Re: [RFC] Kernel version numbering scheme change

[Alex Howells <alex@bytemark.co.uk>]

Hey Valdis

>> Requirements for me to put a kernel on a given server would be:
> 
>>     *  supports the hardware
> The problem is that "supports" is often a fuzzy jello-like substance you
> try to nail to a tree.  You mention the R8169 and e1000 drivers - if they
> bring the device up, but have issues under corner cases, is that "supports"
> or not?

Oh agreed, this is all very "use case" specific.  I'm making all of the
following statements based on the specific hardware we use, and assuming
'stability' based on the kernel/hardware passing a number of tests.

>>     *  no security holes [in options I enable]
> Similarly for "no security holes".  At *BEST*, you'll get "no *known* *major*
> security holes", unless you feel like auditing the entire source tree.  There's
> a whole slew of bugs that we can't even agree if they *are* security bugs or
> just plain bugs - see Linus's rant on the subject a few months back.

Agreed. No *known* *major* security holes is fine here.

>>     *  works reliably, under load/stress.
> And you win the trifecta - I don't think we've *ever* shipped a Linux kernel
> that worked reliably under the proper "beat on the scheduler/VM corner case"
> load/stress testing.  Again, the best you can hope for is "doesn't fall over
> under non-pathological non-corner-case loads when sufficient resources are
> available so the kernel has a fighting chance".  Doing 'make -j100' on a
> single Core2 Duo is gonna be painful, no matter what.

Well the typical tests outlined above are:

    *  random size file creation/deletion, lots of files
    *  memory allocation, and freeing up again
    *  stressing the CPU a bit with one process, then
       forking 25-50 processes to (trivially) test scheduler
    *  testing network I/O by rapidly/concurrently fetching
       many small files via HTTP, and a few large ones.

The end goal is simply to get a server which doesn't crash under
"normal" operating conditions.  The bugs I referred to in
e1000/forcedeth and r8169 either stop it PXE booting (a requirement for
our environment!) or can *easily* be made to oops / stop working.

Alex