On 2017年07月29日 03:09, Vikram Mulukutla wrote:
On 2017-07-28 02:28, Will Deacon wrote:
On Thu, Jul 27, 2017 at 06:10:34PM -0700, Vikram Mulukutla wrote:
<snip>
I think we should have this discussion now - I brought this up
earlier [1]
and I promised a test case that I completely forgot about - but here it
is (attached). Essentially a Big CPU in an acquire-check-release loop
will have an unfair advantage over a little CPU concurrently attempting
to acquire the same lock, in spite of the ticket implementation. If
the Big
CPU needs the little CPU to make forward progress : livelock.
<snip>
One solution was to use udelay(1) in such loops instead of
cpu_relax(), but
that's not very 'relaxing'. I'm not sure if there's something we
could do
within the ticket spin-lock implementation to deal with this.
Does bodging cpu_relax to back-off to wfe after a while help? The event
stream will wake it up if nothing else does. Nasty patch below, but
I'd be
interested to know whether or not it helps.
Will
The patch also helps a lot on my platform. (Though it does cause
deadlock(related with udelay) in uart driver in early boot, and not sure
it's uart driver issue. Just workaround it firstly)
Platform: 4 a53(832MHz) + 4 a73(1.8GHz)
Test condition #1:
a. core2: a53, while loop (spinlock, spin_unlock)
b. core7: a73, while loop (spinlock, spin_unlock, cpu_relax)
Test result: recording the lock acquire times(a53, a73), max lock
acquired time(a53), in 20 seconds
Without cpu_relax bodging patch:
===============================================================
|a53 locked times | a73 locked times | a53 max locked time(us)|
==================|==================|========================|
182| 38371616| 1,951,954|
202| 38427652| 2,261,319|
210| 38477427| 15,309,597|
207| 38494479| 6,656,453|
220| 38422283| 2,064,155|
===============================================================
With cpu_relax bodging patch:
===============================================================
|a53 locked times | a73 locked times | a53 max locked time(us)|
==================|==================|========================|
1849898| 37799379| 131,255|
1574172| 38557653| 38,410|
1924777| 37831725| 42,999|
1477665| 38723741| 52,087|
1865793| 38007741| 783,965|
===============================================================
Also add some workload to the whole system to check the result.
Test condition #2: based on #1
c. core6: a73, 1.8GHz, run "while(1);" loop
With cpu_relax bodging patch:
===============================================================
|a53 locked times | a73 locked times | a53 max locked time(us)|
==================|==================|========================|
20| 42563981| 2,317,070|
10| 42652793| 4,210,944|
9| 42651075| 5,691,834|
28| 42652591| 4,539,555|
10| 42652801| 5,850,639|
===============================================================
Also hotplug out other cores.
Test condition #2: based on #1
d. hotplug out core1/3/4/5/6, keep core0 for scheduling
With cpu_relax bodging patch:
===============================================================
|a53 locked times | a73 locked times | a53 max locked time(us)|
==================|==================|========================|
447| 42652450| 309,549|
515| 42650382| 337,661|
415| 42646669| 628,525|
431| 42651137| 365,862|
464| 42648916| 379,934|
===============================================================
The last two tests are the actual cases where the hard-lockup is
triggered on my platform. So I gathered some data, and it shows that a53
needs much longer time to acquire the lock.
All tests are done in android, black screen with USB cable attached. The
data is not so pretty as Vikram's. It might be related with cpu
topology, core numbers, CCI frequency etc. (I'll do another test with
both a53 and a73 running at 1.2GHz, to check whether it's the core
frequency which leads to the major difference.)
This does seem to help. Here's some data after 5 runs with and without
the patch.
time = max time taken to acquire lock
counter = number of times lock acquired
cpu0: little cpu @ 300MHz, cpu4: Big cpu @2.0GHz
Without the cpu_relax() bodging patch:
=====================================================
cpu0 time | cpu0 counter | cpu4 time | cpu4 counter |
==========|==============|===========|==============|
117893us| 2349144| 2us| 6748236|
571260us| 2125651| 2us| 7643264|
19780us| 2392770| 2us| 5987203|
19948us| 2395413| 2us| 5977286|
19822us| 2429619| 2us| 5768252|
19888us| 2444940| 2us| 5675657|
=====================================================
cpu0: little cpu @ 300MHz, cpu4: Big cpu @2.0GHz
With the cpu_relax() bodging patch:
=====================================================
cpu0 time | cpu0 counter | cpu4 time | cpu4 counter |
==========|==============|===========|==============|
3us| 2737438| 2us| 6907147|
2us| 2742478| 2us| 6902241|
132us| 2745636| 2us| 6876485|
3us| 2744554| 2us| 6898048|
3us| 2741391| 2us| 6882901|
==================================================== >
The patch also seems to have helped with fairness in general
allowing more work to be done if the CPU frequencies are more
closely matched (I don't know if this translates to real world
performance - probably not). The counter values are higher
with the patch.
time = max time taken to acquire lock
counter = number of times lock acquired
cpu0: little cpu @ 1.5GHz, cpu4: Big cpu @2.0GHz
Without the cpu_relax() bodging patch:
=====================================================
cpu0 time | cpu0 counter | cpu4 time | cpu4 counter |
==========|==============|===========|==============|
2us| 5240654| 1us| 5339009|
2us| 5287797| 97us| 5327073|
2us| 5237634| 1us| 5334694|
2us| 5236676| 88us| 5333582|
84us| 5285880| 84us| 5329489|
=====================================================
cpu0: little cpu @ 1.5GHz, cpu4: Big cpu @2.0GHz
With the cpu_relax() bodging patch:
=====================================================
cpu0 time | cpu0 counter | cpu4 time | cpu4 counter |
==========|==============|===========|==============|
140us| 10449121| 1us| 11154596|
1us| 10757081| 1us| 11479395|
83us| 10237109| 1us| 10902557|
2us| 9871101| 1us| 10514313|
2us| 9758763| 1us| 10391849|
=====================================================
Thanks,
Vikram
