[ARM] Fix VFP initialisation issue for SMP systems
authorRussell King <rmk@dyn-67.arm.linux.org.uk>
Tue, 2 Jan 2007 23:40:30 +0000 (23:40 +0000)
committerRussell King <rmk+kernel@arm.linux.org.uk>
Tue, 2 Jan 2007 23:40:30 +0000 (23:40 +0000)
When we install the handlers for context switching, we must enable
VFP on all CPU cores, otherwise undefined (and random) effects
occur.

Signed-off-by: Russell King <rmk+kernel@arm.linux.org.uk>
arch/arm/vfp/vfpmodule.c

index e26cc1f599489f24274bbd0625b49bddc8bd31f4..490d9d18a7d1c166dbef16e988050ec870e406a8 100644 (file)
@@ -264,6 +264,18 @@ void VFP9_bounce(u32 trigger, u32 fpexc, struct pt_regs *regs)
                vfp_raise_exceptions(exceptions, trigger, orig_fpscr, regs);
 }
 
+static void vfp_enable(void *unused)
+{
+       u32 access = get_copro_access();
+
+       /*
+        * Enable full access to VFP (cp10 and cp11)
+        */
+       set_copro_access(access | CPACC_FULL(10) | CPACC_FULL(11));
+}
+
+#include <linux/smp.h>
+
 /*
  * VFP support code initialisation.
  */
@@ -288,6 +300,7 @@ static int __init vfp_init(void)
         * we just need to read the VFPSID register.
         */
        vfpsid = fmrx(FPSID);
+       barrier();
 
        printk(KERN_INFO "VFP support v0.3: ");
        if (VFP_arch) {
@@ -301,6 +314,8 @@ static int __init vfp_init(void)
        } else if (vfpsid & FPSID_NODOUBLE) {
                printk("no double precision support\n");
        } else {
+               smp_call_function(vfp_enable, NULL, 1, 1);
+
                VFP_arch = (vfpsid & FPSID_ARCH_MASK) >> FPSID_ARCH_BIT;  /* Extract the architecture version */
                printk("implementor %02x architecture %d part %02x variant %x rev %x\n",
                        (vfpsid & FPSID_IMPLEMENTER_MASK) >> FPSID_IMPLEMENTER_BIT,