Re: [PATCH v2 2/2] drm/i915/gvt: subsitute kvm_read/write_guest with vfio_dma_rw

[Alex Williamson]

On Sun, 19 Jan 2020 05:06:37 -0500
Yan Zhao <yan.y.zhao@xxxxxxxxx> wrote:

> On Thu, Jan 16, 2020 at 11:37:29PM +0800, Alex Williamson wrote:
> > On Thu, 16 Jan 2020 00:49:41 -0500
> > Yan Zhao <yan.y.zhao@xxxxxxxxx> wrote:
> >   
> > > On Thu, Jan 16, 2020 at 04:06:51AM +0800, Alex Williamson wrote:  
> > > > On Tue, 14 Jan 2020 22:54:55 -0500
> > > > Yan Zhao <yan.y.zhao@xxxxxxxxx> wrote:
> > > >     
> > > > > As a device model, it is better to read/write guest memory using vfio
> > > > > interface, so that vfio is able to maintain dirty info of device IOVAs.
> > > > > 
> > > > > Compared to kvm interfaces kvm_read/write_guest(), vfio_dma_rw() has ~600
> > > > > cycles more overhead on average.
> > > > > 
> > > > > -------------------------------------
> > > > > |    interface     | avg cpu cycles |
> > > > > |-----------------------------------|
> > > > > | kvm_write_guest  |     1554       |
> > > > > | ----------------------------------|
> > > > > | kvm_read_guest   |     707        |
> > > > > |-----------------------------------|
> > > > > | vfio_dma_rw(w)   |     2274       |
> > > > > |-----------------------------------|
> > > > > | vfio_dma_rw(r)   |     1378       |
> > > > > -------------------------------------    
> > > > 
> > > > In v1 you had:
> > > > 
> > > > -------------------------------------
> > > > |    interface     | avg cpu cycles |
> > > > |-----------------------------------|
> > > > | kvm_write_guest  |     1546       |
> > > > | ----------------------------------|
> > > > | kvm_read_guest   |     686        |
> > > > |-----------------------------------|
> > > > | vfio_iova_rw(w)  |     2233       |
> > > > |-----------------------------------|
> > > > | vfio_iova_rw(r)  |     1262       |
> > > > -------------------------------------
> > > > 
> > > > So the kvm numbers remained within +0.5-3% while the vfio numbers are
> > > > now +1.8-9.2%.  I would have expected the algorithm change to at least
> > > > not be worse for small accesses and be better for accesses crossing
> > > > page boundaries.  Do you know what happened?
> > > >    
> > > I only tested the 4 interfaces in GVT's environment, where most of the
> > > guest memory accesses are less than one page.
> > > And the different fluctuations should be caused by the locks.
> > > vfio_dma_rw contends locks with other vfio accesses which are assumed to
> > > be abundant in the case of GVT.  
> > 
> > Hmm, so maybe it's time to convert vfio_iommu.lock from a mutex to a
> > rwsem?  Thanks,
> >   
> 
> hi Alex
> I tested your rwsem patches at (https://lkml.org/lkml/2020/1/16/1869).
> They works without any runtime error at my side. :) 
> However, I found out that the previous fluctuation may be because I didn't
> take read/write counts in to account.
> For example. though the two tests have different avg read/write cycles,
> their average cycles are almost the same.
>  ______________________________________________________________________
> |        | avg read |            | avg write |            |            |
> |        | cycles   | read cnt   | cycles    | write cnt  | avg cycles |
> |----------------------------------------------------------------------|
> | test 1 |   1339   | 29,587,120 |  2258     | 17,098,364 |    1676    |
> | test 2 |   1340   | 28,454,262 |  2238     | 16,501,788 |    1670    |
>  ----------------------------------------------------------------------
> 
> After measuring the exact read/write cnt and cycles of a specific workload,
> I get below findings:
> 
> (1) with single VM running glmark2 inside.
> glmark2: 40M+ read+write cnt, among which 63% is read.
> among reads, 48% is of PAGE_SIZE, the rest is less than a page.
> among writes, 100% is less than a page.
> 
>  __________________________________________________
> |       cycles         | read | write |  avg | inc |
> |--------------------------------------------------|
> | kvm_read/write_page  |  694 |  1506 |  993 |  /  |
> |--------------------------------------------------|
> |  vfio_dma_rw(mutex)  | 1340 |  2248 | 1673 | 680 |
> |--------------------------------------------------|
> | vfio_dma_rw(rwsem r) | 1323 |  2198 | 1645 | 653 |
>  ---------------------------------------------------
> 
> so vfio_dma_rw generally has 650+ more cycles per each read/write.
> While kvm->srcu is of 160 cycles on average with one vm is running, the
> cycles spending on locks for vfio_dma_rw spread like this:
>  ___________________________
> |        cycles       | avg |
> |---------------------------|
> |     iommu->lock     | 117 |
> |---------------------------|
> |   vfio.group_lock   | 108 |
> |---------------------------|
> | group->unbound_lock | 114 |
> |---------------------------|
> |  group->device_lock | 115 |
> |---------------------------|
> |     group->mutex    | 113 |
>  ---------------------------
> 
> I measured the cycles for a mutex without any contention is 104 cycles
> on average (including time for get_cycles() and measured in the same way
> as other locks). So the contention of a single lock in a single vm
> environment is light. probably because there's a vgpu lock hold in GVT already.
> 
> (2) with two VMs each running glmark2 inside.
> The contention increases a little.
> 
>  ___________________________________________________
> |       cycles         | read | write |  avg | inc  |
> |---------------------------------------------------|
> | kvm_read/write_page  | 1035 |  1832 | 1325 |  /   |
> |---------------------------------------------------|
> |  vfio_dma_rw(mutex)  | 2104 |  2886 | 2390 | 1065 |
> |---------------------------------------------------|
> | vfio_dma_rw(rwsem r) | 1965 |  2778 | 2260 | 935  |
>  ---------------------------------------------------
> 
> 
>  -----------------------------------------------
> |     avg cycles       |   one VM   |  two VMs  |
> |-----------------------------------------------|
> |  iommu lock (mutex)  |     117    |   150     |
> |-----------------------------------|-----------|
> | iommu lock (rwsem r) |     117    |   156     |
> |-----------------------------------|-----------|
> |   kvm->srcu          |     160    |   213     |
>  -----------------------------------------------
> 
> In the kvm case, avg cycles increased 332 cycles, while kvm->srcu only costed
> 213 cycles. The rest 109 cycles may be spent on atomic operations.
> But I didn't measure them, as get_cycles() operation itself would influence final
> cycles by ~20 cycles.

It seems like we need to extend the vfio external user interface so
that GVT-g can hold the group and container user references across
multiple calls.  For instance if we had a
vfio_group_get_external_user_from_dev() (based on
vfio_group_get_external_user()) then i915 could get an opaque
vfio_group pointer which it could use to call vfio_group_dma_rw() which
would leave us with only the iommu rw_sem locking.  i915 would release
the reference with vfio_group_put_external_user() when the device is
released.  The same could be done with the pin pages interface to
streamline that as well.  Thoughts?  Thanks,

Alex