Re: [PATCH 18/52] virtio-fs: Map cache using the values from the capabilities

[David Hildenbrand]

On 14.12.18 14:44, Stefan Hajnoczi wrote:
> On Thu, Dec 13, 2018 at 01:38:23PM +0100, Cornelia Huck wrote:
>> On Thu, 13 Dec 2018 13:24:31 +0100
>> David Hildenbrand <david@xxxxxxxxxx> wrote:
>>
>>> On 13.12.18 13:15, Dr. David Alan Gilbert wrote:
>>>> * David Hildenbrand (david@xxxxxxxxxx) wrote:  
>>>>> On 13.12.18 11:00, Dr. David Alan Gilbert wrote:  
>>>>>> * David Hildenbrand (david@xxxxxxxxxx) wrote:  
>>>>>>> On 13.12.18 10:13, Dr. David Alan Gilbert wrote:  
>>>>>>>> * David Hildenbrand (david@xxxxxxxxxx) wrote:  
>>>>>>>>> On 10.12.18 18:12, Vivek Goyal wrote:  
>>>>>>>>>> Instead of assuming we had the fixed bar for the cache, use the
>>>>>>>>>> value from the capabilities.
>>>>>>>>>>
>>>>>>>>>> Signed-off-by: Dr. David Alan Gilbert <dgilbert@xxxxxxxxxx>
>>>>>>>>>> ---
>>>>>>>>>>  fs/fuse/virtio_fs.c | 32 +++++++++++++++++---------------
>>>>>>>>>>  1 file changed, 17 insertions(+), 15 deletions(-)
>>>>>>>>>>
>>>>>>>>>> diff --git a/fs/fuse/virtio_fs.c b/fs/fuse/virtio_fs.c
>>>>>>>>>> index 60d496c16841..55bac1465536 100644
>>>>>>>>>> --- a/fs/fuse/virtio_fs.c
>>>>>>>>>> +++ b/fs/fuse/virtio_fs.c
>>>>>>>>>> @@ -14,11 +14,6 @@
>>>>>>>>>>  #include <uapi/linux/virtio_pci.h>
>>>>>>>>>>  #include "fuse_i.h"
>>>>>>>>>>  
>>>>>>>>>> -enum {
>>>>>>>>>> -	/* PCI BAR number of the virtio-fs DAX window */
>>>>>>>>>> -	VIRTIO_FS_WINDOW_BAR = 2,
>>>>>>>>>> -};
>>>>>>>>>> -
>>>>>>>>>>  /* List of virtio-fs device instances and a lock for the list */
>>>>>>>>>>  static DEFINE_MUTEX(virtio_fs_mutex);
>>>>>>>>>>  static LIST_HEAD(virtio_fs_instances);
>>>>>>>>>> @@ -518,7 +513,7 @@ static int virtio_fs_setup_dax(struct virtio_device *vdev, struct virtio_fs *fs)
>>>>>>>>>>  	struct dev_pagemap *pgmap;
>>>>>>>>>>  	struct pci_dev *pci_dev;
>>>>>>>>>>  	phys_addr_t phys_addr;
>>>>>>>>>> -	size_t len;
>>>>>>>>>> +	size_t bar_len;
>>>>>>>>>>  	int ret;
>>>>>>>>>>  	u8 have_cache, cache_bar;
>>>>>>>>>>  	u64 cache_offset, cache_len;
>>>>>>>>>> @@ -551,17 +546,13 @@ static int virtio_fs_setup_dax(struct virtio_device *vdev, struct virtio_fs *fs)
>>>>>>>>>>          }
>>>>>>>>>>  
>>>>>>>>>>  	/* TODO handle case where device doesn't expose BAR? */  
>>>>>>>>>
>>>>>>>>> For virtio-pmem we decided to not go via BARs as this would effectively
>>>>>>>>> make it only usable for virtio-pci implementers. Instead, we are going
>>>>>>>>> to export the applicable physical device region directly (e.g.
>>>>>>>>> phys_start, phys_size in virtio config), so it is decoupled from PCI
>>>>>>>>> details. Doing the same for virtio-fs would allow e.g. also virtio-ccw
>>>>>>>>> to make eventually use of this.  
>>>>>>>>
>>>>>>>> That makes it a very odd looking PCI device;  I can see that with
>>>>>>>> virtio-pmem it makes some sense, given that it's job is to expose
>>>>>>>> arbitrary chunks of memory.
>>>>>>>>
>>>>>>>> Dave  
>>>>>>>
>>>>>>> Well, the fact that your are
>>>>>>>
>>>>>>> - including <uapi/linux/virtio_pci.h>
>>>>>>> - adding pci related code
>>>>>>>
>>>>>>> in/to fs/fuse/virtio_fs.c
>>>>>>>
>>>>>>> tells me that these properties might be better communicated on the
>>>>>>> virtio layer, not on the PCI layer.
>>>>>>>
>>>>>>> Or do you really want to glue virtio-fs to virtio-pci for all eternity?  
>>>>>>
