On Mon, 2012-06-18 at 15:42 +0530, Jassi Brar wrote:

On 18 June 2012 13:41, Tomi Valkeinen [off-list ref] wrote:

quoted

quoted

Explicitly maintaining HDMI phy power state using a flag is prone to
race and un-necessary when we have a zero-cost alternative of checking
the state before trying to set it.

Why would reading the value from the register be any less racy than
keeping it in memory?

Racy in the sense that h/w doesn't always hop states according to what
a "state" variable would expect it to.

But I don't think the status register is any better. If I'm not
mistaken, when you set the phy pwr to, say, TXON, the status register
will still show the old value. The status register will be right only
after the HW is actually at TXON state.

In this case the hdmi_set_phy_pwr() function will anyway wait until the
HW has changed states, so both approaches should work just fine.

Also in this case, phy_tx_enabled modification is unprotected in
ti_hdmi_4xxx_phy_disable().

So is hdmi_set_phy_pwr(). Note that I'm not saying the current approach
is not racy, but your patch doesn't make it any less racy.

BTW, coming to think about it, I am not sure what we need the
spin_lock_irqsave() protection for in hdmi_check_hpd_state() ?  It
can't control HPD gpio state change and hdmi_set_phy_pwr() seems too
expensive and is already unprotected elsewhere.

It's needed when enabling the hdmi output. In phy_enable() the irq is
requested first, and then the phy_enable() runs hdmi_check_hpd_state().
So there's a chance to run hdmi_check_hpd_state() from both
hpd-interrupt and phy_enable() at the same time.

The hdmi_set_phy_pwr() is not called in many places, but I think there's
indeed a problem there. It is called after free_irq(), but I think
(guess) the irq handler could still be running after free_irq. So those
should be protected by the same spinlock too.

 Tomi