On Mon, Aug 12, 2024 at 1:47 PM Jason Wang [off-list ref] wrote:

On Fri, Aug 9, 2024 at 2:04 AM Dragos Tatulea [off-list ref] wrote:

quoted

On 08.08.24 10:20, Jason Wang wrote:

quoted

We used to call irq_bypass_unregister_producer() in
vhost_vdpa_setup_vq_irq() which is problematic as we don't know if the
token pointer is still valid or not.

Actually, we use the eventfd_ctx as the token so the life cycle of the
token should be bound to the VHOST_SET_VRING_CALL instead of
vhost_vdpa_setup_vq_irq() which could be called by set_status().

Fixing this by setting up  irq bypass producer's token when handling
VHOST_SET_VRING_CALL and un-registering the producer before calling
vhost_vring_ioctl() to prevent a possible use after free as eventfd
could have been released in vhost_vring_ioctl().

Fixes: 2cf1ba9a4d15 ("vhost_vdpa: implement IRQ offloading in vhost_vdpa")
Signed-off-by: Jason Wang <jasowang@redhat.com>
---
Note for Dragos: Please check whether this fixes your issue. I
slightly test it with vp_vdpa in L2.
---
 drivers/vhost/vdpa.c | 12 +++++++++---
 1 file changed, 9 insertions(+), 3 deletions(-)

diff --git a/drivers/vhost/vdpa.c b/drivers/vhost/vdpa.c
index e31ec9ebc4ce..388226a48bcc 100644
--- a/drivers/vhost/vdpa.c
+++ b/drivers/vhost/vdpa.c

@@ -209,11 +209,9 @@ static void vhost_vdpa_setup_vq_irq(struct vhost_vdpa *v, u16 qid)
      if (irq < 0)
              return;

-     irq_bypass_unregister_producer(&vq->call_ctx.producer);
      if (!vq->call_ctx.ctx)
              return;

-     vq->call_ctx.producer.token = vq->call_ctx.ctx;
      vq->call_ctx.producer.irq = irq;
      ret = irq_bypass_register_producer(&vq->call_ctx.producer);
      if (unlikely(ret))

@@ -709,6 +707,12 @@ static long vhost_vdpa_vring_ioctl(struct vhost_vdpa *v, unsigned int cmd,
                      vq->last_avail_idx = vq_state.split.avail_index;
              }
              break;
+     case VHOST_SET_VRING_CALL:
+             if (vq->call_ctx.ctx) {
+                     vhost_vdpa_unsetup_vq_irq(v, idx);
+                     vq->call_ctx.producer.token = NULL;
+             }
+             break;
      }

      r = vhost_vring_ioctl(&v->vdev, cmd, argp);

@@ -747,13 +751,14 @@ static long vhost_vdpa_vring_ioctl(struct vhost_vdpa *v, unsigned int cmd,
                      cb.callback = vhost_vdpa_virtqueue_cb;
                      cb.private = vq;
                      cb.trigger = vq->call_ctx.ctx;
+                     vq->call_ctx.producer.token = vq->call_ctx.ctx;
+                     vhost_vdpa_setup_vq_irq(v, idx);
              } else {
                      cb.callback = NULL;
                      cb.private = NULL;
                      cb.trigger = NULL;
              }
              ops->set_vq_cb(vdpa, idx, &cb);
-             vhost_vdpa_setup_vq_irq(v, idx);
              break;

      case VHOST_SET_VRING_NUM:

@@ -1419,6 +1424,7 @@ static int vhost_vdpa_open(struct inode *inode, struct file *filep)
      for (i = 0; i < nvqs; i++) {
              vqs[i] = &v->vqs[i];
              vqs[i]->handle_kick = handle_vq_kick;
+             vqs[i]->call_ctx.ctx = NULL;
      }
      vhost_dev_init(dev, vqs, nvqs, 0, 0, 0, false,
                     vhost_vdpa_process_iotlb_msg);

No more crashes, but now getting a lot of:
 vhost-vdpa-X: vq Y, irq bypass producer (token 00000000a66e28ab) registration fails, ret =  -16

... seems like the irq_bypass_unregister_producer() that was removed
might still be needed somewhere?

Probably, but I didn't see this when testing vp_vdpa.

When did you meet those warnings? Is it during the boot or migration?

Btw, it would be helpful to check if mlx5_get_vq_irq() works
correctly. I believe it should return an error if the virtqueue
interrupt is not allocated. After a glance at the code, it seems not
straightforward to me.

Thanks

Thanks

quoted

Thanks,
Dragos