Re: [RFC PATCH v3 5/8] vfio/type1: VFIO_IOMMU_ENABLE_IOPF

[Alex Williamson <alex.williamson@redhat.com>]

On Fri, 21 May 2021 14:38:25 +0800
Shenming Lu <lushenming@huawei.com> wrote:

> On 2021/5/19 2:58, Alex Williamson wrote:
> > On Fri, 9 Apr 2021 11:44:17 +0800
> > Shenming Lu <lushenming@huawei.com> wrote:
> >   
> >> Since enabling IOPF for devices may lead to a slow ramp up of performance,
> >> we add an ioctl VFIO_IOMMU_ENABLE_IOPF to make it configurable. And the
> >> IOPF enabling of a VFIO device includes setting IOMMU_DEV_FEAT_IOPF and
> >> registering the VFIO IOPF handler.
> >>
> >> Note that VFIO_IOMMU_DISABLE_IOPF is not supported since there may be
> >> inflight page faults when disabling.
> >>
> >> Signed-off-by: Shenming Lu <lushenming@huawei.com>
> >> ---
> >>  drivers/vfio/vfio_iommu_type1.c | 223 +++++++++++++++++++++++++++++++-
> >>  include/uapi/linux/vfio.h       |   6 +
> >>  2 files changed, 226 insertions(+), 3 deletions(-)
> >>
> >> diff --git a/drivers/vfio/vfio_iommu_type1.c b/drivers/vfio/vfio_iommu_type1.c
> >> index 01e296c6dc9e..7df5711e743a 100644
> >> --- a/drivers/vfio/vfio_iommu_type1.c
> >> +++ b/drivers/vfio/vfio_iommu_type1.c
> >> @@ -71,6 +71,7 @@ struct vfio_iommu {
> >>  	struct rb_root		dma_list;
> >>  	struct blocking_notifier_head notifier;
> >>  	struct mmu_notifier	mn;
> >> +	struct mm_struct	*mm;  
> > 
> > We currently have no requirement that a single mm is used for all
> > container mappings.  Does enabling IOPF impose such a requirement?
> > Shouldn't MAP/UNMAP enforce that?  
> 
> Did you mean that there is a possibility that each vfio_dma in a
> container has a different mm? If so, could we register one MMU
> notifier for each vfio_dma in the MAP ioctl?

We don't prevent it currently.  There needs to be some balance,
typically we'll have one mm, so would it make sense to have potentially
thousands of mmu notifiers registered against the same mm?
 
