[PATCH 2/4] x86/mmx: Use KFPU_MMX for MMX string operations

[Andy Lutomirski]

The default kernel_fpu_begin() doesn't work on systems that support XMM but
haven't yet enabled OSFXSR.  This causes crashes when _mmx_memcpy() is
called too early.

Fix it by using kernel_fpu_begin(KFPU_MMX) explicitly.  This should also be
faster, since it skips both the reasonably fast LDMXCSR and also the rather
slow FNINIT instructions.

Fixes: 7ad816762f9b ("x86/fpu: Reset MXCSR to default in kernel_fpu_begin()")
Reported-by: Krzysztof Mazur <krzysiek@xxxxxxxxxxxx>
Signed-off-by: Andy Lutomirski <luto@xxxxxxxxxx>
---
 arch/x86/lib/mmx_32.c | 10 +++++-----
 1 file changed, 5 insertions(+), 5 deletions(-)

diff --git a/arch/x86/lib/mmx_32.c b/arch/x86/lib/mmx_32.c
index 4321fa02e18d..daa80fa005fb 100644
--- a/arch/x86/lib/mmx_32.c
+++ b/arch/x86/lib/mmx_32.c
@@ -37,7 +37,7 @@ void *_mmx_memcpy(void *to, const void *from, size_t len)
 	p = to;
 	i = len >> 6; /* len/64 */
 
-	kernel_fpu_begin();
+	kernel_fpu_begin_mask(KFPU_MMX);
 
 	__asm__ __volatile__ (
 		"1: prefetch (%0)\n"		/* This set is 28 bytes */
@@ -127,7 +127,7 @@ static void fast_clear_page(void *page)
 {
 	int i;
 
-	kernel_fpu_begin();
+	kernel_fpu_begin_mask(KFPU_MMX);
 
 	__asm__ __volatile__ (
 		"  pxor %%mm0, %%mm0\n" : :
@@ -160,7 +160,7 @@ static void fast_copy_page(void *to, void *from)
 {
 	int i;
 
-	kernel_fpu_begin();
+	kernel_fpu_begin_mask(KFPU_MMX);
 
 	/*
 	 * maybe the prefetch stuff can go before the expensive fnsave...
@@ -247,7 +247,7 @@ static void fast_clear_page(void *page)
 {
 	int i;
 
-	kernel_fpu_begin();
+	kernel_fpu_begin_mask(KFPU_MMX);
 
 	__asm__ __volatile__ (
 		"  pxor %%mm0, %%mm0\n" : :
@@ -282,7 +282,7 @@ static void fast_copy_page(void *to, void *from)
 {
 	int i;
 
-	kernel_fpu_begin();
+	kernel_fpu_begin_mask(KFPU_MMX);
 
 	__asm__ __volatile__ (
 		"1: prefetch (%0)\n"
-- 
2.29.2