>>>>> On Wed, 27 Apr 2005 14:36:22 +0900 (JST), Atsushi Nemoto
>>>>> <anemo@mba.ocn.ne.jp> said:
anemo> With this patch, whole fpu-emulator can be run without
anemo> disabling preempt. I will post a patch to fix preemption issue
anemo> soon.
Here it is.
It also fixes these issues:
* The process might lose fpu BEFORE calling preempt_disable() in
do_fpe().
* The saved fp context might be overwritten if another process took
fpu.
--- linux-mips/arch/mips/kernel/traps.c 2005-04-18 11:42:47.000000000 +0900
+++ linux/arch/mips/kernel/traps.c 2005-04-27 14:04:53.407006244 +0900
@@ -551,6 +551,14 @@
preempt_disable();
+#ifdef CONFIG_PREEMPT
+ if (!is_fpu_owner()) {
+ /* We might lose fpu before disabling preempt... */
+ own_fpu();
+ BUG_ON(!used_math());
+ restore_fp(current);
+ }
+#endif
/*
* Unimplemented operation exception. If we've got the full
* software emulator on-board, let's use it...
@@ -562,11 +570,18 @@
* a bit extreme for what should be an infrequent event.
*/
save_fp(current);
+ /* Ensure 'resume' not overwrite saved fp context again. */
+ lose_fpu();
+
+ preempt_enable();
/* Run the emulator */
sig = fpu_emulator_cop1Handler (0, regs,
¤t->thread.fpu.soft);
+ preempt_disable();
+
+ own_fpu(); /* Using the FPU again. */
/*
* We can't allow the emulated instruction to leave any of
* the cause bit set in $fcr31.
@@ -712,6 +727,8 @@
set_used_math();
}
+ preempt_enable();
+
if (!cpu_has_fpu) {
int sig = fpu_emulator_cop1Handler(0, regs,
¤t->thread.fpu.soft);
@@ -719,8 +736,6 @@
force_sig(sig, current);
}
- preempt_enable();
-
return;
case 2:
|