nanosleep() granularity bumps up in 2.5.64 (was: [PATCH] settimeofday() not synchronised with gettimeofday())
Quoted 1 line · David Mosberger
Are you running ntp?
Yes (I hadn't noticed it) but it was not connected to any server and disabling it doesn't change the results.
Quoted 13 lines
On 2.5:
$ time sleep 16 real 0m16.002s user 0m0.001s sys 0m0.002s$ time sleep 16.02 real 0m25.189s user 0m0.000s sys 0m0.001s
Quoted 4 lines
So clearly something very strange is going on. My suspicion is that the bug was introduced back when x86 switched from 100Hz to 1000Hz ticks, but that's just a guess. Eric, would you be able/willing/interested to look into this?
Sure, I aim at porting the high resolution timers but any annoying bug related to the time can be interesting to remove.
Coincidently Vita has just reported a bug on the lkml which, after a closer look, seems to be the same:
Quoted 5 lines
When playing with
select()timeout values I found that granularity
ofnanosleep()in 2.5.64 kernel bumps to 256 msec. Trying to get finer granularity it ends up sleeping to the next multiple of 256 msecFrom what I understand their is a bug in the timers that causes a big
granularity. The case of Vita is a very good example. Also, after 16s it seems the granuality (slowly?!) jumps from 1/64th s to 16s! :
sleep requested time obtained
14.000000000 14.006201744 15.000000000 15.006647110 16.000000000 16.007089615 17.000000000 18.742679596 18.000000000 32.014190674 19.000000000 32.014190674 20.000000000 32.014190674
I think lines like that from patch-2.5.64 are very suspicious to be related to the bug:
+ base->timer_jiffies = INITIAL_JIFFIES;
+ base->tv1.index = INITIAL_JIFFIES & TVR_MASK;
+ base->tv2.index = (INITIAL_JIFFIES >> TVR_BITS) & TVN_MASK;
+ base->tv3.index = (INITIAL_JIFFIES >> (TVR_BITS+TVN_BITS)) & TVN_MASK;
+ base->tv4.index = (INITIAL_JIFFIES >> (TVR_BITS+2*TVN_BITS)) &TVN_MASK;
+ base->tv5.index = (INITIAL_JIFFIES >> (TVR_BITS+3*TVN_BITS)) &TVN_MASK;
Any idea/sugestion/patch is welcomed. Whatever, I will try to fix this as soon as I'm back from my week end :-)
Eric
Quoted 10 lines · Eric Piel
I think lines like that from patch-2.5.64 are very suspicious to be related to the bug:
(quoted diff) +5
-0+ base->timer_jiffies = INITIAL_JIFFIES; + base->tv1.index = INITIAL_JIFFIES & TVR_MASK; + base->tv2.index = (INITIAL_JIFFIES >> TVR_BITS) & TVN_MASK; + base->tv3.index = (INITIAL_JIFFIES >> (TVR_BITS+TVN_BITS)) & TVN_MASK; + base->tv4.index = (INITIAL_JIFFIES >> (TVR_BITS+2*TVN_BITS)) &
TVN_MASK;(quoted diff) +1
-0+ base->tv5.index = (INITIAL_JIFFIES >> (TVR_BITS+3*TVN_BITS)) &
TVN_MASK;
No, I don't think so. Those lines are for starting `jiffies' at a very high number so we spot jiffie-wrap bugs early on.
Signature · 3 lines
"It's not Hollywood. War is real, war is primarily not about defeat or victory, it is about death. I've seen thousands and thousands of dead bodies. Do you think I want to have an academic debate on this subject?" -- Robert Fisk
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line · Eric Piel
Eric> Sure, I aim at porting the high resolution timers but any Eric> annoying bug related to the time can be interesting to remove.
Great! It would be great to have someone who can focus on that.
Eric> Any idea/sugestion/patch is welcomed. Whatever, I will try to Eric> fix this as soon as I'm back from my week end :-)
Does the x86 show the same behavior? That would be a useful start to narrow down the problem.
--david
Quoted 14 lines · Matthew Wilcox
Quoted 10 lines · Eric Piel
I think lines like that from patch-2.5.64 are very suspicious to be related to the bug:
(quoted diff) +5
-0+ base->timer_jiffies = INITIAL_JIFFIES; + base->tv1.index = INITIAL_JIFFIES & TVR_MASK; + base->tv2.index = (INITIAL_JIFFIES >> TVR_BITS) & TVN_MASK; + base->tv3.index = (INITIAL_JIFFIES >> (TVR_BITS+TVN_BITS)) & TVN_MASK; + base->tv4.index = (INITIAL_JIFFIES >> (TVR_BITS+2*TVN_BITS)) &
TVN_MASK;(quoted diff) +1
-0+ base->tv5.index = (INITIAL_JIFFIES >> (TVR_BITS+3*TVN_BITS)) &
TVN_MASK;No, I don't think so. Those lines are for starting `jiffies' at a very high number so we spot jiffie-wrap bugs early on.
