Message-ID: <1349795389.24256.20.camel@gandalf.local.home>
Date:	Tue, 09 Oct 2012 11:09:49 -0400
From:	Steven Rostedt <rostedt@...dmis.org>
To:	Thomas Gleixner <tglx@...utronix.de>
Cc:	LKML <linux-kernel@...r.kernel.org>,
	linux-rt-users <linux-rt-users@...r.kernel.org>,
	Carsten Emde <cbe@...dl.org>
Subject: Re: [ANNOUNCE] 3.6.1-rt1

On Tue, 2012-10-09 at 15:46 +0200, Thomas Gleixner wrote:
> Dear RT Folks,
> 
> I'm pleased to announce the 3.6.1-rt1 release.

Thomas,

First I want to say, and I'm sure I speak for a lot of people on this,
is "Thank you!". I know how hard it is to deal with the issues of
mainline in a RT specific way, and to balance both the determinism
required by RT with the non-intrusiveness to the work flow of mainline.
When this is done right, both mainline and RT benefit. Interesting
enough, Linus knew this a long time ago, and by denying RT only
enhancements to the kernel, he forced us to improve mainline in
general ;-)

> 
> This is a pretty straight forward move from the 3.4-rt series which
> includes a few significant updates which need to be backported to the
> 3.x-rt stable series:
> 
>    * Make interrupt randomness work again on RT. Based on the 3.x.y
>      stable updates in that area. Should be applicable to all 3.x-rt
>      series with almost no modifications.
> 
>    * RT softirq initialization sequence fix (Steven Rostedt)
> 
>    * Fix for a potential deadlock in mm/slab.c. This had been reported
>      as lockdep splats several times and stupidly ignored as a false
>      positive, but in fact it's a real (though almost impossible to
>      trigger) deadlock lurking.
> 
>    * Use the proper local_lock primitives in mm/page_alloc.c. That's
>      not a real bug, but this fixes an inconsistency which helps
>      debugability and therefore is worthwhile to be backported.
> 
>    * RT-rwlock/rwsem annotations:
> 
>      RT does not allow multiple readers on rwlocks and rwsems. The
>      lockdep annotations did not yet consider that fact. One might
>      think that this is a complete RT specific issue, but it's
>      not. The FIFO fair rwsem/lock modifications in mainline made
>      reader/writer primitives prone to very subtle deadlock problems
>      which cannot be detected by the current lockdep annotations in
>      mainline. The reason is that if a writer interleaves with two
>      readers it will block the second reader from proceeding in order
>      not to allow writer starvation. The restricted RWlocks semantics
>      of RT allow an easy detection of that problem. We already
>      triggered a real deadlock in RT (see:
>      peterz-srcu-crypto-chain.patch) which could result in a hard to
>      trigger, but mainline relevant deadlock. Wait for more
>      interesting problems in that area.
> 
>    * The output of might_sleep debugging is silent about the possible
>      causes vs. the preempt count. Contrary to interrupt disabling
>      there is zero information about what disabled preemption
>      last. Again, not strictly a bugfix, but debuggability is key.
> 
>    * Fix a potentially deadly sto(m)p_machine deadlock. A CPU which
>      calls that code from its inactive state (don't ask me for the
>      ghastly deatils why this is necessary) can run into a contended
>      state of the stomp machine mutex which would cause a rather
>      awkward issue of idle scheduling itself away to idle as the only
>      possible task on that upcoming cpu. Not pretty ....

Here's my road map as everyone just loves them:

  I'm finishing up on releasing the next merge of:

   3.0.45-rt67 and 3.4.13-rt21

  These are only merging the stable 3.0.45 and 3.4.13 without any -rt
  specific changes.

  I'll then backport these fixes to the stable release and release an
  -rc for 3.0.45-rt68 and 3.4.13-rt22

  For 3.2-rt, I'm waiting for the final release of 3.2.31 to be done
  and will be going through the same ordeal with that. That is, I'll
  release a 3.2.31 merged rt only (3.2.31-rt46) and then backport
  and release a -rc for 3.2.31-rt47. This will come later.

> 
> 
> There is also a worth to mention fundamental change in this release:
> 
>    * Split softirq locks

Although this work is not for stable (and shouldn't be), I'm thinking
about backporting these to the 3.2 and 3.4 trees and creating a separate
branch for them. This way, those that want this feature based on the
3.2/3.4 stable trees, can have the same repository to work from.

