Re: [PATCH v4 02/26] iommu/sva: Manage process address spaces

[Jean-Philippe Brucker <jean-philippe@linaro.org>]

On Wed, Feb 26, 2020 at 12:35:06PM +0000, Jonathan Cameron wrote:
> > + * A single Process Address Space ID (PASID) is allocated for each mm. In the
> > + * example, devices use PASID 1 to read/write into address space X and PASID 2
> > + * to read/write into address space Y. Calling iommu_sva_get_pasid() on bond 1
> > + * returns 1, and calling it on bonds 2-4 returns 2.
> > + *
> > + * Hardware tables describing this configuration in the IOMMU would typically
> > + * look like this:
> > + *
> > + *                                PASID tables
> > + *                                 of domain A
> > + *                              .->+--------+
> > + *                             / 0 |        |-------> io_pgtable
> > + *                            /    +--------+
> > + *            Device tables  /   1 |        |-------> pgd X
> > + *              +--------+  /      +--------+
> > + *      00:00.0 |      A |-'     2 |        |--.
> > + *              +--------+         +--------+   \
> > + *              :        :       3 |        |    \
> > + *              +--------+         +--------+     --> pgd Y
> > + *      00:01.0 |      B |--.                    /
> > + *              +--------+   \                  |
> > + *      00:01.1 |      B |----+   PASID tables  |
> > + *              +--------+     \   of domain B  |
> > + *                              '->+--------+   |
> > + *                               0 |        |-- | --> io_pgtable
> > + *                                 +--------+   |
> > + *                               1 |        |   |
> > + *                                 +--------+   |
> > + *                               2 |        |---'
> > + *                                 +--------+
> > + *                               3 |        |
> > + *                                 +--------+
> > + *
> > + * With this model, a single call binds all devices in a given domain to an
> > + * address space. Other devices in the domain will get the same bond implicitly.
> > + * However, users must issue one bind() for each device, because IOMMUs may
> > + * implement SVA differently. Furthermore, mandating one bind() per device
> > + * allows the driver to perform sanity-checks on device capabilities.
> 
> > + *
> > + * In some IOMMUs, one entry of the PASID table (typically the first one) can
> > + * hold non-PASID translations. In this case PASID 0 is reserved and the first
> > + * entry points to the io_pgtable pointer. In other IOMMUs the io_pgtable
> > + * pointer is held in the device table and PASID 0 is available to the
> > + * allocator.
> 
> Is it worth hammering home in here that we can only do this because the PASID space
> is global (with exception of PASID 0)?  It's a convenient simplification but not
> necessarily a hardware restriction so perhaps we should remind people somewhere in here?

I could add this four paragraphs up:

"A single Process Address Space ID (PASID) is allocated for each mm. It is
a choice made for the Linux SVA implementation, not a hardware
restriction."

> > + */
> > +
> > +struct io_mm {
> > +	struct list_head		devices;
> > +	struct mm_struct		*mm;
> > +	struct mmu_notifier		notifier;
> > +
> > +	/* Late initialization */
> > +	const struct io_mm_ops		*ops;
> > +	void				*ctx;
> > +	int				pasid;
> > +};
> > +
> > +#define to_io_mm(mmu_notifier)	container_of(mmu_notifier, struct io_mm, notifier)
> > +#define to_iommu_bond(handle)	container_of(handle, struct iommu_bond, sva)
> 
> Code ordering wise, do we want this after the definition of iommu_bond?
> 
> For both of these it's a bit non obvious what they come 'from'.
> I wouldn't naturally assume to_io_mm gets me from notifier to the io_mm
> for example.  Not sure it matters though if these are only used in a few
> places.

Right, I can rename the first one to mn_to_io_mm(). The second one I think
might be good enough.


> > +static struct iommu_sva *
> > +io_mm_attach(struct device *dev, struct io_mm *io_mm, void *drvdata)
> > +{
> > +	int ret = 0;
> 
> I'm fairly sure this is set in all paths below.  Now, of course the
> compiler might not think that in which case fair enough :)
> 
> > +	bool attach_domain = true;
> > +	struct iommu_bond *bond, *tmp;
> > +	struct iommu_domain *domain, *other;
> > +	struct iommu_sva_param *param = dev->iommu_param->sva_param;
> > +
> > +	domain = iommu_get_domain_for_dev(dev);
> > +
> > +	bond = kzalloc(sizeof(*bond), GFP_KERNEL);
> > +	if (!bond)
> > +		return ERR_PTR(-ENOMEM);
> > +
> > +	bond->sva.dev	= dev;
> > +	bond->drvdata	= drvdata;
> > +	refcount_set(&bond->refs, 1);
> > +	RCU_INIT_POINTER(bond->io_mm, io_mm);
> > +
> > +	mutex_lock(&iommu_sva_lock);
> > +	/* Is it already bound to the device or domain? */
> > +	list_for_each_entry(tmp, &io_mm->devices, mm_head) {
> > +		if (tmp->sva.dev != dev) {
> > +			other = iommu_get_domain_for_dev(tmp->sva.dev);
> > +			if (domain == other)
> > +				attach_domain = false;
> > +
> > +			continue;
> > +		}
> > +
> > +		if (WARN_ON(tmp->drvdata != drvdata)) {
> > +			ret = -EINVAL;
> > +			goto err_free;
> > +		}
> > +
> > +		/*
> > +		 * Hold a single io_mm reference per bond. Note that we can't
> > +		 * return an error after this, otherwise the caller would drop
> > +		 * an additional reference to the io_mm.
> > +		 */
> > +		refcount_inc(&tmp->refs);
> > +		io_mm_put(io_mm);
> > +		kfree(bond);
> 
> Free outside the lock would be ever so slightly more logical given we allocated
> before taking the lock.
> 
> > +		mutex_unlock(&iommu_sva_lock);
> > +		return &tmp->sva;
> > +	}
> > +
> > +	list_add_rcu(&bond->mm_head, &io_mm->devices);
> > +	param->nr_bonds++;
> > +	mutex_unlock(&iommu_sva_lock);
> > +
> > +	ret = io_mm->ops->attach(bond->sva.dev, io_mm->pasid, io_mm->ctx,
> > +				 attach_domain);
> > +	if (ret)
> > +		goto err_remove;
> > +
> > +	return &bond->sva;
> > +
> > +err_remove:
> > +	/*
> > +	 * At this point concurrent threads may have started to access the
> > +	 * io_mm->devices list in order to invalidate address ranges, which
> > +	 * requires to free the bond via kfree_rcu()
> > +	 */
> > +	mutex_lock(&iommu_sva_lock);
> > +	param->nr_bonds--;
> > +	list_del_rcu(&bond->mm_head);
> > +
> > +err_free:
> > +	mutex_unlock(&iommu_sva_lock);
> > +	kfree_rcu(bond, rcu_head);
> 
> I don't suppose it matters really but we don't need the rcu free if
> we follow the err_free goto.  Perhaps we are cleaner in this case
> to not use a unified exit path but do that case inline?

Agreed, though I moved the kzalloc() later as suggested by Jacob, I think
it looks a little better and simplifies the error paths

Thanks,
Jean