Re: [PATCH v1 1/8] x86: Always use eager-fpu

[Andrew Cooper <andrew.cooper3@xxxxxxxxxx>]

On 19/03/2026 1:29 pm, Ross Lagerwall wrote:
> Lazy FPU avoids some work during a context switch but pushes the costs
> elsewhere:

I'd phrase this as "more expensive costs elsewhere".

We're trading off a slightly-longer XRSTOR now, for

1) in PV guests, an #NM exception
2) in HVM guests, an #NM exception and VMExit

and the longer XRSTOR.

Lazy is only a win even in theory if the sum of time handing #NM is less
than sum of time doing the longer XRSTOR, and with ...

> * For a workload running some Windows VMs, I measured about 83% of
>   context switches out had used the FPU so most of the time the FPU
>   save/restore is not avoided, just delayed.

... this "No, 83% of the time", falls firmly into "no not a win" category.

> * Setting/clearing the cr0.TS bit is serializing and reportedly slower
>   than the processor optimized xsave/restore.
> * Linux uses PKRU so a partial xsave/restore is performed on each
>   context switch anyway, followed by a second xsave/restore at some
>   point during execution.

"This interferes with the 'modified' optimisation that hardware uses to
try and reduce the cost of the following XSAVE".

>
> There is no measurable performance benefit for using lazy FPU and it
> adds unwanted complexity so remove the option and always use eager-fpu.
>
> Signed-off-by: Ross Lagerwall <ross.lagerwall@xxxxxxxxxx>

I think we want one other bullet point.

A key difference between 32bit and 64bit OSes is that %xmm is in the
base featureset for 64bit an thus get ubiquitous use in userspace.  This
is likely why we hit 83%.

Otherwise, Reviewed-by: Andrew Cooper <andrew.cooper3@xxxxxxxxxx>

> diff --git a/docs/misc/hypfs-paths.pandoc b/docs/misc/hypfs-paths.pandoc
> index e86f7d0dbef9..1553cb0bcb7f 100644
> --- a/docs/misc/hypfs-paths.pandoc
> +++ b/docs/misc/hypfs-paths.pandoc
> @@ -108,12 +108,10 @@ A populated Xen hypervisor file system might look like 
> the following example:
>              active-hvm/      directory for mitigations active in hvm doamins
>                  msr-spec-ctrl "No" or "Yes"
>                  rsb          "No" or "Yes"
> -                eager-fpu    "No" or "Yes"
>                  md-clear     "No" or "Yes"
>              active-pv/       directory for mitigations active in pv doamins
>                  msr-spec-ctrl "No" or "Yes"
>                  rsb          "No" or "Yes"
> -                eager-fpu    "No" or "Yes"
>                  md-clear     "No" or "Yes"
>                  xpti         "No" or list of "dom0", "domU", "PCID-on"
>                  l1tf-shadow  "No" or list of "dom0", "domU"

Juergen, do we want to nuke this whole paragraph?  I recall that we
never took the patch wiring up the speculation controls.

If so, it will be better to split this into a separate patch, rather
than to wonder why we've got a hypfs docs change with no associated code
change.

~Andrew