On 4/22/2025 1:38 AM, Jürgen Groß wrote:
On 22.04.25 10:21, Xin Li (Intel) wrote:
To eliminate the indirect call overhead introduced by the pv_ops API,
use the alternatives mechanism to read PMC:
Which indirect call overhead? The indirect call is patched via the
alternative mechanism to a direct one.
See below.
1) When built with !CONFIG_XEN_PV, X86_FEATURE_XENPV becomes a
disabled feature, preventing the Xen PMC read code from being
built and ensuring the native code is executed unconditionally.
Without CONFIG_XEN_PV CONFIG_PARAVIRT_XXL is not selected, resulting in
native code anyway.
Yes, this is kept in this patch, but in a little different way.
2) When built with CONFIG_XEN_PV:
2.1) If not running on the Xen hypervisor (!X86_FEATURE_XENPV),
the kernel runtime binary is patched to unconditionally
jump to the native PMC read code.
2.2) If running on the Xen hypervisor (X86_FEATURE_XENPV), the
kernel runtime binary is patched to unconditionally jump
to the Xen PMC read code.
Consequently, remove the pv_ops PMC read API.
I don't see the value of this patch.
It adds more #ifdef and code lines without any real gain.
In case the x86 maintainers think it is still worth it, I won't object.
I think we want to totally bypass pv_ops in the case 2.1).
Do you mean the indirect call is patched to call native code *directly*
for 2.1? I don't know it, can you please elaborate?
AFAIK, Xen PV has been the sole user of pv_ops for nearly 20 years. This
raises significant doubts about whether pv_ops provides Linux with the
value of being a well-abstracted "CPU" or "Platform". And the x86
maintainers have said that it's a maintenance nightmare.
Thanks!
Xin
