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

Re: [Xen-devel] [PATCH v6 09/12] xen: add runtime parameter access support to hypfs

On 06.03.20 10:22, Jan Beulich wrote:
On 06.03.2020 10:20, Jürgen Groß wrote:
On 06.03.20 10:04, Jan Beulich wrote:
On 06.03.2020 09:47, Jürgen Groß wrote:
On 06.03.20 09:20, Jan Beulich wrote:
On 06.03.2020 07:42, Jürgen Groß wrote:
On 05.03.20 09:26, Jan Beulich wrote:
On 05.03.2020 07:01, Jürgen Groß wrote:
On 04.03.20 17:56, Jan Beulich wrote:
On 04.03.2020 17:31, Jürgen Groß wrote:
On 04.03.20 16:19, Jan Beulich wrote:
On 04.03.2020 16:07, Jürgen Groß wrote:
On 04.03.20 12:32, Jan Beulich wrote:
On 26.02.2020 13:47, Juergen Gross wrote:
+static void update_ept_param_append(const char *str, int val)
+    char *pos = opt_ept_setting + strlen(opt_ept_setting);
+    snprintf(pos, sizeof(opt_ept_setting) - (pos - opt_ept_setting),
+             ",%s=%d", str, val);
+static void update_ept_param(void)
+    snprintf(opt_ept_setting, sizeof(opt_ept_setting), "pml=%d", opt_ept_pml);
+    if ( opt_ept_ad >= 0 )
+        update_ept_param_append("ad", opt_ept_ad);

This won't correctly reflect reality: If you look at
vmx_init_vmcs_config(), even a negative value means "true" here,
unless on a specific Atom model. I think init_ept_param() wants
to have that erratum workaround logic moved there, such that
you can then assme the value to be non-negative here.

But isn't not mentioning it in the -1 case correct? -1 means: do the
correct thing on the current hardware.

Well, I think the output here should represent effective settings,

The minimum requirement is to reflect the effective parameters, like
cmdline is doing for boot-time only parameters. With runtime parameters
we had no way of telling what was set, and this is now possible.

and a sub-item should be suppressed only if a setting has no effect
at all in the current setup, like ...

+    if ( opt_ept_exec_sp >= 0 )
+        update_ept_param_append("exec-sp", opt_ept_exec_sp);

I agree for this one - if the value is still -1, it has neither
been set nor is its value of any interest.

... here.

I think we should not mix up specified parameters and effective
settings. In case an effective setting is of common interest it should
be reported via a specific node (like e.g. specific mitigation settings
where the cmdline is not providing enough details).

But then a boolean option that wasn't specified on the command line
should produce no output at all. And hence we'd need a way to tell
whether an option was set from command line for _all_ of them. I
don't think this would be very helpful.

I disagree here.

This is important only for cases where the hypervisor treats the
parameter as a tristate: true/false/unspecified. In all cases where
the bool value is really true or false it can be reported as such.

The problem I'm having with this is the resulting inconsistency:
When we write the variable with 0 or 1 in case we find it to be
-1 after command line parsing, the externally visible effect will
be different from the case where we leave it to be -1 yet still
treat it as (pseudo-)boolean. This, however, is an implementation
detail, while imo the hypfs presentation should not depend on
such implementation details.

Reporting 0/1 for e.g. "ad" if opt_ept_ad==-1 would add a latent problem
if any other action would be derived from the parameter variable being

So either opt_ept_ad should be modified to change it to 0/1 instead of
only setting the VCMS flag,

That's what I did suggest.

or the logic should be kept as is in this
patch. IMO changing the setting of opt_ept_ad should be done in another
patch if this is really wanted.

And of course I don't mind at all doing so in a prereq patch.
It's just that the patch here provides a good place _where_ to
actually do such an adjustment.

I was thinking of something like this:

--- a/xen/arch/x86/hvm/vmx/vmcs.c
+++ b/xen/arch/x86/hvm/vmx/vmcs.c
@@ -313,12 +313,12 @@ static int vmx_init_vmcs_config(void)
             rdmsrl(MSR_IA32_VMX_EPT_VPID_CAP, _vmx_ept_vpid_cap);

+        if ( /* Work around Erratum AVR41 on Avoton processors. */
+             boot_cpu_data.x86 == 6 && boot_cpu_data.x86_model == 0x4d &&
+             opt_ept_ad < 0 )
+            opt_ept_ad = 0;
             if ( !opt_ept_ad )
                 _vmx_ept_vpid_cap &= ~VMX_EPT_AD_BIT;
-        else if ( /* Work around Erratum AVR41 on Avoton processors. */
-                  boot_cpu_data.x86 == 6 && boot_cpu_data.x86_model == 0x4d &&
-                  opt_ept_ad < 0 )
-            _vmx_ept_vpid_cap &= ~VMX_EPT_AD_BIT;

              * Additional sanity checking before using EPT:

And I was specifically hoping to avoid doing this in a non-__init

Should be fairly easy (see attached patch).

Why not put the opt_ept_ad adjustment right into start_vmx(),
just like e.g. the opt_ept_exec_sp gets also done there? Pulling
the setting up of the 'v' key handler risks installing it ahead
of a potential future later error exit from start_vmx(). But I'm

Is this really problematic?

Not _really_, but still. In particular I'd prefer the 'v' key to
not even be listed among 'h' key output in such a case.

Now this is an optimization for a supposedly never to happen error
case only theoretically possible with future code additions.

In order to prepare for this case I don't think we should export
opt_ept_ad and put the setting of it at the very first thing to do
in start_vmx().


Xen-devel mailing list



Lists.xenproject.org is hosted with RackSpace, monitoring our
servers 24x7x365 and backed by RackSpace's Fanatical Support®.