Thread (9 messages) 9 messages, 3 authors, 2021-08-23

Re: [PATCH v2] usb: chipidea: local_irq_save/restore added for hw_ep_prime

From: Sebastian Andrzej Siewior <bigeasy@linutronix.de>
Date: 2021-08-18 16:17:58
Also in: linux-rt-users, lkml

On 2021-08-17 18:53:13 [+0900], Jeaho Hwang wrote:
hw_ep_prime sometimes fails if irq occurs while it rus on RT kernel.
How/ why does it fail? Which IRQ occurs? Does it also occur without RT
and with threadirqs enabled?
local_irq_save/restore is added inside the function to gurantee atomicity.
only effective for preempt_rt since hw_ep_prime is called inside top half
or spin_lock_irqsave. No effect is expected for standard linux.
How is that helping?
#1 
  udc_irq() -> isr_tr_complete_handler() -> isr_tr_complete_low ->
   _hardware_dequeue() -> reprime_dtd() -> hw_ep_prime()

udc_irq() acquires ci->lock.

#2 
  ep_queue -> _ep_queue() ->_hardware_enqueue() -> hw_ep_prime()

ep_queue acquires hwep->lock. Which is actually ci->lock.

So if I read this right then hw_ep_prime() may not be interrupted in the
middle of its operation (but preempted) because each path is protected
by the lock.

isr_tr_complete_low() drops hwep->lock and acquires it again so it that
phase another thread may acquire it.
quoted hunk ↗ jump to hunk
Signed-off-by: Jeaho Hwang <redacted>
diff --git a/drivers/usb/chipidea/udc.c b/drivers/usb/chipidea/udc.c
index 8834ca613721..a624eddb3e22 100644
--- a/drivers/usb/chipidea/udc.c
+++ b/drivers/usb/chipidea/udc.c
@@ -191,22 +191,31 @@ static int hw_ep_get_halt(struct ci_hdrc *ci, int num, int dir)
 static int hw_ep_prime(struct ci_hdrc *ci, int num, int dir, int is_ctrl)
 {
 	int n = hw_ep_bit(num, dir);
+	unsigned long flags;
+	int ret = 0;
 
 	/* Synchronize before ep prime */
 	wmb();
 
-	if (is_ctrl && dir == RX && hw_read(ci, OP_ENDPTSETUPSTAT, BIT(num)))
+	/* irq affects this routine so irq should be disabled on RT.
+	 * on standard kernel, irq is already disabled by callers.
The important part is _how_ it is affected. If locking works then
nothing should read/ write the HW register. If the lock is briefly
dropped then another thread _may_ read/ write the registers but not
within this function.

If this function here is sensitive to timing (say the cpu_relax() loop
gets interrupt for 1ms) then it has to be documented as such.
+	 */
+	local_irq_save(flags);
+	if (is_ctrl && dir == RX && hw_read(ci, OP_ENDPTSETUPSTAT, BIT(num))) {
+		local_irq_restore(flags);
 		return -EAGAIN;
+	}
 
 	hw_write(ci, OP_ENDPTPRIME, ~0, BIT(n));
 
 	while (hw_read(ci, OP_ENDPTPRIME, BIT(n)))
 		cpu_relax();
 	if (is_ctrl && dir == RX && hw_read(ci, OP_ENDPTSETUPSTAT, BIT(num)))
-		return -EAGAIN;
+		ret = -EAGAIN;
 
+	local_irq_restore(flags);
 	/* status shoult be tested according with manual but it doesn't work */
-	return 0;
+	return ret;
 }
 
 /**
Sebastian
Keyboard shortcuts
hback out one level
jnext message in thread
kprevious message in thread
ldrill in
Escclose help / fold thread tree
?toggle this help