Skip to contentSkip to conversation

w1: Disable irqs in critical section

RFC In review Jan Weitzel 3 participants 6 messages Sep 7, 2011 – Sep 14, 2011 W1 dallas's 1-wire bus

lore
Share & tools
lore
#0 lore raw

RFC 1/1 w1: Disable irqs in critical section

Interrupting w1_delay in w1_read_bit results in missing the low level on the w1 line and receiving "1" instead of "0".
Adding local_irq_save / local_irq_restore around the critical section

---
 drivers/w1/w1_io.c |    5 +++++
 1 files changed, 5 insertions(+), 0 deletions(-)
drivers/w1/w1_io.c +5 -0
index 8e8b64c..a059eaa 100644
@@ -158,13 +158,18 @@ EXPORT_SYMBOL_GPL(w1_write_8);
 static u8 w1_read_bit(struct w1_master *dev)
 {
 	int result;
+	unsigned long flags;
 
+	/* sample timing is critical here */
+	local_irq_save(flags);
 	dev->bus_master->write_bit(dev->bus_master->data, 0);
 	w1_delay(6);
 	dev->bus_master->write_bit(dev->bus_master->data, 1);
 	w1_delay(9);
 
 	result = dev->bus_master->read_bit(dev->bus_master->data);
+	local_irq_restore(flags);
+
 	w1_delay(55);
 
 	return result & 0x1;
Signature · 1 line
1.7.0.4
#1 lore raw
Quoted 3 lines · Jan Weitzel

Interrupting w1_delay in w1_read_bit results in missing the low level on the w1 line and receiving "1" instead of "0".
Adding local_irq_save / local_irq_restore around the critical section

This means that CPU will be essentially stuck for 15 useconds for every bit transferred, doesn't really look like a good idea.

Are you absolutely sure that missing bit is because of timings and not some other bug?

Signature · 1 line
	Evgeniy Polyakov
#2 lore raw

Am Mittwoch, den 07.09.2011, 21:50 +0400 schrieb Evgeniy Polyakov:

Quoted 11 lines · Jan Weitzel
Quoted 3 lines

Interrupting w1_delay in w1_read_bit results in missing the low level on the w1 line and receiving "1" instead of "0".
Adding local_irq_save / local_irq_restore around the critical section

This means that CPU will be essentially stuck for 15 useconds for every bit transferred, doesn't really look like a good idea.

Are you absolutely sure that missing bit is because of timings and not some other bug?

I trigger a gpio line after the samplepoint. I case of a wrong bit the sample is taken after the "0 gap". The cycle time (samplt to sample)is increased form about 80µs to 95µs.
I did the measurement with the w1-gpio driver on a OMAP4 board.

Jan

#3 lore raw
Quoted 9 lines · On Thu, 08 Sep 2011 07:46:21 +0200 Jan Weitzel <[email protected]> wrote:

Am Mittwoch, den 07.09.2011, 21:50 +0400 schrieb Evgeniy Polyakov:

Quoted 8 lines · Jan Weitzel
Quoted 3 lines

Interrupting w1_delay in w1_read_bit results in missing the low level on the w1 line and receiving "1" instead of "0".
Adding local_irq_save / local_irq_restore around the critical section

This means that CPU will be essentially stuck for 15 useconds for every bit transferred, doesn't really look like a good idea.

Quoted 9 lines
Quoted 3 lines

Are you absolutely sure that missing bit is because of timings and not some other bug?

I trigger a gpio line after the samplepoint. I case of a wrong bit the sample is taken after the "0 gap". The cycle time (samplt to sample)is increased form about 80__s to 95__s.
I did the measurement with the w1-gpio driver on a OMAP4 board.

I'm not clear on how w1 actually works. Is it a bit-banging protocol in which the timing is provided by the host CPU? If so then yes, we should carefully disable interrupts in places where an interrupt would disrupt critical timing. (But what to do about NMIs and SMIs?)

Disabling interrupts for 15us is pretty obnoxious, but there's a 55us delay there with interrupts enabled, so the overall effect shouldn't be too bad.

Finally, can we fine-tune the interrupt-disabled section a bit? For example, can the local_irq_disable() be moved to after the write_bit(..., 0)?

#4 lore raw

Am Dienstag, den 13.09.2011, 15:41 -0700 schrieb Andrew Morton:

Quoted 16 lines
Quoted 13 lines · On Thu, 08 Sep 2011 07:46:21 +0200 Jan Weitzel <[email protected]> wrote:

Am Mittwoch, den 07.09.2011, 21:50 +0400 schrieb Evgeniy Polyakov:

Quoted 11 lines · Jan Weitzel
Quoted 3 lines

Interrupting w1_delay in w1_read_bit results in missing the low level on the w1 line and receiving "1" instead of "0".
Adding local_irq_save / local_irq_restore around the critical section

This means that CPU will be essentially stuck for 15 useconds for every bit transferred, doesn't really look like a good idea.

Are you absolutely sure that missing bit is because of timings and not some other bug?

Quoted 8 lines
Quoted 5 lines

I trigger a gpio line after the samplepoint. I case of a wrong bit the sample is taken after the "0 gap". The cycle time (samplt to sample)is increased form about 80__s to 95__s.
I did the measurement with the w1-gpio driver on a OMAP4 board.

I'm not clear on how w1 actually works. Is it a bit-banging protocol in which the timing is provided by the host CPU? If so then

correct.

Quoted 3 lines

yes, we should carefully disable interrupts in places where an interrupt would disrupt critical timing. (But what to do about NMIs and SMIs?)

fortunately on our embedded platform we have no SMIs

Quoted 8 lines

Disabling interrupts for 15us is pretty obnoxious, but there's a 55us delay there with interrupts enabled, so the overall effect shouldn't be too bad.

Finally, can we fine-tune the interrupt-disabled section a bit? For example, can the local_irq_disable() be moved to after the write_bit(..., 0)?

The falling edge from 1 to 0 starts timing for the slave. After the 6µs the slave should drive the line low (In case the information bit is "0"). So there is no additional edge on the line. If there is an interrupt and the low pulse from master is longer, slave did't see it. Master need to generate the 1 -> 0 transition and measure after 15µs

I tested it anyhow. Behaviour is slightly better than without disabling interrupts. Only disabling it for 15µs fix the problem.

Jan

#5 lore raw
Quoted 4 lines · Jan Weitzel
Quoted 4 lines

Finally, can we fine-tune the interrupt-disabled section a bit? For example, can the local_irq_disable() be moved to after the write_bit(..., 0)?

Quoted 8 lines

The falling edge from 1 to 0 starts timing for the slave. After the 6µs the slave should drive the line low (In case the information bit is "0"). So there is no additional edge on the line. If there is an interrupt and the low pulse from master is longer, slave did't see it. Master need to generate the 1 -> 0 transition and measure after 15µs

I tested it anyhow. Behaviour is slightly better than without disabling interrupts. Only disabling it for 15µs fix the problem.

Sad story, but if it is the only solution, then we should go with it.

Andrew, please push this patch upstream.

Thank you Jan.

Signature · 1 line
	Evgeniy Polyakov