> 
>      In the pre 3,x-RT versions we spawned a separate thread for each
>      softirq on each CPU. This served the PER_CPUness requirements,
>      but did not provide any means against priority inversions
>      vs. softirqs.
> 
>      With the start of the 3.0-rt series I decided to drop the per
>      softirq threads for simplicity reasons as I had to deal with all
>      the fallout of the migration disabling design I had taken course
>      to.
> 
>      I got several complaints about the missing softirq thread split
>      since then and a few patches to reestablish them. I refused to
>      take those patches for a simple reason: configuration. It's
>      extremly hard to get the parameters right for a RT system in
>      general. Adding something which is obscure as soft interrupts to
>      the system designers todo list is a bad idea.
> 
>      Now I spent quite some time on analysing the most urgent issues
>      on RT:
> 
>      	  throughpout  versus  deterministism
> 
>      The interested observer may have noticed that deterministic
>      behaviour and throughput are mutually exclusive properties, but
>      in the 2.6 based RT series the split softirq implementation at
>      least allowed some mitigation of this problem by adjusting the
>      priorities, while the 3.x RT series did not provide a user
>      tunable knob at all. Though the untunable behaviour of the 3.x RT
>      series behaved in general better than the untuned 2.6 RT at least
>      in terms of throughput.
> 
>      The reason is that 3.x RT put a big focus on dealing with the
>      increasing PER_CPUness of data in the mainline kernel. The
>      migrate_disable based ability of executing pending soft
>      interrupts in any thread context which had raised a soft
>      interrupt made RT a bit more similar to the mainline behaviour,
>      but did not provide any serious means of controlling that
>      behaviour.
> 
>      My new approach of split softirq locks is another (sigh) futile
>      attempt to deal with the current (non)existing softirq semantics
>      of the mainline kernel.
> 
>      What's the meaning of soft interrupt processing in Linux today?
> 
>      First of all, it's a conglomorate of mostly unrelated jobs, which
>      run in the context of a randomly chosen victim w/o the ability to
>      put any control on them. Softirq processing happens in three
>      contexts:
> 
>      - Return from hard interrupt context. Basically the same as the
>        hard interrupt context except that interrupts are enabled.
> 
>      - In the context of a thread which reenables softirq processing
>        via local_bh_enable or *_unlock_bh. Interestingly enough
>        reenabling interrupts does not have the same effect, though
>        disabling interrupts prevents softirq processing as well and
>        there are places which raise soft interrupt in interrupt
>        disabled regions which delegates them to ksoftirqd or to the
>        next random context which happens to reach a softirq processing
>        context before ksoftirqd.
> 
>      - ksoftirqd. The invocation of ksoftirqd is not well defined. It
>        happens when the above two contexts looped extensivly in the
>        softirq processing or if a softirq gets raised outside of a
>        hard interrupt context in a bottom half enabled region. Now
>        even in the case that ksoftirqd has been woken up there is no
>        guarantee, that it will actually process softinterrupts at all
>        because the other two contexts can be invoked (again) before
>        ksoftirqd gets scheduled in.
> 
>      Quite a set of imprecise rules and unclear semantics which
>      explain the pain RT has with softirq processing.
> 
>      The pre 3.0-RT approach of delegating all softirq processing to
>      separate per softirq threads is only a partial solution to the
>      problem and introduces a hard to configure set of softirq thread
>      scheduling policy and priority questions. Aside of that it does
>      not allow to process soft interrupts from the tail of the
>      interrupt threads or softirq enabling code.
> 
>      The 3.x-RT approach of allowing the softirq processing from the
>      tail of interrupt threads or softirq enabling code gave us a
>      throughput enhancement and got rid of the configuration
>      complexity, but we lost the ability to optimize for specific use
>      cases (e.g. deterministic networking).
>      
>      After studying the softirq behaviour I came to the conclusion
>      that it might be interesting to try a different approach.
>      Especially networking handles the softirq processing either in
>      the interrupt tail or from bh enabling thread contexts.
> 
>      So instead of splitting the softirq threads I split the softirq
>      locks so different softirqs can be handled seperately. If a
>      softirq is raised in the context of a thread, then its noted in
>      the task struct and when the thread leaves the bh disabled
>      section it handles this particular soft interrupt in its own
>      context. This removes the burden of running completely unrelated
>      softirqs like timers, tasklets etc. from a context which raised a
>      network soft interrupt. That way the softirq processing is
>      coupled to the originating thread and its scheduling properties,
>      so the need for finding optimal parameters should be gone.

Very interesting. I haven't looked at the patches yet (will do that
after I finish with the stable merge releases), but I started looking
into the softirq changes as well, and came up with something almost
identical. I talked a little with Carsten about this, and he told me to
wait for your release, which I then did, and I'm glad I did :-)

I was looking specifically at the network softirqs as well, and started
some patches to separate out the softirqs with the task (sounds similar
to what you did). But before that, I also played with the
local_softirq_lock. For the end of interrupt processing only (where it
should always be safe to lock), if the trylock fails, I grabbed it and
then released it. Because if a lower priority task is currently running
the softirq that the higher priority interrupt wants to run, it would at
least priority boost the lower thread, and the higher priority interrupt
could run its softirq at its priority. Maybe this can still be added?


> 
>      Now this only works for soft interrupts which are raised in the
>      context of a thread. Unfortunately there is no way to do the same
>      for soft interrupts which are raised in hard interrupt context
>      (e.g. RCU, timers). They have no thread associated and are
>      therefor delegated to ksoftirqd. This is ok, except that it does
>      not help people who want to use signal based timers, but that
>      problem needs to be solved by moving the complex handling into
>      the context of the thread which is going to receive the signal
>      and should vanish from the softirq processing completely.
>      
>      In principle we should have even in mainline a clear separation
>      of which soft interrupts are disabled by a particular code region
>      instead of disabling them wholesale. Though the nicest solution
>      would be to get rid of them completely :)

I've started looking at playing with the NAPI code again, and trying to
see if I can add an ENAPI interface (Even Newer API), where the driver
uses its own interrupt thread, and instead of having the polling in the
network softirq, it can do the polling in its own thread.

>  
>      Give it proper testing and lemme know whether this solves your
>      particular problems which arised from giving up the separate
>      softirq threads. Don't complain about signal based timers - see
>      above!
> 
> The RT patch against 3.6.1 can be found here:
> 
>   http://www.kernel.org/pub/linux/kernel/projects/rt/3.6/patch-3.6.1-rt1.patch.xz
> 
> The split quilt queue is available at:
> 
>   http://www.kernel.org/pub/linux/kernel/projects/rt/3.4/patches-3.6.1-rt1.tar.xz
> 
> Enjoy,

Will do!

-- Steve


--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@...r.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/

Powered by blists - more mailing lists