Re: [Qemu-devel] [PATCH v7 2/6] virtio-pmem: Add virtio pmem driver

[Pankaj Gupta]

> 
> On 4/25/19 10:00 PM, Pankaj Gupta wrote:
> 
> > +void host_ack(struct virtqueue *vq)
> > +{
> > +   unsigned int len;
> > +   unsigned long flags;
> > +   struct virtio_pmem_request *req, *req_buf;
> > +   struct virtio_pmem *vpmem = vq->vdev->priv;
> > +
> > +   spin_lock_irqsave(&vpmem->pmem_lock, flags);
> > +   while ((req = virtqueue_get_buf(vq, &len)) != NULL) {
> > +           req->done = true;
> > +           wake_up(&req->host_acked);
> > +
> > +           if (!list_empty(&vpmem->req_list)) {
> > +                   req_buf = list_first_entry(&vpmem->req_list,
> > +                                   struct virtio_pmem_request, list);
> > +                   list_del(&vpmem->req_list);
> 
> Shouldn't it be rather `list_del(vpmem->req_list.next)`? We are trying to
> unlink
> first element of the list and `vpmem->req_list` is just the list head.

This looks correct. We are not deleting head but first entry in 'req_list'
which is device corresponding list of pending requests.

Please see below:

/**
 * Retrieve the first list entry for the given list pointer.
 *
 * Example:
 * struct foo *first;
 * first = list_first_entry(&bar->list_of_foos, struct foo, list_of_foos);
 *
 * @param ptr The list head
 * @param type Data type of the list element to retrieve
 * @param member Member name of the struct list_head field in the list element.
 * @return A pointer to the first list element.
 */
#define list_first_entry(ptr, type, member) \
    list_entry((ptr)->next, type, member)

> 
> > +int virtio_pmem_flush(struct nd_region *nd_region)
> > +{
> > +   int err;
> > +   unsigned long flags;
> > +   struct scatterlist *sgs[2], sg, ret;
> > +   struct virtio_device *vdev = nd_region->provider_data;
> > +   struct virtio_pmem *vpmem = vdev->priv;
> > +   struct virtio_pmem_request *req;
> > +
> > +   might_sleep();
> > +   req = kmalloc(sizeof(*req), GFP_KERNEL);
> > +   if (!req)
> > +           return -ENOMEM;
> > +
> > +   req->done = req->wq_buf_avail = false;
> > +   strcpy(req->name, "FLUSH");
> > +   init_waitqueue_head(&req->host_acked);
> > +   init_waitqueue_head(&req->wq_buf);
> > +   sg_init_one(&sg, req->name, strlen(req->name));
> > +   sgs[0] = &sg;
> > +   sg_init_one(&ret, &req->ret, sizeof(req->ret));
> > +   sgs[1] = &ret;
> > +
> > +   spin_lock_irqsave(&vpmem->pmem_lock, flags);
> > +   err = virtqueue_add_sgs(vpmem->req_vq, sgs, 1, 1, req, GFP_ATOMIC);
> > +   if (err) {
> > +           dev_err(&vdev->dev, "failed to send command to virtio pmem 
> > device\n");
> > +
> > +           list_add_tail(&vpmem->req_list, &req->list);
> > +           spin_unlock_irqrestore(&vpmem->pmem_lock, flags);
> > +
> > +           /* When host has read buffer, this completes via host_ack */
> > +           wait_event(req->wq_buf, req->wq_buf_avail);
> > +           spin_lock_irqsave(&vpmem->pmem_lock, flags);
> > +   }
> 
> Aren't the arguments in `list_add_tail` swapped? The element we are adding

No, this is intentional. 'vpmem->req_list' maintains a list of pending requests
for entire pmem device.  'req->list'is per request list and maintains pending
request on virtio queue add failure. I think we don't need this list.

> should
> be first, the list should be second. Also, shouldn't we resubmit the request
> after
> waking up from `wait_event(req->wq_buf, req->wq_buf_avail)`?

Yes. we should. Good point.

> 
> I propose rewriting it like that:
> 
> diff --git a/drivers/nvdimm/virtio_pmem.c b/drivers/nvdimm/virtio_pmem.c
> index 66b582f751a3..ff0556b04e86 100644
> --- a/drivers/nvdimm/virtio_pmem.c
> +++ b/drivers/nvdimm/virtio_pmem.c
> @@ -25,7 +25,7 @@ void host_ack(struct virtqueue *vq)
>               if (!list_empty(&vpmem->req_list)) {
>                       req_buf = list_first_entry(&vpmem->req_list,
>                                       struct virtio_pmem_request, list);
> -                     list_del(&vpmem->req_list);
> +                     list_del(vpmem->req_list.next);

Don't think its correct.

>                       req_buf->wq_buf_avail = true;
>                       wake_up(&req_buf->wq_buf);
>               }
> @@ -59,17 +59,33 @@ int virtio_pmem_flush(struct nd_region *nd_region)
>       sgs[1] = &ret;
>  
>       spin_lock_irqsave(&vpmem->pmem_lock, flags);
> -     err = virtqueue_add_sgs(vpmem->req_vq, sgs, 1, 1, req, GFP_ATOMIC);
> -     if (err) {
> -             dev_err(&vdev->dev, "failed to send command to virtio pmem 
> device\n");
> +     /*
> +      * If virtqueue_add_sgs returns -ENOSPC then req_vq virtual queue does 
> not
> +      * have free descriptor slots. We add the request to req_list and wait
> +      * for host_ack to wake us up when free slots are available.
> +      */
> +     while ((err = virtqueue_add_sgs(vpmem->req_vq, sgs, 1, 1, req, 
> GFP_ATOMIC))
> == -ENOSPC) {
> +             dev_err(&vdev->dev, "failed to send command to virtio pmem 
> device, no free
> slots in the virtqueue, postponing request\n");
> +             req->wq_buf_avail = false;
>  
> -             list_add_tail(&vpmem->req_list, &req->list);
> +             list_add_tail(&req->list, &vpmem->req_list);
>               spin_unlock_irqrestore(&vpmem->pmem_lock, flags);
>  
>               /* When host has read buffer, this completes via host_ack */
>               wait_event(req->wq_buf, req->wq_buf_avail);
>               spin_lock_irqsave(&vpmem->pmem_lock, flags);
>       }
> +
> +     /*
> +      * virtqueue_add_sgs failed with error different than -ENOSPC, we can't
> +      * do anything about that.
> +      */
> +     if (err) {
> +             dev_info(&vdev->dev, "failed to send command to virtio pmem 
> device, error
> code %d\n", err);
> +             spin_unlock_irqrestore(&vpmem->pmem_lock, flags);
> +             err = -EIO;
> +             goto ret;
> +     }
>       err = virtqueue_kick(vpmem->req_vq);
>       spin_unlock_irqrestore(&vpmem->pmem_lock, flags);
> 
> 
> Let me know if it looks reasonable to you.

Don't think this is fulfilling entire logic correctly. But thanks, I spotted a 
bug in my code :)
Will fix it. 

> 
> Thank you,
> Jakub Staron
> 
>