Permit init(8) use its own cpuset group.

[Alexander V. Chernikov]

On 05.06.2014 23:59, John Baldwin wrote:
> On Thursday, June 05, 2014 3:46:15 pm Alexander V. Chernikov wrote:
>> On 05.06.2014 18:09, John Baldwin wrote:
>>> On Wednesday, June 04, 2014 3:16:59 pm Alexander V. Chernikov wrote:
>>>> On 04.06.2014 19:06, John Baldwin wrote:
>>>>> On Monday, June 02, 2014 12:48:50 pm Konstantin Belousov wrote:
>>>>>> On Mon, Jun 02, 2014 at 06:52:10PM +0400, Alexander V. Chernikov wrote:
>>>>>>> Hello list!
>>>>>>>
>>>>>>> Currently init(8) uses group 1 which is root group.
>>>>>>> Modifications of this group affects both kernel and userland threads.
>>>>>>> Additionally, such modifications are impossible, for example, in 
>>> presence
>>>>>>> of multi-queue NIC drivers (like igb or ixgbe) which binds their 
> threads
>>>>> to
>>>>>>> particular cpus.
>>>>>>>
>>>>>>> Proposed change ("init_cpuset" loader tunable) permits changing cpu
>>>>>>> masks for
>>>>>>> userland more easily. Restricting user processes to migrate to/from 
> CPU
>>>>>>> cores
>>>>>>> used for network traffic processing is one of the cases.
>>>>>>>
>>>>>>> Phabricator: https://phabric.freebsd.org/D141 (the same version 
> attached
>>>>>>> inline)
>>>>>>>
>>>>>>> If there are no objections, I'll commit this next week.
>>>>>> Why is the tunable needed ?
>>>>> Because some people already depend on doing 'cpuset -l 0 -s 1'.  It is 
>>> also
>>>>> documented in our manpages that processes start in cpuset 1 by default 
> so
>>>>> that you can use 'cpuset -l 0 -s 1' to move all processes, etc.
>>>>>
>>>>> For the stated problem (bound ithreads in drivers), I would actually 
> like 
>>> to
>>>>> fix ithreads that are bound to a specific CPU to create a different 
> cpuset
>>>>> instead so they don't conflict with set 1.
>>>> Yes, this seem to be much better approach.
>>>> Please take a look on the new patch (here or in the phabricator).
>>>> Comments:
>>>>
>>>> Use different approach for modifyable root set:
>>>>
>>>>   * Make sets in 0..15 internal
>>>>   * Define CPUSET_SET1 & CPUSET_ITHREAD in that range
>>>>   * Add cpuset_lookup_builtin() to retrieve such cpu sets by id
>>>>   * Create additional root set for ithreads
>>>>   * Use this set in ithread_create()
>>>>   * Change intr_setaffinity() to use cpuset_iroot (do we really need 
> this)?
>>>>
>>>> We can probably do the same for kprocs, but I'm unsure if we really need 
> it.
>>>
>>> I imagined something a bit simpler.  Just create a new set in 
> intr_event_bind
>>> and bind the ithread to the new set.  No need to have more magic set ids, 
> etc. 
>> Well, we also have userland which can modify given changes via `cpuset
>> -x', so we need to be able to add some more logic on set
>> allocation/keeping. Additionally, we can try to do the same via `cpuset
>> -t', so introducing something like cpuset_setIthread() and hooking into
>> intr_event_bind() won't probably be enough. At least I can't think out a
>> quick and easy way to do this.
> 
> cpuset -x calls intr_event_bind().  If you just do it there you fix both 
> places.
1:04 [0] ra# procstat -t 12 | grep irq275
   12 100121 intr             irq275: ix0:qu1    2  127 wait    -
1:04 [0] ra# cpuset -g -x 275
irq 275 mask: 2
1:04 [0] ra# cpuset -g -t 100121
tid 100121 mask: 2
1:04 [0] ra# cpuset -l 3 -t 100121
-------------------------^^^------
1:05 [0] ra# cpuset -g -t 100121
tid 100121 mask: 3

> 
>>> That also means that an ithread that isn't bound to a specific CPU via 
> either 
>>> 'cpuset -x' or BUS_BIND_INTR() will honor 'cpuset -s 1' like other
>>> kernel processes.  I think that's probably fine and sensible.  The issue 
> is
>> Well, it is questionable. Kernel threads are a bit different in terms of
>> TLB changes, memory working set and so on. (Personally I'd prefer to
>> separate user / kthreads / ithreads to different sets in HEAD but that's
>> another story).
>>
>> Anyway, we probably can (and should) MFC a bit different version which
>> tries to change several sets at once if user supplied set 1 as argument.
> 
> No, I don't think we need umpteen special sets.  I think we just need to fix
> this one specific case of bound ithreads and everything else will work fine.  
> If someone wants to move kprocs out of set 1, they can already do that with 
> the existing tools via cpuset -C, etc.
>