>>>>>> No, these need cleaning up; and the split within the bar
>>>>>> is probably going to change to be communicated via virtio layer
>>>>>> rather than pci capabilities.  However, I don't want to make our PCI
>>>>>> device look odd, just to make portability to non-PCI devices - so it's
>>>>>> right to make the split appropriately, but still to use PCI bars
>>>>>> for what they were designed for.
>>>>>>
>>>>>> Dave  
>>>>>
>>>>> Let's discuss after the cleanup. In general I am not convinced this is
>>>>> the right thing to do. Using virtio-pci for anything else than pure
>>>>> transport smells like bad design to me (well, I am no virtio expert
>>>>> after all ;) ). No matter what PCI bars were designed for. If we can't
>>>>> get the same running with e.g. virtio-ccw or virtio-whatever, it is
>>>>> broken by design (or an addon that is tightly glued to virtio-pci, if
>>>>> that is the general idea).  
>>>>
>>>> I'm sure we can find alternatives for virtio-*, so I wouldn't expect
>>>> it to be glued to virtio-pci.
>>>>
>>>> Dave  
>>>
>>> As s390x does not have the concept of memory mapped io (RAM is RAM,
>>> nothing else), this is not architectured. vitio-ccw can therefore not
>>> define anything similar like that. However, in virtual environments we
>>> can do whatever we want on top of the pure transport (e.g. on the virtio
>>> layer).
>>>
>>> Conny can correct me if I am wrong.
>>
>> I don't think you're wrong, but I haven't read the code yet and I'm
>> therefore not aware of the purpose of this BAR.
>>
>> Generally, if there is a memory location shared between host and guest,
>> we need a way to communicate its location, which will likely differ
>> between transports. For ccw, I could imagine a new channel command
>> dedicated to exchanging configuration information (similar to what
>> exists today to communicate the locations of virtqueues), but I'd
>> rather not go down this path.
>>
>> Without reading the code/design further, can we use one of the
>> following instead of a BAR:
>> - a virtqueue;
>> - something in config space?
>> That would be implementable by any virtio transport.
> 
> The way I think about this is that we wish to extend the VIRTIO device
> model with the concept of shared memory.  virtio-fs, virtio-gpu, and
> virtio-vhost-user all have requirements for shared memory.
> 
> This seems like a transport-level issue to me.  PCI supports
> memory-mapped I/O and that's the right place to do it.  If you try to
> put it into config space or the virtqueue, you'll end up with something
> that cannot be realized as a PCI device because it bypasses PCI bus
> address translation.
> 
> If CCW needs a side-channel, that's fine.  But that side-channel is a
> CCW-specific mechanism and probably doesn't apply to all other
> transports.
> 
> Stefan
> 

I think the problem is more fundamental. There is no iommu. Whatever
shared region you want to indicate, you want it to be assigned a memory
region in guest physical memory. Like a DIMM/NVDIMM. And this should be
different to the concept of a BAR. Or am I missing something?

I am ok with using whatever other channel to transport such information.
But I believe this is different to a typical BAR. (I wish I knew more
about PCI internals ;) ).

I would also like to know how shared memory works as of now for e.g.
virtio-gpu.

-- 

Thanks,

David / dhildenb