On 21/09/2020 14:58, John Garry wrote:
On 21/09/2020 14:43, Will Deacon wrote:
On Fri, Aug 21, 2020 at 09:54:20PM +0800, John Garry wrote:
As mentioned in [0], the CPU may consume many cycles processing
arm_smmu_cmdq_issue_cmdlist(). One issue we find is the cmpxchg()
loop to
get space on the queue takes a lot of time once we start getting many
CPUs contending - from experiment, for 64 CPUs contending the cmdq,
success rate is ~ 1 in 12, which is poor, but not totally awful.
This series removes that cmpxchg() and replaces with an atomic_add,
same as how the actual cmdq deals with maintaining the prod pointer.
> I'm still not a fan of this.
:(
Could you try to adapt the hacks I sent before,
please? I know they weren't quite right (I have no hardware to test
on), but
the basic idea is to fall back to a spinlock if the cmpxchg() fails. The
queueing in the spinlock implementation should avoid the contention.
OK, so if you're asking me to try this again, then I can do that, and
see what it gives us.
JFYI, to prove that this is not a problem which affects only our HW, I
managed to test an arm64 platform from another vendor. Generally I see
the same issue, and this patchset actually helps that platform even more.
CPUs Before After % Increase
Huawei D06 8 282K 302K 7%
Other 379K 420K 11%
Huawei D06 16 115K 193K 68K
Other 102K 291K 185K
Huawei D06 32 36K 80K 122%
Other 41K 156K 280%
Huawei D06 64 11K 30K 172%
Other 6K 47K 683%
I tested with something like [1], so unit is map+unmaps per cpu per
second - higher is better.
My D06 is memory poor, so would expect higher results otherwise (with
more memory). Indeed, my D05 has memory on all nodes and performs better.
Anyway, I see that the implementation here is not perfect, and I could
not get suggested approach to improve performance significantly. So back
to the drawing board...
Thanks,
John
[1]
https://lore.kernel.org/linux-iommu/[email protected]/
