On 7/8/20 1:10 AM, Nicholas Piggin wrote:
Excerpts from Waiman Long's message of July 8, 2020 1:33 pm:
On 7/7/20 1:57 AM, Nicholas Piggin wrote:
Yes, powerpc could certainly get more performance out of the slow
paths, and then there are a few parameters to tune.
We don't have a good alternate patching for function calls yet, but
that would be something to do for native vs pv.
And then there seem to be one or two tunable parameters we could
experiment with.
The paravirt locks may need a bit more tuning. Some simple testing
under KVM shows we might be a bit slower in some cases. Whether this
is fairness or something else I'm not sure. The current simple pv
spinlock code can do a directed yield to the lock holder CPU, whereas
the pv qspl here just does a general yield. I think we might actually
be able to change that to also support directed yield. Though I'm
not sure if this is actually the cause of the slowdown yet.
Regarding the paravirt lock, I have taken a further look into the
current PPC spinlock code. There is an equivalent of pv_wait() but no
pv_kick(). Maybe PPC doesn't really need that.
So powerpc has two types of wait, either undirected "all processors" or
directed to a specific processor which has been preempted by the
hypervisor.
The simple spinlock code does a directed wait, because it knows the CPU
which is holding the lock. In this case, there is a sequence that is
used to ensure we don't wait if the condition has become true, and the
target CPU does not need to kick the waiter it will happen automatically
(see splpar_spin_yield). This is preferable because we only wait as
needed and don't require the kick operation.
Thanks for the explanation.
The pv spinlock code I did uses the undirected wait, because we don't
know the CPU number which we are waiting on. This is undesirable because
it's higher overhead and the wait is not so accurate.
I think perhaps we could change things so we wait on the correct CPU
when queued, which might be good enough (we could also put the lock
owner CPU in the spinlock word, if we add another format).
The LS byte of the lock word is used to indicate locking status. If we
have less than 255 cpus, we can put the (cpu_nr + 1) into the lock byte.
The special 0xff value can be used to indicate a cpu number >= 255 for
indirect yield. The required change to the qspinlock code will be
minimal, I think.
Attached are two
additional qspinlock patches that adds a CONFIG_PARAVIRT_QSPINLOCKS_LITE
option to not require pv_kick(). There is also a fixup patch to be
applied after your patchset.
I don't have access to a PPC LPAR with shared processor at the moment,
so I can't test the performance of the paravirt code. Would you mind
adding my patches and do some performance test on your end to see if it
gives better result?
Great, I'll do some tests. Any suggestions for what to try?
I will just like to see if it will produce some better performance
result compared with your current version.
Cheers,
Longman