> >>  	unsigned int		dma_avail;
> >>  	unsigned int		vaddr_invalid_count;
> >>  	uint64_t		pgsize_bitmap;
> >> @@ -81,6 +82,7 @@ struct vfio_iommu {
> >>  	bool			dirty_page_tracking;
> >>  	bool			pinned_page_dirty_scope;
> >>  	bool			container_open;
> >> +	bool			iopf_enabled;
> >>  };
> >>  
> >>  struct vfio_domain {
> >> @@ -461,6 +463,38 @@ vfio_find_iopf_group(struct iommu_group *iommu_group)
> >>  	return node ? iopf_group : NULL;
> >>  }
> >>  
> >> +static void vfio_link_iopf_group(struct vfio_iopf_group *new)
> >> +{
> >> +	struct rb_node **link, *parent = NULL;
> >> +	struct vfio_iopf_group *iopf_group;
> >> +
> >> +	mutex_lock(&iopf_group_list_lock);
> >> +
> >> +	link = &iopf_group_list.rb_node;
> >> +
> >> +	while (*link) {
> >> +		parent = *link;
> >> +		iopf_group = rb_entry(parent, struct vfio_iopf_group, node);
> >> +
> >> +		if (new->iommu_group < iopf_group->iommu_group)
> >> +			link = &(*link)->rb_left;
> >> +		else
> >> +			link = &(*link)->rb_right;
> >> +	}
> >> +
> >> +	rb_link_node(&new->node, parent, link);
> >> +	rb_insert_color(&new->node, &iopf_group_list);
> >> +
> >> +	mutex_unlock(&iopf_group_list_lock);
> >> +}
> >> +
> >> +static void vfio_unlink_iopf_group(struct vfio_iopf_group *old)
> >> +{
> >> +	mutex_lock(&iopf_group_list_lock);
> >> +	rb_erase(&old->node, &iopf_group_list);
> >> +	mutex_unlock(&iopf_group_list_lock);
> >> +}
> >> +
> >>  static int vfio_lock_acct(struct vfio_dma *dma, long npage, bool async)
> >>  {
> >>  	struct mm_struct *mm;
> >> @@ -2363,6 +2397,68 @@ static void vfio_iommu_iova_insert_copy(struct vfio_iommu *iommu,
> >>  	list_splice_tail(iova_copy, iova);
> >>  }
> >>  
> >> +static int vfio_dev_domian_nested(struct device *dev, int *nested)  
> > 
> > 
> > s/domian/domain/
> > 
> >   
> >> +{
> >> +	struct iommu_domain *domain;
> >> +
> >> +	domain = iommu_get_domain_for_dev(dev);
> >> +	if (!domain)
> >> +		return -ENODEV;
> >> +
> >> +	return iommu_domain_get_attr(domain, DOMAIN_ATTR_NESTING, nested);  
> > 
> > 
> > Wouldn't this be easier to use if it returned bool, false on -errno?  
> 
> Make sense.
> 
> > 
> >   
> >> +}
> >> +
> >> +static int vfio_iommu_type1_dma_map_iopf(struct iommu_fault *fault, void *data);
> >> +
> >> +static int dev_enable_iopf(struct device *dev, void *data)
> >> +{
> >> +	int *enabled_dev_cnt = data;
> >> +	int nested;
> >> +	u32 flags;
> >> +	int ret;
> >> +
> >> +	ret = iommu_dev_enable_feature(dev, IOMMU_DEV_FEAT_IOPF);
> >> +	if (ret)
> >> +		return ret;
> >> +
> >> +	ret = vfio_dev_domian_nested(dev, &nested);
> >> +	if (ret)
> >> +		goto out_disable;
> >> +
> >> +	if (nested)
> >> +		flags = FAULT_REPORT_NESTED_L2;
> >> +	else
> >> +		flags = FAULT_REPORT_FLAT;
> >> +
> >> +	ret = iommu_register_device_fault_handler(dev,
> >> +				vfio_iommu_type1_dma_map_iopf, flags, dev);
> >> +	if (ret)
> >> +		goto out_disable;
> >> +
> >> +	(*enabled_dev_cnt)++;
> >> +	return 0;
> >> +
> >> +out_disable:
> >> +	iommu_dev_disable_feature(dev, IOMMU_DEV_FEAT_IOPF);
> >> +	return ret;
> >> +}
> >> +
> >> +static int dev_disable_iopf(struct device *dev, void *data)
> >> +{
> >> +	int *enabled_dev_cnt = data;
> >> +
> >> +	if (enabled_dev_cnt && *enabled_dev_cnt <= 0)
> >> +		return -1;  
> > 
> > 
> > Use an errno.>
> >   
> >> +
> >> +	WARN_ON(iommu_unregister_device_fault_handler(dev));
> >> +	WARN_ON(iommu_dev_disable_feature(dev, IOMMU_DEV_FEAT_IOPF));
> >> +
> >> +	if (enabled_dev_cnt)
> >> +		(*enabled_dev_cnt)--;  
> > 
> > 
> > Why do we need these counters?  
> 
> We use this counter to record the number of IOPF enabled devices, and if
> any device fails to be enabled, we will exit the loop (three nested loop)
> and go to the unwind process, which will disable the first @enabled_dev_cnt
> devices.
> 
> >   
> >> +
> >> +	return 0;
> >> +}
