Re: [PATCH RFC 1/1] genirq: Make threaded handler use irq affinity for managed interrupt

From: Ming Lei <ming.lei@redhat.com>
To: John Garry <john.garry@huawei.com>
Cc: Marc Zyngier <maz@kernel.org>,
	tglx@linutronix.de, "chenxiang (M)" <chenxiang66@hisilicon.com>,
	bigeasy@linutronix.de, linux-kernel@vger.kernel.org,
	hare@suse.com, hch@lst.de, axboe@kernel.dk, bvanassche@acm.org,
	peterz@infradead.org, mingo@redhat.com,
	Zhang Yi <yi.zhang@redhat.com>
Subject: Re: [PATCH RFC 1/1] genirq: Make threaded handler use irq affinity for managed interrupt
Date: Fri, 3 Jan 2020 08:46:25 +0800	[thread overview]
Message-ID: <20200103004625.GA5219@ming.t460p> (raw)
In-Reply-To: <72a6a738-f04b-3792-627a-fbfcb7b297e1@huawei.com>

On Thu, Jan 02, 2020 at 10:35:31AM +0000, John Garry wrote:
> On 25/12/2019 00:48, Ming Lei wrote:
> > On Tue, Dec 24, 2019 at 11:20:25AM +0000, Marc Zyngier wrote:
> > > On 2019-12-24 01:59, Ming Lei wrote:
> > > > On Mon, Dec 23, 2019 at 10:47:07AM +0000, Marc Zyngier wrote:
> > > > > On 2019-12-23 10:26, John Garry wrote:
> > > > > > > > > > I've also managed to trigger some of them now that I have
> > > > > > > > > access to
> > > > > > > > > > a decent box with nvme storage.
> > > > > > > > > 
> > > > > > > > > I only have 2x NVMe SSDs when this occurs - I should not be
> > > > > > > > > hitting this...
> > > > > > > > > 
> > > > > > > > > Out of curiosity, have you tried
> > > > > > > > > > with the SMMU disabled? I'm wondering whether we hit some
> > > > > > > > > livelock
> > > > > > > > > > condition on unmapping buffers...
> > > > > > > > > 
> > > > > > > > > No, but I can give it a try. Doing that should lower the CPU
> > > > > > > > > usage, though,
> > > > > > > > > so maybe masks the issue - probably not.
> > > > > > > > 
> > > > > > > > Lots of CPU lockup can is performance issue if there isn't
> > > > > > > > obvious bug.
> > > > > > > > 
> > > > > > > > I am wondering if you may explain it a bit why enabling SMMU
> > > > > may
> > > > > > > > save
> > > > > > > > CPU a it?
> > > > > > > The other way around. mapping/unmapping IOVAs doesn't comes for
> > > > > > > free.
> > > > > > > I'm trying to find out whether the NVMe map/unmap patterns
> > > > > trigger
> > > > > > > something unexpected in the SMMU driver, but that's a very long
> > > > > > > shot.
> > > > > > 
> > > > > > So I tested v5.5-rc3 with and without the SMMU enabled, and
> > > > > without
> > > > > > the SMMU enabled I don't get the lockup.
> > > > > 
> > > > > OK, so my hunch wasn't completely off... At least we have something
> > > > > to look into.
> > > > > 
> > > > > [...]
> > > > > 
> > > > > > Obviously this is not conclusive, especially with such limited
> > > > > > testing - 5 minute runs each. The CPU load goes up when disabling
> > > > > the
> > > > > > SMMU, but that could be attributed to extra throughput (1183K ->
> > > > > > 1539K) loading.
> > > > > > 
> > > > > > I do notice that since we complete the NVMe request in irq
> > > > > context,
> > > > > > we also do the DMA unmap, i.e. talk to the SMMU, in the same
> > > > > context,
> > > > > > which is less than ideal.
> > > > > 
> > > > > It depends on how much overhead invalidating the TLB adds to the
> > > > > equation, but we should be able to do some tracing and find out.
> > > > > 
> > > > > > I need to finish for the Christmas break today, so can't check
> > > > > this
> > > > > > much further ATM.
> > > > > 
> > > > > No worries. May I suggest creating a new thread in the new year,
> > > > > maybe
> > > > > involving Robin and Will as well?
> > > > 
> > > > Zhang Yi has observed the CPU lockup issue once when running heavy IO on
> > > > single nvme drive, and please CC him if you have new patch to try.
> > > 
> > > On which architecture? John was indicating that this also happen on x86.
> > 
> > ARM64.
> > 
> > To be honest, I never see such CPU lockup issue on x86 in case of running
> > heavy IO on single NVMe drive.
> > 
> > > 
> > > > Then looks the DMA unmap cost is too big on aarch64 if SMMU is involved.
> > > 
> > > So far, we don't have any data suggesting that this is actually the case.
> > > Also, other workloads (such as networking) do not exhibit this behaviour,
> > > while being least as unmap-heavy as NVMe is.
> > 
> > Maybe it is because networking workloads usually completes IO in softirq
> > context, instead of hard interrupt context.
> > 
> > > 
> > > If the cross-architecture aspect is confirmed, this points more into
> > > the direction of an interaction between the NVMe subsystem and the
> > > DMA API more than an architecture-specific problem.
> > > 
> > > Given that we have so far very little data, I'd hold off any conclusion.
> > 
> > We can start to collect latency data of dma unmapping vs nvme_irq()
> > on both x86 and arm64.
> > 
> > I will see if I can get a such box for collecting the latency data.
> 
> To reiterate what I mentioned before about IOMMU DMA unmap on x86, a key
> difference is that by default it uses the non-strict (lazy) mode unmap, i.e.
> we unmap in batches. ARM64 uses general default, which is strict mode, i.e.
> every unmap results in an IOTLB fluch.
> 
> In my setup, if I switch to lazy unmap (set iommu.strict=0 on cmdline), then
> no lockup.
> 
> Are any special IOMMU setups being used for x86, like enabling strict mode?
> I don't know...

BTW, I have run the test on one 224-core ARM64 with one 32-hw_queue NVMe, the
softlock issue can be triggered in one minute.

nvme_irq() often takes ~5us to complete on this machine, then there is really
risk of cpu lockup when IOPS is > 200K.

The soft lockup can be triggered too if 'iommu.strict=0' is passed in,
just takes a bit longer by starting more IO jobs.

In above test, I submit IO to one single NVMe drive from 4 CPU cores via 8 or
12 jobs(iommu.strict=0), meantime make the nvme interrupt handled just in one
dedicated CPU core.

Is there lock contention among iommu dma map and unmap callback?

Thanks,
Ming