Integrating and Validating dynticks and Preemptable RCU [LWN.net]

Integrating and Validating dynticks and Preemptable RCU

Posted Apr 24, 2008 19:15 UTC (Thu) by ds2horner (subscriber, #13438)
Parent article: Integrating and Validating dynticks and Preemptable RCU

There are a few typos that (when fixed ) may help the read follow better:

Interrupt entry is handled similarly by rcu_irq_exit():
should be 
Interrupt exit is handled similarly by rcu_irq_exit():

and in the same section

(preceding the rcu_irq_enter() invocation).
should be
(preceding the rcu_irq_exit() invocation).


the definition of in_interrupt would help (but likely out of scope) as would an explanation of
MNI behaviour (does it complete handing before any masked interupts are enabled?)  

However, the most puzzling to me was why the code (which finally got fixed)
 
 10   if ((curr == snap) && ((curr & 0x1) == 0))
 11     return 0;
 12   if ((curr - snap) > 2 || (curr & 0x1) == 0)
 13     return 0;

isn't coded as 

 10   if ((curr & 0x1) == 0) || (curr - snap) > 2)
 11     return 0;
 
Which I would read as:
 If we are in dynticks or have passed through dynticks since snapshot this cpu's good to go.

(The simplification occurred to me in the original code, but meant semantically much less due
to the potentially expired snapshot state)

(Log in to post comments)

passed through dynticks

Posted Apr 26, 2008 1:05 UTC (Sat) by PaulMcKenney (subscriber, #9624) [Link]

Excellent points, all on the money!!!  I clearly should submit more of my code to LWN for
review!!!

I made the changes to spin, and all but the last (big) one completed
successfully.  This last one is in progress, and will take a bit to complete.  I will let you
know how it goes.

Thank you very much for your careful and productive review!

passed through dynticks

Posted Apr 26, 2008 3:12 UTC (Sat) by ds2horner (subscriber, #13438) [Link]

Thank you for your encouraging words.

I was reluctant to add more - ( still working through the logic and especially the Promela
modeling verification process) - until I got some confirmation that I was grasping the
concepts.

I would like to suggest a further code simplification in rcu_try_flip_waitmb_needed.

  if ((curr == snap) && ((curr & 0x1) == 0))
    return 0;
  if (curr != snap)
    return 0;

reduces to:

if ((curr & 0x1) == 0 || (curr != snap) )
    return 0;

If we are in dynticks we have not done any RCU activity since entering (and have done the
required memory barrier on entry, so nothing to synch (is there a miniscule race here between
the setting of the dynticks_progress_counter and invoking of the barrier on the remote cpu?))
or we have transitioned dynticks states and therefore invoked the required mb.


And some more editting/profreading corrections:

like blank line 4 in rcu_irq_enter throwing off the description in the following paragraph.
should be 

while lines 5 and 6 ... (not 4 and 5)  
and
Lines 7 and 8 .. (not 6 and 7) 

Do you want me to report such text corrections?

I am still working my way through this.

I greatly appreciate such articles in LWN 
- the description around the code greatly helps in validating possible corrections or
recommendations.

passed through dynticks

Posted Apr 27, 2008 0:13 UTC (Sun) by PaulMcKenney (subscriber, #9624) [Link]

Very good -- much simpler formulation!  (A bit faster as well, but as this is the slow path,
the simplicity is more important.)  Passes Promela/spin, and also rcutorture.

Feel free to post feedback, or email it to me if you prefer.  Don't worry about line-number
mismatches, as they will all change anyway.

Would you be willing to review before publication on the next one?  If so, please drop me an
email.

RCU reviews

Posted Apr 27, 2008 4:03 UTC (Sun) by ds2horner (subscriber, #13438) [Link]

I would be honoured to pre-view the next LWN article.
However, I believe I am mostly out of my league here. 
I hope to not disappoint you.

I understand memory barriers are quite expensive (on most current machines that implement
them), so I have some suggestions for optimizations in rcu_try_flip_waitmb_needed and
rcu_try_flip_waitack_needed. 

And I understand mb is both read and write barrier. I plan on looking into the possibility of
using (the potentially) less expensive read (or write) barriers, if you think they may
benefit.

I did not see your email anywhere on LWN - even in the new "GuestArticles index", so I will
look for your email on lkml. Having seen no other posts here, perhaps we are OK going private
correspondence.

Thanks again for these informative (and thought provoking) articles.