svn commit: r277213 - in head: share/man/man9 sys/kern sys/ofed/include/linux sys/sys

[Randy Stewart]

Hans:

We (netflix) run in production 35% of the internet with these very things
you identify no lock an all. We *do* have some issue we are looking at but so far
I have *never* connected the dots the way you were claiming that would
cause a crash. I can see where TCP would do incorrect retransmissions but
I did *not* see a crash. Now granted my look was quick at this, but that
was due to time constraints and the holidays. I am going to put myself full-time
on this to see if I can understand both how you got at “there is a panic in tcp” and
it must fully be the callout-subsystem thus we need to re-write large parts of it.

You *may* be correct in a re-write is needed, you *may* be completely incorrect.
In either case I plan to dig into this and find out.

R
> On Jan 22, 2015, at 3:39 AM, Hans Petter Selasky <hps at selasky.org> wrote:
> 
> On 01/22/15 09:10, Konstantin Belousov wrote:
>> On Thu, Jan 22, 2015 at 08:14:26AM +0100, Hans Petter Selasky wrote:
>>> On 01/22/15 06:26, Warner Losh wrote:
>>>  >
>>>>> The code simply needs an update. It is not broken in any ways - right? If it is not broken, fixing it is not that urgent.
>>>> 
>>>> Radically changing the performance characteristics is breaking the code. Performance regression in the TCP stack is urgent to fix.
>> 
>>> Not being able to enumerate what all the consumers are that use this and
>>> provide an analysis about why they aren?t important to fix is a bug in
>>> your process, and in your interaction with the project. We simply do not
>>> operate that way.
>> Right, I completely agree with this statement.
>> 
>> 
>>> Hi,
>>> 
>>> My plan is to work out a patch for the TCP stack today, which only
>>> change the callout_init() call or its function. This should not need any
>>> particular review. I'll let adrian test and review, because I think he
>>> is closer to me timezone wise and you're standing on my head saying its
>>> urgent. If he is still not happy, I can back my change out. Else it
>>> remains in -current AS-IS.
>> TCP regresssion was noted, so it is brought in front.  There is nothing
>> else which makes TCP issue different from other (hidden) issues.
>> 
>> ===========================
>>> MFC to 10-stable I can delay for sure until
>>> all issues you report to me are fixed.
>> ===========================
>> 
>> Sigh, you still do not understand.  It is your duty to identify all pieces
>> which break after your change.  After that, we can argue whether each of
>> them is critical or not to allow the migration. But this must have been
>> done before the KPI change hit the tree.
>> 
> 
> Hi,
> 
> Are you saying that pieces of code that runs completely unlocked using "volatile" as only synchronization mechanism is better than what I would call a temporary and hopefully short TCP stack performance loss?
> 
> I don't understand? How frequently do you reboot your boxes? Maybe one every day? And you don't care?
> 
> --HPS
> 
> 
> 

-----
Randall Stewart
randall at lakerest.net