Skip to contentSkip to conversation

Add explanation of udelay() inaccuracy

PATCH Applied Russell King 2 participants 2 messages Jan 19, 2017 – Feb 16, 2017 High-resolution timers, timer wheel, clockevents

lore
Share & tools
lore
#0 lore raw

PATCH 1/1 Add explanation of udelay() inaccuracy

There seems to be some misunderstanding that udelay() and friends will always guarantee the specified delay. This is a false understanding. When udelay() is based on CPU cycles, it can return early for many reasons which are detailed by Linus' reply to me in a thread in 2011:

http://lists.openwall.net/linux-kernel/2011/01/12/372

However, a udelay test module was created in 2014 which allows udelay() to only be 0.5% fast, which is outside of the CPU-cycles udelay() results I measured back in 2011, which were deemed to be in the "we don't care" region.

test_udelay() should be fixed to reflect the real allowable tolerance on udelay(), rather than 0.5%.

Cc David Riley <[email protected]>
Cc John Stultz <[email protected]>

---
The issue of "udelay being too slow" came up towards the end of last year again, caused again by people believing that udelay() should provide a delay of at least the requested value. We need to make sure that the inaccuracies in this interface are well documented and understood, and that we don't lead people down the path of believing (through the addition of this test_udelay code) that udelay() is more accurate than it really is. 0.5% is way _too_ tight a specification, and as timer interrupt handling gets heavier, so the inaccuracy in the CPU cycles based udelay() gets bigger.

 include/linux/delay.h | 11 +++++++++++
 1 file changed, 11 insertions(+)
include/linux/delay.h +11 -0
index a6ecb34cf547..2ecb3c46b20a 100644
@@ -5,6 +5,17 @@
  * Copyright (C) 1993 Linus Torvalds
  *
  * Delay routines, using a pre-computed "loops_per_jiffy" value.
+ *
+ * Please note that ndelay(), udelay() and mdelay() may return early for
+ * several reasons:
+ *  1. computed loops_per_jiffy too low (due to the time taken to
+ *     execute the timer interrupt.)
+ *  2. cache behaviour affecting the time it takes to execute the
+ *     loop function.
+ *  3. CPU clock rate changes.
+ *
+ * Please see this thread:
+ *   http://lists.openwall.net/linux-kernel/2011/01/09/56
  */
 
 #include <linux/kernel.h>
Signature · 1 line
2.7.4
#1 lore raw

Hi!

Quoted 4 lines

+++ b/include/linux/delay.h

include/linux/delay.h +0 -0
@@ -5,6 +5,17 @@
  * Copyright (C) 1993 Linus Torvalds
  *

include/linux/delay.h line 16

Quoted 10 lines
include/linux/delay.h +9 -0
  * Delay routines, using a pre-computed "loops_per_jiffy" value.
+ *
+ * Please note that ndelay(), udelay() and mdelay() may return early for
+ * several reasons:
+ *  1. computed loops_per_jiffy too low (due to the time taken to
+ *     execute the timer interrupt.)
+ *  2. cache behaviour affecting the time it takes to execute the
+ *     loop function.
+ *  3. CPU clock rate changes.
+ *

Hmm. Formulated like this, it would mean that udelay(100) can return in 10usec (because of clock rate changes). No way can drivers work reliably in that case.

Can we formulate something more useful? We don't want driver writers to delay 10 times more "just for cpufreq", right?

Pavel

Signature · 2 lines
(english) http://www.livejournal.com/~pavelmachek
(cesky, pictures) http://atrey.karlin.mff.cuni.cz/~pavel/picture/horses/blog.html