Re: [PATCH 1/8] iommu: Introduce a replace API for device pasid

From: Yi Liu
Date: Thu Mar 21 2024 - 07:23:41 EST

On 2024/3/21 14:16, Yi Liu wrote:
On 2024/3/20 20:38, Jason Gunthorpe wrote:
On Tue, Mar 19, 2024 at 03:29:39PM +0800, Yi Liu wrote:
On 2024/3/19 00:52, Jason Gunthorpe wrote:
On Wed, Mar 13, 2024 at 04:11:41PM +0800, Yi Liu wrote:

yes. how about your opinion? @Jason. I noticed the set_dev_pasid callback
and pasid_array update is under the group->lock, so update it should be
fine to adjust the order to update pasid_array after set_dev_pasid returns.

Yes, it makes some sense

But, also I would like it very much if we just have the core pass in
the actual old domain as a an addition function argument.

ok, this works too. For normal attach, just pass in a NULL old domain.

I think we have some small mistakes in multi-device group error
unwinding for remove because the global xarray can't isn't actually
going to be correct in all scenarios.

do you mean the __iommu_remove_group_pasid() call in the below function?
Currently, it is called when __iommu_set_group_pasid() failed. However,
__iommu_set_group_pasid() may need to do remove itself when error happens,
so the helper can be more self-contained. Or you mean something else?

Yes..

int iommu_attach_device_pasid(struct iommu_domain *domain,
                  struct device *dev, ioasid_t pasid)
{
    /* Caller must be a probed driver on dev */
    struct iommu_group *group = dev->iommu_group;
    void *curr;
    int ret;

    if (!domain->ops->set_dev_pasid)
        return -EOPNOTSUPP;

    if (!group)
        return -ENODEV;

    if (!dev_has_iommu(dev) || dev_iommu_ops(dev) != domain->owner)
        return -EINVAL;

    mutex_lock(&group->mutex);
    curr = xa_cmpxchg(&group->pasid_array, pasid, NULL, domain, GFP_KERNEL);
    if (curr) {
        ret = xa_err(curr) ? : -EBUSY;
        goto out_unlock;
    }

    ret = __iommu_set_group_pasid(domain, group, pasid);

So here we have the xa set to the new domain

aha, yes. The underlying driver won't be able to get the correct domain
in the remove_dev_pasid callback.

    if (ret) {
        __iommu_remove_group_pasid(group, pasid);

And here we still have it set to the new domain even though some of
the devices within the group failed to attach. The logic needs to be
more like the main domain attach path where iterate and then undo only
what we did

yes, the correct way is to undo what have been done before the fail
device. However, I somehow remember that pasid capability is only
available when the group is singleton. So iterate all devices of the
devices just means one device in fact. If this is true, then the
current code is fine although a bit confusing.

And the whole thing is easier to reason about if an input argument
specifies the current attached domain instead of having the driver
read it from the xarray.

yep, will correct it as a fix patch.

Hi Jason,

It appears there are two solutions here.

First, only undo the devices that have set_dev_pasid successfully in
the __iommu_set_group_pasid(), so the problematic
__iommu_remove_group_pasid() call at line 3378 [1] would go away.
This also makes the helper more self-contained. Draft patch in [2]

Second, pass in the domain to remove_dev_pasid(). Draft patch in [3]

Either of the above two should be able to solve the mistake you mentioned.
BTW. They are orthogonal, so it's also possible to apply both of them.
Which one is your preference then?

[1] https://github.com/torvalds/linux/blob/b3603fcb79b1036acae10602bffc4855a4b9af80/drivers/iommu/iommu.c#L3378
[2] https://github.com/yiliu1765/iommufd/commit/bb83270767ab460978a3452750f5e032b43e59bf
[3] https://github.com/yiliu1765/iommufd/commit/bd270b7b6619b8ba35dfaf9870df8728e370163f

Regards,
Yi Liu