The nanosleep() bug narrowed down to 2.5.63-bk2. That's version, the "initial
jiffies" patch went in. And yes, it's on i686 machine.
Cheers,
Vita
Re: [BUG] nanosleep() granularity bumps up in 2.5.64
Quoted 16 lines · Vitezslav Samel
Quoted 14 lines · Matthew Wilcox
Quoted 10 lines · Eric Piel
I think lines like that from patch-2.5.64 are very suspicious to be related to the bug:
(quoted diff) +5
-0+ base->timer_jiffies = INITIAL_JIFFIES; + base->tv1.index = INITIAL_JIFFIES & TVR_MASK; + base->tv2.index = (INITIAL_JIFFIES >> TVR_BITS) & TVN_MASK; + base->tv3.index = (INITIAL_JIFFIES >> (TVR_BITS+TVN_BITS)) & TVN_MASK; + base->tv4.index = (INITIAL_JIFFIES >> (TVR_BITS+2*TVN_BITS)) &
TVN_MASK;(quoted diff) +1
-0+ base->tv5.index = (INITIAL_JIFFIES >> (TVR_BITS+3*TVN_BITS)) &
TVN_MASK;No, I don't think so. Those lines are for starting `jiffies' at a very high number so we spot jiffie-wrap bugs early on.
Quoted 2 lines
The
nanosleep()bug narrowed down to 2.5.63-bk2. That's version, the "initial
jiffies" patch went in. And yes, it's on i686 machine.
You can easily check whether it's connected with this change by setting INITIAL_JIFFIES to zero. This should exactly recover the previous situation.
I.e., something like the following (untested, hand-crafted) patch:
include/linux/time.h +1 -1
@@ -28,7 +28,7 @@* Have the 32 bit jiffies value wrap 5 minutes after boot * so jiffies wrap bugs show up earlier. */ -#define INITIAL_JIFFIES ((unsigned int) (-300*HZ)) +#define INITIAL_JIFFIES 0/* * Change timeval to jiffies, trying to avoid the
Tim
[BUG & WORKAROUND] nanosleep() granularity bumps up in 2.5.64
Quoted 8 lines · Tim Schmielau
Quoted 2 lines · Vitezslav Samel
The
nanosleep()bug narrowed down to 2.5.63-bk2. That's version, the "initial
jiffies" patch went in. And yes, it's on i686 machine.You can easily check whether it's connected with this change by setting
INITIAL_JIFFIESto zero. This should exactly recover the previous situation.
OK. I've done the test myself and I plead guilty. As a temporary workaround you can apply the following patch:
include/linux/time.h renamed from include/linux/time.h.orig +1 -1
@@ -31,7 +31,7 @@* Have the 32 bit jiffies value wrap 5 minutes after boot * so jiffies wrap bugs show up earlier. */ -#define INITIAL_JIFFIES ((unsigned int) (-300*HZ)) +#define INITIAL_JIFFIES 0/* * Change timeval to jiffies, trying to avoid the
Still, after half an hour of glancing at the code I can't see my mistake. I've re-checked that the problem does not occur with the original "initial jiffies" patch for 2.4. So I must have missed a (subtle?) difference between 2.4 and 2.5 when I did the forward-port.
Sorry,
Tim
PATCH 1/1 fix nanosleep() granularity bumps
Quoted 2 lines · Tim Schmielau
I've re-checked that the problem does not occur with the original "initial jiffies" patch for 2.4.
Stupid me - 2.4 with the patch has the same problem, it just takes 10x as long to show up, a range that was not covered by the testcase.
The actual problem is that I didn't fully understand the, well, 'elaborated' way of counting jiffies in the timer cascade:
When starting out with timer_jiffies=0, the timer cascade is (unneccessarily) triggered on the first timer interrupt, incrementing all the higher indices.
When starting with any other initial jiffies value, we miss that and end up with all higher indices being off by one.
Fix is as follows:
kernel/timer.c +8 -4
@@ -1183,10 +1183,14 @@ base->timer_jiffies =INITIAL_JIFFIES ; base->tv1.index =INITIAL_JIFFIES &TVR_MASK ; - base->tv2.index = (INITIAL_JIFFIES >>TVR_BITS ) &TVN_MASK ; - base->tv3.index = (INITIAL_JIFFIES >> (TVR_BITS +TVN_BITS )) &TVN_MASK ; - base->tv4.index = (INITIAL_JIFFIES >> (TVR_BITS +2 *TVN_BITS )) &TVN_MASK ; - base->tv5.index = (INITIAL_JIFFIES >> (TVR_BITS +3 *TVN_BITS )) &TVN_MASK ; + base->tv2.index = ((INITIAL_JIFFIES >>TVR_BITS ) + + (INITIAL_JIFFIES !=0 )) &TVN_MASK ; + base->tv3.index = ((INITIAL_JIFFIES >> (TVR_BITS +TVN_BITS )) + + (INITIAL_JIFFIES !=0 )) &TVN_MASK ; + base->tv4.index = ((INITIAL_JIFFIES >> (TVR_BITS +2 *TVN_BITS )) + + (INITIAL_JIFFIES !=0 )) &TVN_MASK ; + base->tv5.index = ((INITIAL_JIFFIES >> (TVR_BITS +3 *TVN_BITS )) + + (INITIAL_JIFFIES !=0 )) &TVN_MASK ; } static int __devinit timer_cpu_notify(struct notifier_block *self,
Sorry for all the trouble,
Tim
Re: [PATCH] fix nanosleep() granularity bumps
Quoted 9 lines · Tim Schmielau
Quoted 2 lines · Tim Schmielau
I've re-checked that the problem does not occur with the original "initial jiffies" patch for 2.4.