> >> +
> >>  static int vfio_iommu_type1_attach_group(void *iommu_data,
> >>  					 struct iommu_group *iommu_group)
> >>  {
> >> @@ -2376,6 +2472,8 @@ static int vfio_iommu_type1_attach_group(void *iommu_data,
> >>  	struct iommu_domain_geometry geo;
> >>  	LIST_HEAD(iova_copy);
> >>  	LIST_HEAD(group_resv_regions);
> >> +	int iopf_enabled_dev_cnt = 0;
> >> +	struct vfio_iopf_group *iopf_group = NULL;
> >>  
> >>  	mutex_lock(&iommu->lock);
> >>  
> >> @@ -2453,6 +2551,24 @@ static int vfio_iommu_type1_attach_group(void *iommu_data,
> >>  	if (ret)
> >>  		goto out_domain;
> >>  
> >> +	if (iommu->iopf_enabled) {
> >> +		ret = iommu_group_for_each_dev(iommu_group, &iopf_enabled_dev_cnt,
> >> +					       dev_enable_iopf);
> >> +		if (ret)
> >> +			goto out_detach;
> >> +
> >> +		iopf_group = kzalloc(sizeof(*iopf_group), GFP_KERNEL);
> >> +		if (!iopf_group) {
> >> +			ret = -ENOMEM;
> >> +			goto out_detach;
> >> +		}
> >> +
> >> +		iopf_group->iommu_group = iommu_group;
> >> +		iopf_group->iommu = iommu;
> >> +
> >> +		vfio_link_iopf_group(iopf_group);  
> > 
> > This seems backwards, once we enable iopf we can take a fault, so the
> > structure to lookup the data for the device should be setup first.  I'm
> > still not convinced this iopf_group rb tree is a good solution to get
> > from the device to the iommu either.  vfio-core could traverse dev ->
> > iommu_group -> vfio_group -> container -> iommu_data.  
> 
> Make sense.
> 
> > 
> >   
> >> +	}
> >> +
> >>  	/* Get aperture info */
> >>  	iommu_domain_get_attr(domain->domain, DOMAIN_ATTR_GEOMETRY, &geo);
> >>  
> >> @@ -2534,9 +2650,11 @@ static int vfio_iommu_type1_attach_group(void *iommu_data,
> >>  	vfio_test_domain_fgsp(domain);
> >>  
> >>  	/* replay mappings on new domains */
> >> -	ret = vfio_iommu_replay(iommu, domain);
> >> -	if (ret)
> >> -		goto out_detach;
> >> +	if (!iommu->iopf_enabled) {
> >> +		ret = vfio_iommu_replay(iommu, domain);
> >> +		if (ret)
> >> +			goto out_detach;
> >> +	}  
> > 
> > Test in vfio_iommu_replay()?  
> 
> Not yet, I will do more tests later.
> 
> >   
> >>  
> >>  	if (resv_msi) {
> >>  		ret = iommu_get_msi_cookie(domain->domain, resv_msi_base);
> >> @@ -2567,6 +2685,15 @@ static int vfio_iommu_type1_attach_group(void *iommu_data,
> >>  	iommu_domain_free(domain->domain);
> >>  	vfio_iommu_iova_free(&iova_copy);
> >>  	vfio_iommu_resv_free(&group_resv_regions);
> >> +	if (iommu->iopf_enabled) {
> >> +		if (iopf_group) {
> >> +			vfio_unlink_iopf_group(iopf_group);
> >> +			kfree(iopf_group);
> >> +		}
> >> +
> >> +		iommu_group_for_each_dev(iommu_group, &iopf_enabled_dev_cnt,
> >> +					 dev_disable_iopf);  
> > 
> > Backwards, disable fault reporting before unlinking lookup.  
> 
> Make sense.
> 
> >   
> >> +	}
> >>  out_free:
> >>  	kfree(domain);
> >>  	kfree(group);
> >> @@ -2728,6 +2855,19 @@ static void vfio_iommu_type1_detach_group(void *iommu_data,
> >>  		if (!group)
> >>  			continue;
> >>  
> >> +		if (iommu->iopf_enabled) {
> >> +			struct vfio_iopf_group *iopf_group;
> >> +
> >> +			iopf_group = vfio_find_iopf_group(iommu_group);
> >> +			if (!WARN_ON(!iopf_group)) {
> >> +				vfio_unlink_iopf_group(iopf_group);
> >> +				kfree(iopf_group);
> >> +			}
> >> +
> >> +			iommu_group_for_each_dev(iommu_group, NULL,
> >> +						 dev_disable_iopf);  
> > 
> > 
> > Backwards.  Also unregistering the fault handler can fail if there are
> > pending faults.  It appears the fault handler and iopf lookup could also
> > race with this function, ex. fault handler gets an iopf object before
> > it's removed from the rb-tree, blocks trying to acquire iommu->lock,
> > disable_iopf fails, detach proceeds, fault handler has use after free
> > error.  
> 
> We have WARN_ON for the failure of the unregistering.
> 
> Yeah, it seems that using vfio_iopf_group is not a good choice...
