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

[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

 



On Fri, Dec 14, 2018 at 02:50:58PM +0100, Cornelia Huck wrote:
> On Fri, 14 Dec 2018 13:44:34 +0000
> Stefan Hajnoczi <stefanha@xxxxxxxxxx> 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.
> 
> But virtio-gpu works with ccw right now (I haven't checked what it
> uses); can virtio-fs use an equivalent method?

virtio-gpu does not use shared memory yet but it needs to in the future.

> If there's a more generic case to be made for extending virtio devices
> with a way to handle shared memory, a ccw for that would be fine. I
> just want to avoid adding new ccws for everything as the namespace is
> not infinite.

Yes, virtio-vhost-user needs it too.  I think it makes sense for shared
memory resources to be part of the VIRTIO device model.

Stefan

Attachment: signature.asc
Description: PGP signature


[Index of Archives]     [Linux Ext4 Filesystem]     [Union Filesystem]     [Filesystem Testing]     [Ceph Users]     [Ecryptfs]     [AutoFS]     [Kernel Newbies]     [Share Photos]     [Security]     [Netfilter]     [Bugtraq]     [Yosemite News]     [MIPS Linux]     [ARM Linux]     [Linux Security]     [Linux Cachefs]     [Reiser Filesystem]     [Linux RAID]     [Samba]     [Device Mapper]     [CEPH Development]

  Powered by Linux