On Mon, Jan 06, 2014 at 06:21:47PM +0100, Sander Klein wrote:
> On 06.01.2014 15:10, Willy Tarreau wrote:
> >I would go even further (using git). What I understand here is that the 
> >issue
> >was introduced after the epoll optimization and is hidden by this one. 
> >So I'd
> >rather start by reverting that patch and then looking up for another 
> >faulty
> >patch after those :
> >
> >  1) create a new branch called test1 starting at the first faulty 
> >commit :
> >
> >     git checkout -b test1 2f877304
> >
> >  2) apply the revert patch first :
> >
> >     git cherry-pick 3ef5af3d
> >
> >  3) OK now both the faulty patch and the revert are merged, it makes 
> >sense
> >     to confirm that the bug is still not there.
> >
> >  4) now rebase all further patches on top of these ones : Git will 
> >re-apply
> >     all other patches after the ones above. You will thus have a 
> >working
> >     version to start from :
> >
> >     git checkout -b test2 master
> >     git rebase test1
> >
> >  5) ensure that branch test2 is wrong by doing a test
> >
> >  6) bisect the code from test1 which was verified to be good at 3) and 
> >test2
> >     which was verified to be bad at 5) :
> >
> >     git bisect start test2 test1
> >
> >It will offer you another patch which introduced the regression hidden 
> >by the
> >one above.
> 
> I will do the bisect as soon as I have time.
> 
> On a side note, today I had an issue with another loadbalancer running 
> ss-20140101 which showed almost the same behavior as the 'NTLM' bug I 
> was having. (hanging connection, or waiting a loooong time and the 
> giving a corrupt file) This bug only happened with certain downloads 
> (jpg's) with http compression enabled. If a browser requested the file 
> without the compression header everything was fine.
> 
> Downgrading to dev19 also fixed this issue. I don't know if this could 
> be related somehow.

I don't know either but that's good to know! It is indeed possible that
the server would not correctly chunk responses and that it makes a
difference after some of the changes that were required for keep-alive,
but that's just an hypothesis.

Willy


Reply via email to