Re: [PATCH 10/10] vfio/qat: Add vfio_pci driver for Intel QAT VF devices

From: Jason Gunthorpe <jgg@nvidia.com>
To: "Zeng, Xin" <xin.zeng@intel.com>
Cc: "herbert@gondor.apana.org.au" <herbert@gondor.apana.org.au>,
	"alex.williamson@redhat.com" <alex.williamson@redhat.com>,
	"yishaih@nvidia.com" <yishaih@nvidia.com>,
	"shameerali.kolothum.thodi@huawei.com"
	<shameerali.kolothum.thodi@huawei.com>,
	"Tian, Kevin" <kevin.tian@intel.com>,
	"linux-crypto@vger.kernel.org" <linux-crypto@vger.kernel.org>,
	"kvm@vger.kernel.org" <kvm@vger.kernel.org>,
	qat-linux <qat-linux@intel.com>,
	"Cao, Yahui" <yahui.cao@intel.com>
Subject: Re: [PATCH 10/10] vfio/qat: Add vfio_pci driver for Intel QAT VF devices
Date: Fri, 9 Feb 2024 08:10:45 -0400	[thread overview]
Message-ID: <20240209121045.GP10476@nvidia.com> (raw)
In-Reply-To: <DM4PR11MB550222F7A5454DF9DBEE7FEC884B2@DM4PR11MB5502.namprd11.prod.outlook.com>

On Fri, Feb 09, 2024 at 08:23:32AM +0000, Zeng, Xin wrote:
> Thanks for your comments, Jason.
> On Tuesday, February 6, 2024 8:55 PM, Jason Gunthorpe <jgg@nvidia.com> wrote:
> > > +
> > > +	ops = mdev->ops;
> > > +	if (!ops || !ops->init || !ops->cleanup ||
> > > +	    !ops->open || !ops->close ||
> > > +	    !ops->save_state || !ops->load_state ||
> > > +	    !ops->suspend || !ops->resume) {
> > > +		ret = -EIO;
> > > +		dev_err(&parent->dev, "Incomplete device migration ops
> > structure!");
> > > +		goto err_destroy;
> > > +	}
> > 
> > Why are there ops pointers here? I would expect this should just be
> > direct function calls to the PF QAT driver.
> 
> I indeed had a version where the direct function calls are Implemented in
> QAT driver, while when I look at the functions, most of them 
> only translate the interface to the ops pointer. That's why I put
> ops pointers directly into vfio variant driver.

But why is there an ops indirection at all? Are there more than one
ops?

> > > +static void qat_vf_pci_aer_reset_done(struct pci_dev *pdev)
> > > +{
> > > +	struct qat_vf_core_device *qat_vdev = qat_vf_drvdata(pdev);
> > > +
> > > +	if (!qat_vdev->core_device.vdev.mig_ops)
> > > +		return;
> > > +
> > > +	/*
> > > +	 * As the higher VFIO layers are holding locks across reset and using
> > > +	 * those same locks with the mm_lock we need to prevent ABBA
> > deadlock
> > > +	 * with the state_mutex and mm_lock.
> > > +	 * In case the state_mutex was taken already we defer the cleanup work
> > > +	 * to the unlock flow of the other running context.
> > > +	 */
> > > +	spin_lock(&qat_vdev->reset_lock);
> > > +	qat_vdev->deferred_reset = true;
> > > +	if (!mutex_trylock(&qat_vdev->state_mutex)) {
> > > +		spin_unlock(&qat_vdev->reset_lock);
> > > +		return;
> > > +	}
> > > +	spin_unlock(&qat_vdev->reset_lock);
> > > +	qat_vf_state_mutex_unlock(qat_vdev);
> > > +}
> > 
> > Do you really need this? I thought this ugly thing was going to be a
> > uniquely mlx5 thing..
> 
> I think that's still required to make the migration state synchronized
> if the VF is reset by other VFIO emulation paths. Is it the case? 
> BTW, this implementation is not only in mlx5 driver, but also in other
> Vfio pci variant drivers such as hisilicon acc driver and pds
> driver.

It had to specifically do with the mm lock interaction that, I
thought, was going to be unique to the mlx driver. Otherwise you could
just directly hold the state_mutex here.

Yishai do you remember the exact trace for the mmlock entanglement?

Jason