[hrtimers] Re: Regression affecting b43 LP-PHY card

[hrtimers] Re: Regression affecting b43 LP-PHY card

[Rafał Miłecki]

W dniu 10 maja 2011 00:52 użytkownik Rafał Miłecki <zajec5@gmail.com> napisał:
> Juan owns Lenovo affected by well-known LP-PHY DMA errors. His testing
> procedure is following:
> modprobe wl; connect; download sth small; rmmod wl;
> modprobe b43; download 2GB
>
> When working on DMA errors we discovered that wireless-testing is not
> working well for him. Even after performing above procedure his
> machine disconnects quickly and he is not able to reconnect. We tested
> 2.6.39-rc6 from tarball and it was working fine. I'd like to highlight
> here, that we were switching between mainline and wireless-testing few
> times. It is not a random issue.
>
> I suspected this regression could be caused by my recent ssb patches.
> So I reverted all of them but this didn't help.
>
> In this situation we decided to bisect. I was a little afraid of last
> merges so we took older 2.6.38 as GOOD (we tested this twice) and
> wireless-testing commit before my ssb changes as BAD. Today Juan
> finished bisecting kernel:
> http://pastebin.com/HSKbRzpB
>
> According to his bisection the first bad commit is
> e06383db9ec591696a06654257474b85bac1f8cb [0]:
> hrtimers: extend hrtimer base code to handle more then 2 clockids
>
> Does it make any sense to you? Could this be some timing issue?
>
> It was too late to test this today, we (Juan) will work on this
> tomorrow. It's impossible to revert this commit from HEAD of
> wireless-testing, so my idea is to checkout commit, test, revert,
> test.
>
> Did anyone else experience any similar problems with latest wireless-testing?
>
>
> http://git.kernel.org/?p=linux/kernel/git/torvalds/linux-2.6.git;a=commit;h=e06383db9ec591696a06654257474b85bac1f8cb

Today Juan checkouted commit e06383db9ec591696a06654257474b85bac1f8cb
and tested it. He was disconnected really soon.

Then he reverted e06383db9ec591696a06654257474b85bac1f8cb and tested
again. Connection was stable, he downloaded 2GB file over network.


John S.: your commit does not touch Broadcom card directly, but it
seems it somehow affects it. I suspect there can be some timing issue.
Do you have any idea what could it be, how can we debug this?

-- 
Rafał

Re: [hrtimers] Re: Regression affecting b43 LP-PHY card

[John Stultz]

On Tue, 2011-05-10 at 20:57 +0200, Rafał Miłecki wrote:
> W dniu 10 maja 2011 00:52 użytkownik Rafał Miłecki <zajec5@gmail.com> napisał:
> > In this situation we decided to bisect. I was a little afraid of last
> > merges so we took older 2.6.38 as GOOD (we tested this twice) and
> > wireless-testing commit before my ssb changes as BAD. Today Juan
> > finished bisecting kernel:
> > http://pastebin.com/HSKbRzpB
> >
> > According to his bisection the first bad commit is
> > e06383db9ec591696a06654257474b85bac1f8cb [0]:
> > hrtimers: extend hrtimer base code to handle more then 2 clockids
> >
> > Does it make any sense to you? Could this be some timing issue?
> >
> > It was too late to test this today, we (Juan) will work on this
> > tomorrow. It's impossible to revert this commit from HEAD of
> > wireless-testing, so my idea is to checkout commit, test, revert,
> > test.
> >
> > Did anyone else experience any similar problems with latest wireless-testing?
> >
> >
> > http://git.kernel.org/?p=linux/kernel/git/torvalds/linux-2.6.git;a=commit;h=e06383db9ec591696a06654257474b85bac1f8cb
> 
> Today Juan checkouted commit e06383db9ec591696a06654257474b85bac1f8cb
> and tested it. He was disconnected really soon.
> 
> Then he reverted e06383db9ec591696a06654257474b85bac1f8cb and tested
> again. Connection was stable, he downloaded 2GB file over network.
> 
> 
> John S.: your commit does not touch Broadcom card directly, but it
> seems it somehow affects it. I suspect there can be some timing issue.
> Do you have any idea what could it be, how can we debug this?

Sorry for the trouble!

My commit exposed a few spots where hrtimers were being initialized
before hrtimer_init is called, which caused problems. Thomas provided a
solution that makes such behavior still function ok:
http://git.kernel.org/?p=linux/kernel/git/torvalds/linux-2.6.git;a=commitdiff;h=ce31332d3c77532d6ea97ddcb475a2b02dd358b4

Let me know if the issue is still reproducible with Linus' latest git
tree.

thanks
-john

Re: [hrtimers] Re: Regression affecting b43 LP-PHY card

[Rafał Miłecki]

