Re: [PATCH v3 4/8] iommufd: Add iommufd fault object

[Baolu Lu]

On 3/23/24 1:22 AM, Jason Gunthorpe wrote:
> On Wed, Mar 20, 2024 at 04:18:05PM +0000, Shameerali Kolothum Thodi wrote:
> > What I have noticed is that,
> > -read interface works fine and I can receive struct tiommu_hwpt_pgfault data.
> > -But once Guest handles the page faults and returns the page response,
> >   the write to fault fd never reaches the kernel. The sequence is like below,
> >   
> >    sqe = io_uring_get_sqe(ring);
> >    io_uring_prep_write(sqe, hwpt->fault_fd, resp, sizeof(*resp), 0);
> >    io_uring_sqe_set_data(sqe, resp);
> >    io_uring_submit(ring);
> >    ret = io_uring_wait_cqe(ring, &cqe);
> >    ....
> > Please find the function here[2]
> >
> > The above cqe wait never returns and hardware times out without receiving
> > page response. My understanding of io_uring default op is that it tries to
> > issue an sqe as non-blocking first. But it looks like the above write sequence
> > ends up in kernel poll_wait() as well.Not sure how we can avoid that for
> > write.
> Ah, right, it is because poll can't be choosy about read/write, it has
> to work equally for both directions. iommufd_fault_fops_poll() never
> returns EPOLLOUT
>
> It should just always return EPOLLOUT because we don't have any queue
> to manage.

Are you suggesting the poll file operation to be like below?

static __poll_t iommufd_fault_fops_poll(struct file *filep,
                                        struct poll_table_struct *wait)
{
        struct iommufd_fault *fault = filep->private_data;
        __poll_t pollflags = EPOLLOUT;

        poll_wait(filep, &fault->wait_queue, wait);
        mutex_lock(&fault->mutex);
        if (!list_empty(&fault->deliver))
                pollflags = EPOLLIN | EPOLLRDNORM;
        mutex_unlock(&fault->mutex);

        return pollflags;
}

The diff is,

diff --git a/drivers/iommu/iommufd/fault.c b/drivers/iommu/iommufd/fault.c
index ede16702d433..a33f8aa92575 100644
--- a/drivers/iommu/iommufd/fault.c
+++ b/drivers/iommu/iommufd/fault.c
@@ -175,7 +175,7 @@ static __poll_t iommufd_fault_fops_poll(struct file 
*filep,
                                        struct poll_table_struct *wait)
 {
        struct iommufd_fault *fault = filep->private_data;
-       __poll_t pollflags = 0;
+       __poll_t pollflags = EPOLLOUT;

        poll_wait(filep, &fault->wait_queue, wait);
        mutex_lock(&fault->mutex);


I was originally thinking that poll file operation is specifically
designed for polling on read events associated with IOMMU faults.

Best regards,
baolu