> 
> > 
> >   
> >> +		}
> >> +
> >>  		vfio_iommu_detach_group(domain, group);
> >>  		update_dirty_scope = !group->pinned_page_dirty_scope;
> >>  		list_del(&group->next);
> >> @@ -2846,6 +2986,11 @@ static void vfio_iommu_type1_release(void *iommu_data)
> >>  
> >>  	vfio_iommu_iova_free(&iommu->iova_list);
> >>  
> >> +	if (iommu->iopf_enabled) {
> >> +		mmu_notifier_unregister(&iommu->mn, iommu->mm);
> >> +		mmdrop(iommu->mm);
> >> +	}
> >> +
> >>  	kfree(iommu);
> >>  }
> >>  
> >> @@ -3441,6 +3586,76 @@ static const struct mmu_notifier_ops vfio_iommu_type1_mn_ops = {
> >>  	.invalidate_range	= mn_invalidate_range,
> >>  };
> >>  
> >> +static int vfio_iommu_type1_enable_iopf(struct vfio_iommu *iommu)
> >> +{
> >> +	struct vfio_domain *d;
> >> +	struct vfio_group *g;
> >> +	struct vfio_iopf_group *iopf_group;
> >> +	int enabled_dev_cnt = 0;
> >> +	int ret;
> >> +
> >> +	if (!current->mm)
> >> +		return -ENODEV;
> >> +
> >> +	mutex_lock(&iommu->lock);
> >> +
> >> +	mmgrab(current->mm);  
> > 
> > 
> > As above, this is a new and undocumented requirement.
> > 
> >   
> >> +	iommu->mm = current->mm;
> >> +	iommu->mn.ops = &vfio_iommu_type1_mn_ops;
> >> +	ret = mmu_notifier_register(&iommu->mn, current->mm);
> >> +	if (ret)
> >> +		goto out_drop;
> >> +
> >> +	list_for_each_entry(d, &iommu->domain_list, next) {
> >> +		list_for_each_entry(g, &d->group_list, next) {
> >> +			ret = iommu_group_for_each_dev(g->iommu_group,
> >> +					&enabled_dev_cnt, dev_enable_iopf);
> >> +			if (ret)
> >> +				goto out_unwind;
> >> +
> >> +			iopf_group = kzalloc(sizeof(*iopf_group), GFP_KERNEL);
> >> +			if (!iopf_group) {
> >> +				ret = -ENOMEM;
> >> +				goto out_unwind;
> >> +			}
> >> +
> >> +			iopf_group->iommu_group = g->iommu_group;
> >> +			iopf_group->iommu = iommu;
> >> +
> >> +			vfio_link_iopf_group(iopf_group);
> >> +		}
> >> +	}
> >> +
> >> +	iommu->iopf_enabled = true;
> >> +	goto out_unlock;
> >> +
> >> +out_unwind:
> >> +	list_for_each_entry(d, &iommu->domain_list, next) {
> >> +		list_for_each_entry(g, &d->group_list, next) {
> >> +			iopf_group = vfio_find_iopf_group(g->iommu_group);
> >> +			if (iopf_group) {
> >> +				vfio_unlink_iopf_group(iopf_group);
> >> +				kfree(iopf_group);
> >> +			}
> >> +
> >> +			if (iommu_group_for_each_dev(g->iommu_group,
> >> +					&enabled_dev_cnt, dev_disable_iopf))
> >> +				goto out_unregister;  
> > 
> > 
> > This seems broken, we break out of the unwind if any group_for_each
> > fails??  
> 
> The iommu_group_for_each_dev function will return a negative value if
> we have disabled all previously enabled devices (@enabled_dev_cnt).

What's the harm in calling disable IOPF on a device where it's already
disabled?  This is why I'm not sure the value of keeping a counter.  It
also presumes that for_each_dev iterates in the same order every time.
 
> >> +		}
> >> +	}
> >> +
> >> +out_unregister:
> >> +	mmu_notifier_unregister(&iommu->mn, current->mm);
> >> +
> >> +out_drop:
> >> +	iommu->mm = NULL;
> >> +	mmdrop(current->mm);
> >> +
> >> +out_unlock:
> >> +	mutex_unlock(&iommu->lock);
> >> +	return ret;
> >> +}  
> > 
> > 
> > Is there an assumption that the user does this before performing any
> > mapping?  I don't see where current mappings would be converted which
> > means we'll have pinned pages leaked and vfio_dma objects without a
> > pinned page bitmap.  
> 
> Yeah, we have an assumption that this ioctl (ENABLE_IOPF) must be called
> before any MAP/UNMAP ioctl...
> 
> I will document these later. :-)


That can't be fixed with documentation, if it's a requirement, the code
needs to enforce it.  But also why is it a requirement?  Theoretically
we could unpin and de-account existing mappings and allow the mmu
notifier to handle these as well, right?  Thanks,

Alex