Re: [PATCH v2 4/6] iommufd: Deliver fault messages to user space

[Baolu Lu]

On 2023/12/1 23:24, Jason Gunthorpe wrote:
> On Thu, Oct 26, 2023 at 10:49:28AM +0800, Lu Baolu wrote:
>
> > +static ssize_t hwpt_fault_fops_write(struct file *filep,
> > +				     const char __user *buf,
> > +				     size_t count, loff_t *ppos)
> > +{
> > +	size_t response_size = sizeof(struct iommu_hwpt_page_response);
> > +	struct hw_pgtable_fault *fault = filep->private_data;
> > +	struct iommu_hwpt_page_response response;
> > +	struct iommufd_hw_pagetable *hwpt;
> > +	struct iopf_group *iter, *group;
> > +	struct iommufd_device *idev;
> > +	size_t done = 0;
> > +	int rc = 0;
> > +
> > +	if (*ppos || count % response_size)
> > +		return -ESPIPE;
> > +
> > +	mutex_lock(&fault->mutex);
> > +	while (!list_empty(&fault->response) && count > done) {
> > +		rc = copy_from_user(&response, buf + done, response_size);
> > +		if (rc)
> > +			break;
> > +
> > +		/* Get the device that this response targets at. */
> > +		idev = container_of(iommufd_get_object(fault->ictx,
> > +						       response.dev_id,
> > +						       IOMMUFD_OBJ_DEVICE),
> > +				    struct iommufd_device, obj);
> > +		if (IS_ERR(idev)) {
> > +			rc = PTR_ERR(idev);
> > +			break;
> > +		}
>
> See here it might be better to have a per-fd list of outstanding
> faults per-fd and then the cookie would just index that list, then you
> get everything in one shot instead of having to do a xarray looking
> and then a linear list search

Yours is more efficient. I will do it that way in the next version.

> > +static const struct file_operations hwpt_fault_fops = {
> > +	.owner		= THIS_MODULE,
> > +	.read		= hwpt_fault_fops_read,
> > +	.write		= hwpt_fault_fops_write,
> > +};
>
> nonseekable_open() behavior should be integrated into this

Sure.

> > +static int hw_pagetable_get_fault_fd(struct hw_pgtable_fault *fault)
> > +{
> > +	struct file *filep;
> > +	int fdno;
> > +
> > +	fdno = get_unused_fd_flags(O_CLOEXEC);
> > +	if (fdno < 0)
> > +		return fdno;
> > +
> > +	filep = anon_inode_getfile("[iommufd-pgfault]", &hwpt_fault_fops,
> > +				   fault, O_RDWR);
> > +	if (IS_ERR(filep)) {
> > +		put_unused_fd(fdno);
> > +		return PTR_ERR(filep);
> > +	}
> > +
> > +	fd_install(fdno, filep);
> > +	fault->fault_file = filep;
> > +	fault->fault_fd = fdno;
>
> fd_install must be the very last thing before returning success from a
> system call because we cannot undo it.

Yes.

> There are other failure paths before here and the final return
>
> Jason

Best regards,
baolu