2011/5/10 John Stultz <john.stultz@linaro.org>:
> On Tue, 2011-05-10 at 20:57 +0200, Rafał Miłecki wrote:
>> W dniu 10 maja 2011 00:52 użytkownik Rafał Miłecki <zajec5@gmail.com> napisał:
>> > In this situation we decided to bisect. I was a little afraid of last
>> > merges so we took older 2.6.38 as GOOD (we tested this twice) and
>> > wireless-testing commit before my ssb changes as BAD. Today Juan
>> > finished bisecting kernel:
>> > http://pastebin.com/HSKbRzpB
>> >
>> > According to his bisection the first bad commit is
>> > e06383db9ec591696a06654257474b85bac1f8cb [0]:
>> > hrtimers: extend hrtimer base code to handle more then 2 clockids
>> >
>> > Does it make any sense to you? Could this be some timing issue?
>> >
>> > It was too late to test this today, we (Juan) will work on this
>> > tomorrow. It's impossible to revert this commit from HEAD of
>> > wireless-testing, so my idea is to checkout commit, test, revert,
>> > test.
>> >
>> > Did anyone else experience any similar problems with latest wireless-testing?
>> >
>> >
>> > http://git.kernel.org/?p=linux/kernel/git/torvalds/linux-2.6.git;a=commit;h=e06383db9ec591696a06654257474b85bac1f8cb
>>
>> Today Juan checkouted commit e06383db9ec591696a06654257474b85bac1f8cb
>> and tested it. He was disconnected really soon.
>>
>> Then he reverted e06383db9ec591696a06654257474b85bac1f8cb and tested
>> again. Connection was stable, he downloaded 2GB file over network.
>>
>>
>> John S.: your commit does not touch Broadcom card directly, but it
>> seems it somehow affects it. I suspect there can be some timing issue.
>> Do you have any idea what could it be, how can we debug this?
>
> Sorry for the trouble!
>
> My commit exposed a few spots where hrtimers were being initialized
> before hrtimer_init is called, which caused problems. Thomas provided a
> solution that makes such behavior still function ok:
> http://git.kernel.org/?p=linux/kernel/git/torvalds/linux-2.6.git;a=commitdiff;h=ce31332d3c77532d6ea97ddcb475a2b02dd358b4
>
> Let me know if the issue is still reproducible with Linus' latest git
> tree.

We were using wireless-testing git tree, commit:
1e664a777e5eb4b23e65e76fbeadd2376fe8d8d8

I can not see ce31332d3c77532d6ea97ddcb475a2b02dd358b4 it git log.
I'll apply and test, thanks!

-- 
Rafał

Re: [hrtimers] Re: Regression affecting b43 LP-PHY card

[Rafał Miłecki]

2011/5/10 Rafał Miłecki <zajec5@gmail.com>:
> 2011/5/10 John Stultz <john.stultz@linaro.org>:
>> On Tue, 2011-05-10 at 20:57 +0200, Rafał Miłecki wrote:
>>> W dniu 10 maja 2011 00:52 użytkownik Rafał Miłecki <zajec5@gmail.com> napisał:
>>> > In this situation we decided to bisect. I was a little afraid of last
>>> > merges so we took older 2.6.38 as GOOD (we tested this twice) and
>>> > wireless-testing commit before my ssb changes as BAD. Today Juan
>>> > finished bisecting kernel:
>>> > http://pastebin.com/HSKbRzpB
>>> >
>>> > According to his bisection the first bad commit is
>>> > e06383db9ec591696a06654257474b85bac1f8cb [0]:
>>> > hrtimers: extend hrtimer base code to handle more then 2 clockids
>>> >
>>> > Does it make any sense to you? Could this be some timing issue?
>>> >
>>> > It was too late to test this today, we (Juan) will work on this
>>> > tomorrow. It's impossible to revert this commit from HEAD of
>>> > wireless-testing, so my idea is to checkout commit, test, revert,
>>> > test.
>>> >
>>> > Did anyone else experience any similar problems with latest wireless-testing?
>>> >
>>> >
>>> > http://git.kernel.org/?p=linux/kernel/git/torvalds/linux-2.6.git;a=commit;h=e06383db9ec591696a06654257474b85bac1f8cb
>>>
>>> Today Juan checkouted commit e06383db9ec591696a06654257474b85bac1f8cb
>>> and tested it. He was disconnected really soon.
>>>
>>> Then he reverted e06383db9ec591696a06654257474b85bac1f8cb and tested
>>> again. Connection was stable, he downloaded 2GB file over network.
>>>
>>>
>>> John S.: your commit does not touch Broadcom card directly, but it
>>> seems it somehow affects it. I suspect there can be some timing issue.
>>> Do you have any idea what could it be, how can we debug this?
>>
>> Sorry for the trouble!
>>
>> My commit exposed a few spots where hrtimers were being initialized
>> before hrtimer_init is called, which caused problems. Thomas provided a
>> solution that makes such behavior still function ok:
>> http://git.kernel.org/?p=linux/kernel/git/torvalds/linux-2.6.git;a=commitdiff;h=ce31332d3c77532d6ea97ddcb475a2b02dd358b4
>>
>> Let me know if the issue is still reproducible with Linus' latest git
>> tree.
>
> We were using wireless-testing git tree, commit:
> 1e664a777e5eb4b23e65e76fbeadd2376fe8d8d8
>
> I can not see ce31332d3c77532d6ea97ddcb475a2b02dd358b4 it git log.
> I'll apply and test, thanks!

I can confirm updated wireless-testing resolves this issue!

Too bad Juan spent 2-3 days on bisecting on his Atom... but at least
we came and met ready fix. It could be pain to find a solution with
out feedback only. I'm aware it was not very detailed.

Thanks for your help :)

-- 
Rafał