Stupid me - 2.4 with the patch has the same problem, it just takes 10x as long to show up, a range that was not covered by the testcase.
Quoted 8 lines
The actual problem is that I didn't fully understand the, well, 'elaborated' way of counting jiffies in the timer cascade:
When starting out with timer_jiffies=0, the timer cascade is (unneccessarily) triggered on the first timer interrupt, incrementing all the higher indices.
When starting with any other initial jiffies value, we miss that and end up with all higher indices being off by one.
Suggest the attached patch as a fix instead - easier to understand I think and works for every possible start value. This is what I made for Andrea Arcangeli many years ago...
kernel/timer.c +8 -5
@@ -1182,11 +1182,14 @@INIT_LIST_HEAD (base->tv1.vec + j); base->timer_jiffies =INITIAL_JIFFIES ; - base->tv1.index =INITIAL_JIFFIES &TVR_MASK ; - base->tv2.index = (INITIAL_JIFFIES >>TVR_BITS ) &TVN_MASK ; - base->tv3.index = (INITIAL_JIFFIES >> (TVR_BITS +TVN_BITS )) &TVN_MASK ; - base->tv4.index = (INITIAL_JIFFIES >> (TVR_BITS +2 *TVN_BITS )) &TVN_MASK ; - base->tv5.index = (INITIAL_JIFFIES >> (TVR_BITS +3 *TVN_BITS )) &TVN_MASK ; + base->tv1.index = (1 + (INITIAL_JIFFIES -1 )) &TVR_MASK ; + base->tv2.index = (1 + ((INITIAL_JIFFIES -1 ) >>TVR_BITS )) &TVN_MASK ; + base->tv3.index = (1 + ((INITIAL_JIFFIES -1 ) + >> (TVR_BITS +TVN_BITS ))) &TVN_MASK ; + base->tv4.index = (1 + ((INITIAL_JIFFIES -1 ) + >> (TVR_BITS +2 *TVN_BITS ))) &TVN_MASK ; + base->tv5.index = (1 + ((INITIAL_JIFFIES -1 ) + >> (TVR_BITS +3 *TVN_BITS ))) &TVN_MASK ; } static int __devinit timer_cpu_notify(struct notifier_block *self,
- Finn Arne
Re: [PATCH] fix nanosleep() granularity bumps
Quoted 1 line · Tim Schmielau
[...]
Quoted 17 lines
Suggest the attached patch as a fix instead - easier to understand I think and works for every possible start value. This is what I made for Andrea Arcangeli many years ago...
diff -ur
linux-2.5.65/kernel/timer.clinux-2.5.65-new/kernel/timer.ckernel/timer.c +1
-5@@ -1182,11 +1182,14 @@INIT_LIST_HEAD (base->tv1.vec + j); base->timer_jiffies =INITIAL_JIFFIES ; - base->tv1.index =INITIAL_JIFFIES &TVR_MASK ; - base->tv2.index = (INITIAL_JIFFIES >>TVR_BITS ) &TVN_MASK ; - base->tv3.index = (INITIAL_JIFFIES >> (TVR_BITS +TVN_BITS )) &TVN_MASK ; - base->tv4.index = (INITIAL_JIFFIES >> (TVR_BITS +2 *TVN_BITS )) &TVN_MASK ; - base->tv5.index = (INITIAL_JIFFIES >> (TVR_BITS +3 *TVN_BITS )) &TVN_MASK ; + base->tv1.index = (1 + (INITIAL_JIFFIES -1 )) &TVR_MASK ;
kernel/timer.c line 1195
Quoted 10 lines
kernel/timer.c +7
-0+ base->tv2.index = (1 + ((INITIAL_JIFFIES -1 ) >>TVR_BITS )) &TVN_MASK ; + base->tv3.index = (1 + ((INITIAL_JIFFIES -1 ) + >> (TVR_BITS +TVN_BITS ))) &TVN_MASK ; + base->tv4.index = (1 + ((INITIAL_JIFFIES -1 ) + >> (TVR_BITS +2 *TVN_BITS ))) &TVN_MASK ; + base->tv5.index = (1 + ((INITIAL_JIFFIES -1 ) + >> (TVR_BITS +3 *TVN_BITS ))) &TVN_MASK ; } static int __devinit timer_cpu_notify(struct notifier_block *self,
Too late - the whole tv[1-5].index duplication is gone already after a cleanup done by George Anzinger.
Tim