On Fri, Feb 11, 2011 at 4:12 AM, Catalin Marinas
[off-list ref] wrote:

Colin,

On Thu, 2011-02-10 at 21:31 +0000, Colin Cross wrote:

quoted

+static int vfp_idle_notifier(struct notifier_block *self, unsigned long cmd,
+ ? ? ? void *v)
+{
+ ? ? ? u32 fpexc = fmrx(FPEXC);
+ ? ? ? unsigned int cpu = smp_processor_id();
+
+ ? ? ? if (cmd != CPU_PM_ENTER)
+ ? ? ? ? ? ? ? return NOTIFY_OK;
+
+ ? ? ? /* The VFP may be reset in idle, save the state */
+ ? ? ? if ((fpexc & FPEXC_EN) && last_VFP_context[cpu]) {
+ ? ? ? ? ? ? ? vfp_save_state(last_VFP_context[cpu], fpexc);
+ ? ? ? ? ? ? ? last_VFP_context[cpu]->hard.cpu = cpu;
+ ? ? ? }

Should we only handle the case where the VFP is enabled? At context
switch we disable the VFP and re-enable it when an application tries to
use it but it will remain disabled even the application hasn't used the
VFP. So switching to the idle thread would cause the VFP to be disabled
but the state not necessarily saved.

Right

On SMP systems, we save the VFP at every context switch to deal with the
thread migration (though I have a plan to make this lazily on SMP as
well). On UP however, we don't save the VFP registers at context switch,
we just disable it and save it lazily if used later in a different task

Something like below (untested):

? ? ? ?if (last_VFP_context[cpu]) {
? ? ? ? ? ? ? ?vfp_save_state(last_VFP_context[cpu], fpexc);
? ? ? ? ? ? ? ?/* force a reload when coming back from idle */
? ? ? ? ? ? ? ?last_VFP_context[cpu] = NULL;
? ? ? ? ? ? ? ?fmxr(FPEXC, fpexc & ~FPEXC_EN);
? ? ? ?}

The last line (disabling) may not be necessary if we know that it comes
back from idle as disabled.

It shouldn't be necessary, the context switch into the idle thread
should have disabled it, but it doesn't hurt.  We should also disable
it when exiting idle.

I wonder whether the current vfp_pm_suspend() function needs fixing for
UP systems as well. It is find if the hardware preserves the VFP
registers (which may not be the case).

I think there is a case where the VFP registers can be lost in suspend
on UP platforms that don't save the VFP registers in their platform
suspend.  If a thread is using the VFP, and then context switches to a
thread that does not use VFP but triggers suspend by writing to
/sys/power/state, vfp_pm_suspend will be called with the VFP disabled
but the registers not saved.  I think this would work:

	/* save state for resumption */
	if (last_VFP_context[ti->cpu]) {
		printk(KERN_DEBUG "%s: saving vfp state\n", __func__);
		vfp_save_state(last_VFP_context[ti->cpu], fpexc);

		/* disable, just in case */
		fmxr(FPEXC, fpexc & ~FPEXC_EN);
	}

If the thread that wrote to /sys/power/state is using VFP,
last_VFP_context will be the same as ti->vfpstate, so we can always
save last_VFP_context.