Re: [PATCH v2 1/2] x86/fpu: Extend kernel_fpu_begin_mask() to initialize AMX state

[Dave Hansen <dave.hansen@intel.com>]

On 5/8/24 17:29, Chang S. Bae wrote:
> +void kernel_fpu_reset(void)
> +{
> +       kernel_fpu_begin();
> +       if (cpu_feature_enabled(X86_FEATURE_AMX_TILE))
> +               tile_release();
> +       kernel_fpu_end();
> +}
> +EXPORT_SYMBOL(kernel_fpu_reset);
> +
...
> --- a/drivers/platform/x86/intel/ifs/runtest.c
> +++ b/drivers/platform/x86/intel/ifs/runtest.c
> @@ -188,6 +188,8 @@ static int doscan(void *data)
>         /* Only the first logical CPU on a core reports result */
>         first = cpumask_first(cpu_smt_mask(cpu));
> 
> +       kernel_fpu_reset();
> +
>         wait_for_sibling_cpu(&scan_cpus_in, NSEC_PER_SEC);

Remember, kernel_fpu_begin/end() mark a section of code that needs the
FPU.  Once code calls kernel_fpu_end(), it no longer owns the FPU and
all bets are off.  A interrupt could theoretically come in and do
whatever it wants.

I _assume_ that this is practically impossible since the stop_machine()
infrastructure keeps interrupts at bay.  But it's rather subtle.

I'd probably just do this:

+	kernel_fpu_begin();
+	// AMX *MUST* be in the init state for the wrmsr() to work.
+	// But, the more in the init state, the less state the test
+	// has to save and restore.  Just zap everything.
+	restore_fpregs_from_fpstate(&init_fpstate,	
+				    fpu_user_cfg.max_features);
+
        wrmsrl(MSR_ACTIVATE_SCAN, params->activate->data);
        rdmsrl(MSR_SCAN_STATUS, status.data);

+	kernel_fpu_end();

That's dirt simple.  It doesn't require new infrastructure.  It doesn't
call an opaque new helper.  It doesn't require a feature check.  It
probably makes the IFS test run faster.  It will also magically work for
any fancy new feature that comes along which *ALSO* needs to be in its
init state ... with zero changes to this code.  For bonus points, this
code is quite universal.  It will work, as-is, in a bunch of kernel
contexts if future deranged kernel developer copies and pastes it.  The
code you suggested above can race unless it's called under
stop_machine() and isn't safe to copy elsewhere.

Three lines of code:

	1. IFS declares its need to own the FPU for a moment, like any
	   other kernel_fpu_begin() user.  It's not a special snowflake.
	   It is boring.
	2. IFS zaps the FPU state
	3. IFS gives up the FPU

Am I out of my mind?  What am I missing?  Why bother with _anything_
more complicated than this?