Re: [Qemu-devel] [RFC 0/7] Rework vhost memory region updates

[Igor Mammedov]

On Thu, 30 Nov 2017 15:18:55 +0000
"Dr. David Alan Gilbert" <address@hidden> wrote:

> * Igor Mammedov (address@hidden) wrote:
> > On Thu, 30 Nov 2017 13:06:29 +0000
> > "Dr. David Alan Gilbert" <address@hidden> wrote:
> >   
> > > * Igor Mammedov (address@hidden) wrote:  
> > > > On Thu, 30 Nov 2017 12:47:20 +0000
> > > > "Dr. David Alan Gilbert" <address@hidden> wrote:
> > > >     
> > > > > * Igor Mammedov (address@hidden) wrote:    
> > > > > > On Thu, 30 Nov 2017 12:08:06 +0000
> > > > > > "Dr. David Alan Gilbert" <address@hidden> wrote:
> > > > > >       
> > > > > > > * Igor Mammedov (address@hidden) wrote:      
> > > > > > > > On Wed, 29 Nov 2017 18:50:19 +0000
> > > > > > > > "Dr. David Alan Gilbert (git)" <address@hidden> wrote:
> > > > > > > >         
> > > > > > > > > From: "Dr. David Alan Gilbert" <address@hidden>
> > > > > > > > > 
> > > > > > > > > Hi,
> > > > > > > > >   This is an experimental set that reworks the way the vhost
> > > > > > > > > code handles changes in physical address space layout that
> > > > > > > > > came from a discussion with Igor.        
> > > > > > > > Thanks for looking into it.
> > > > > > > > 
> > > > > > > >          
> > > > > > > > > Instead of updating and trying to merge sections of address
> > > > > > > > > space on each add/remove callback, we wait until the commit 
> > > > > > > > > phase
> > > > > > > > > and go through and rebuild a list by walking the Flatview of
> > > > > > > > > memory and end up producing an ordered list.
> > > > > > > > > We compare the list to the old list to trigger updates.
> > > > > > > > > 
> > > > > > > > > Note, only very lightly tested so far, I'm just trying to see 
> > > > > > > > > if it's
> > > > > > > > > the right shape.
> > > > > > > > > 
> > > > > > > > > Igor, is this what you were intending?        
> > > > > > > > 
> > > > > > > > I was thinking about a little less intrusive approach
> > > > > > > > 
> > > > > > > > where vhost_region_add/del are modified to maintain
> > > > > > > > sorted by GPA array of mem_sections, vhost_dev::mem is dropped
> > > > > > > > altogether and vhost_memory_region array is build/used/freed
> > > > > > > > on every vhost_commit().
> > > > > > > > Maintaining sorted array should roughly cost us O(2 log n) if
> > > > > > > > binary search is used.
> > > > > > > > 
> > > > > > > > However I like your idea with iterator even more as it have
> > > > > > > > potential to make it even faster O(n) if we get rid of
> > > > > > > > quadratic and relatively complex vhost_update_compare_list().   
> > > > > > > >      
> > > > > > > 
> > > > > > > Note vhost_update_compare_list is complex,
> > > > > > > but it is O(n) - it's
> > > > > > > got nested loops, but the inner loop moves forward and oldi never
> > > > > > > gets reset back to zero.      
> > > > > > While skimming through patches I've overlooked it.
> > > > > > 
> > > > > > Anyways,
> > > > > > why memcmp(old_arr, new_arr) is not sufficient
> > > > > > to detect a change in memory map?      
> > > > > 
> > > > > It tells you that you've got a change, but doesn't give
> > > > > the start/end of the range that's changed, and those
> > > > > are used by vhost_commit to limit the work of
> > > > > vhost_verify_ring_mappings.    
> > > > Isn't memmap list a sorted and
> > > >  dev->mem_changed_[start|end]_addr are the lowest|highest
> > > > addresses of whole map?
> > > > 
> > > > If it's, so wouldn't getting values directly from 
> > > > the 1st/last entries of array be sufficient?    
> > > 
> > > THat wasn't my understanding from the existing code;
> > > my understanding was that changed_start_addr is set by
> > > vhost_region_add->vhost_set_memory when a new region is added
> > > (or removed) and is set to the limit of the section added.
> > > But perhaps I'm misunderstanding.  
> > changed_*_addr is actually lower/upper bound of memory transaction
> > and in practice it often includes several memory sections that
> > get mapped during transaction (between begin - commit).
> > 
> > but then again,
> >   - how expensive vhost_verify_ring_mappings() is?
> >   - do we really need optimization here (perhaps finding out changed range 
> > is more expensive)?
> >   - how about calling vhost_verify_ring_mappings() unconditionally?  
> 
> My worry is that:
>     vhost_verify_ring_mappings
>        vhost_verify_ring_part_mapping
>           vhost_verify_ring_part_mapping
>              vhost_memory_map & vhost_memory_unmap
>                (non-iommu case...)
>                cpu_physical_memory_map & cpu_physical_memory_unmap
>                  address_space_map/address_space_unmap
>                     flatview_translate etc
> 
> so it's not cheap at all - I *think* it should end up doing very little
> after it's gone all the way through that because it should already be
> mapped; but still it's not trivial.
neither trivial finding out changed range.
How often it will be called and what actual time it takes
for vhost_verify_ring_mappings and vhost_update_compare_list to complete.

note 
vhost_memory_map() would be run only on ranges that
overlap with rings (typically 1), while vhost_update_compare_list()
would go over all ranges.
So question is does optimization really saves us anything?


> 
> Dave
> 
> >   
> > > (The logic in vhost_verify_ring_mappings doesn't make sense
> > > to me either though; if vhost_verify_ring_part_mapping returns 0
> > > on success, why is it doing   if (!r) { break; }  surely it
> > > should be  if (r) { break; })  
> > it looks like a bug (CCing Greg)
> > 
> > before (f1f9e6c5 vhost: adapt vhost_verify_ring_mappings() to virtio 1 ring 
> > layout)
> > logic used to be
> > 
> >    if changed_*_addr doesn't contain ring
> >       "IGNORE as we don't care"
> >    
> >    if changed_*_addr contain ring AND ring can't be mapped at the same place
> >       ABORT
> > 
> > with f1f9e6c5 we have 3 rings so on any of them following could happen
> >    if "IGNORE as we don't care"
> >      break => false success 
> >      since it's possible that the remaining rings in vq do overlap and 
> > didn't get checked
> >   
> --
> Dr. David Alan Gilbert / address@hidden / Manchester, UK