Huck wrote:
> On Mon, 22 Jul 2019 13:20:18 +0200
> Christian Borntraeger <borntraeger@xxxxxxxxxx> wrote:
> 
>> On 22.07.19 12:56, Dr. David Alan Gilbert wrote:
>>> * Christian Borntraeger (borntraeger@xxxxxxxxxx) wrote:  
>>>>
>>>>
>>>> On 18.07.19 16:30, Dan Williams wrote:  
>>>>> On Thu, Jul 18, 2019 at 6:15 AM Vivek Goyal <vgoyal@xxxxxxxxxx> wrote:  
>>>>>>
>>>>>> On Wed, Jul 17, 2019 at 07:27:25PM +0200, Halil Pasic wrote:  
>>>>>>> On Wed, 15 May 2019 15:27:03 -0400
>>>>>>> Vivek Goyal <vgoyal@xxxxxxxxxx> wrote:
>>>>>>>  
>>>>>>>> From: Stefan Hajnoczi <stefanha@xxxxxxxxxx>
>>>>>>>>
>>>>>>>> Setup a dax device.
>>>>>>>>
>>>>>>>> Use the shm capability to find the cache entry and map it.
>>>>>>>>
>>>>>>>> The DAX window is accessed by the fs/dax.c infrastructure and must have
>>>>>>>> struct pages (at least on x86).  Use devm_memremap_pages() to map the
>>>>>>>> DAX window PCI BAR and allocate struct page.
>>>>>>>>  
>>>>>>>
>>>>>>> Sorry for being this late. I don't see any more recent version so I will
>>>>>>> comment here.
>>>>>>>
>>>>>>> I'm trying to figure out how is this supposed to work on s390. My concern
>>>>>>> is, that on s390 PCI memory needs to be accessed by special
>>>>>>> instructions. This is taken care of by the stuff defined in
>>>>>>> arch/s390/include/asm/io.h. E.g. we 'override' __raw_writew so it uses
>>>>>>> the appropriate s390 instruction. However if the code does not use the
>>>>>>> linux abstractions for accessing PCI memory, but assumes it can be
>>>>>>> accessed like RAM, we have a problem.
>>>>>>>
>>>>>>> Looking at this patch, it seems to me, that we might end up with exactly
>>>>>>> the case described. For example AFAICT copy_to_iter() (3) resolves to
>>>>>>> the function in lib/iov_iter.c which does not seem to cater for s390
>>>>>>> oddities.
>>>>>>>
>>>>>>> I didn't have the time to investigate this properly, and since virtio-fs
>>>>>>> is virtual, we may be able to get around what is otherwise a
>>>>>>> limitation on s390. My understanding of these areas is admittedly
>>>>>>> shallow, and since I'm not sure I'll have much more time to
>>>>>>> invest in the near future I decided to raise concern.
>>>>>>>
>>>>>>> Any opinions?  
>>>>>>
>>>>>> Hi Halil,
>>>>>>
>>>>>> I don't understand s390 and how PCI works there as well. Is there any
>>>>>> other transport we can use there to map IO memory directly and access
>>>>>> using DAX?
>>>>>>
>>>>>> BTW, is DAX supported for s390.
>>>>>>
>>>>>> I am also hoping somebody who knows better can chip in. Till that time,
>>>>>> we could still use virtio-fs on s390 without DAX.  
>>>>>
>>>>> s390 has so-called "limited" dax support, see CONFIG_FS_DAX_LIMITED.
>>>>> In practice that means that support for PTE_DEVMAP is missing which
>>>>> means no get_user_pages() support for dax mappings. Effectively it's
>>>>> only useful for execute-in-place as operations like fork() and ptrace
>>>>> of dax mappings will fail.  
>>>>
>>>>
>>>> This is only true for the dcssblk device driver (drivers/s390/block/dcssblk.c
>>>> and arch/s390/mm/extmem.c). 
>>>>
>>>> For what its worth, the dcssblk looks to Linux like normal memory (just above the
>>>> previously detected memory) that can be used like normal memory. In previous time
>>>> we even had struct pages for this memory - this was removed long ago (when it was
>>>> still xip) to reduce the memory footprint for large dcss blocks and small memory
>>>> guests.
>>>> Can the CONFIG_FS_DAX_LIMITED go away if we have struct pages for that memory?
>>>>
>>>> Now some observations: 
>>>> - dcssblk is z/VM only (not KVM)
>>>> - Setting CONFIG_FS_DAX_LIMITED globally as a Kconfig option depending on wether
>>>>   a device driver is compiled in or not seems not flexible enough in case if you
>>>>   have device driver that does have struct pages and another one that doesn't
>>>> - I do not see a reason why we should not be able to map anything from QEMU
>>>>   into the guest real memory via an additional KVM memory slot. 
>>>>   We would need to handle that in the guest somehow (and not as a PCI bar),
>>>>   register this with struct pages etc.
> 
> You mean for ccw, right? I don't think we want pci to behave
> differently than everywhere else.

Yes for virtio-ccw. We would need to have a look at how virtio-ccw can create a memory
mapping with struct pages, so that DAX will work.(Dan, it is just struct pages that 
you need, correct?)


> 
>>>> - we must then look how we can create the link between the guest memory and the
>>>>   virtio-fs driver. For virtio-ccw we might be able to add a new ccw command or
>>>>   whatever. Maybe we could also piggy-back on some memory hotplug work from David
>>>>   Hildenbrand (add cc).
>>>>
>>>> Regarding limitations on the platform:
>>>> - while we do have PCI, the virtio devices are usually plugged via the ccw bus.
>>>>   That implies no PCI bars. I assume you use those PCI bars only to implicitely 
>>>>   have the location of the shared memory
>>>>   Correct?  
>>>
>>> Right.  
>>
>> So in essence we just have to provide a vm_get_shm_region callback in the virtio-ccw
>> guest code?
>>
>> How many regions do we have to support? One region per device? Or many?
>> Even if we need more, this should be possible with a 2 new CCWs, e.g READ_SHM_BASE(id)
>> and READ_SHM_SIZE(id)
> 
> I'd just add a single CCW with a control block containing id and size.
> 
> The main issue is where we put those regions, and what happens if we
> use both virtio-pci and virtio-ccw on the same machine.

Then these 2 devices should get independent memory regions that are added in an
independent (but still exclusive) way.
> 
>>
>>
>>>   
>>>> - no real memory mapped I/O. Instead there are instructions that work on the mmio.
>>>>   As I understand things, this is of no concern regarding virtio-fs as you do not
>>>>   need mmio in the sense that a memory access of the guest to such an address 
>>>>   triggers an exit. You just need the shared memory as a mean to have the data
>>>>   inside the guest. Any notification is done via normal virtqueue mechanisms
>>>>   Correct?  
>>>
>>> Yep.  
>>
>