On Thu, Jan 14, 2021 at 01:36:57PM +0100, Krzysztof Mazur wrote:
The OSFXSR must be set only on CPUs with SSE. There are some CPUs with 3DNow!, but without SSE and FXSR (like AMD Geode LX, which is still used in many embedded systems). So, I've changed that to:
if (unlikely(in_interrupt()) || (boot_cpu_has(X86_FEATURE_XMM) && unlikely(!(cr4_read_shadow() & X86_CR4_OSFXSR))))
Why?
X86_CR4_OSFXSR won't ever be set on those CPUs but the test will be performed anyway. So there's no need for boot_cpu_has().
However, except for some embedded systems, probably almost nobody uses that code today, and the most imporant is avoiding future breakage.
Yes, exactly. K7 is a don't-care performance-wise nowadays.
And I'm not sure which approach is better. Because that CR4 test tests for a feature that is not used in mmx_memcpy(), but it's used in kernel_fpu_begin(). And in future kernel_fpu_begin() may change and require also other stuff. So I think that the best approach would be delay any FPU optimized memcpy() after fpu__init_system() is executed.
I'd prefer the simplest approach as this code is almost never used. So no need to complicate things unnecessarily.
Subject: [PATCH] x86/lib: don't use mmx_memcpy() to early
s/to/too/
The MMX 3DNow! optimized memcpy() is used very early, even before FPU is initialized in the kernel. It worked fine, but commit 7ad816762f9bf89e940e618ea40c43138b479e10 ("x86/fpu: Reset MXCSR to default in kernel_fpu_begin()") broke that. After that commit the kernel_fpu_begin() assumes that FXSR is enabled in the CR4 register on all processors with SSE. Because memcpy() is used before FXSR is enabled, the kernel crashes just after "Booting the kernel." message. It affects all kernels with CONFIG_X86_USE_3DNOW (enabled when some AMD/Cyrix processors are selected) on processors with SSE (like AMD K7, which supports both MMX 3DNow! and SSE).
Fixes: 7ad816762f9b ("x86/fpu: Reset MXCSR to default in kernel_fpu_begin()")
...
Thx.