On Thu, Oct 23, 2025 at 10:59:01AM +0200, Eugenio Perez Martin wrote: > On Fri, Oct 10, 2025 at 2:34 PM Michael S. Tsirkin <[email protected]> wrote: > > > > ... > > > > > > > > @@ -2045,11 +2046,11 @@ static int vdpa_dev_add(struct vdpa_mgmt_dev > > > *mdev, const char *name, > > > if (ret) > > > return ret; > > > > > > - mutex_lock(&dev->domain_lock); > > > + write_lock(&dev->domain_lock); > > > if (!dev->domain) > > > dev->domain = vduse_domain_create(VDUSE_IOVA_SIZE - 1, > > > dev->bounce_size); > > > - mutex_unlock(&dev->domain_lock); > > > + write_unlock(&dev->domain_lock); > > > if (!dev->domain) { > > > put_device(&dev->vdev->vdpa.dev); > > > return -ENOMEM; > > > > > > Let's look at this example: > > > > So now you are invoking this under an rw lock: > > > > > > > > > > struct vduse_iova_domain * > > vduse_domain_create(unsigned long iova_limit, size_t bounce_size) > > { > > struct vduse_iova_domain *domain; > > struct file *file; > > struct vduse_bounce_map *map; > > unsigned long pfn, bounce_pfns; > > int ret; > > > > bounce_pfns = PAGE_ALIGN(bounce_size) >> BOUNCE_MAP_SHIFT; > > if (iova_limit <= bounce_size) > > return NULL; > > > > domain = kzalloc(sizeof(*domain), GFP_KERNEL); > > if (!domain) > > return NULL; > > > > > > ... > > > > > > > > Which unless I am mistaken will produce a lockdep splat and deadlock. > > > > Can you expand on this? All that code works with and without lockdep.
GFP_KERNEL can sleep and if that happens and another thread runs and will try to take the lock, it will spin forever. > > > > So it looks like the previous version did not compile > > and this one looks DOA. What's up? > > > > In case it is useful, I forgot to enable some vendor vdpa drivers of > HW I don't have, and that was the reason why it didn't compile. But I > didn't send it without testing it of course. > > > At this stage please include information about configs you > > tested, and how. > > > > And any locking changes should also be tested with lockdep enabled > > please. > > > > This is the text I'm planning to include in the next series, please > let me know if you want me to expand: > > Tested by creating a VDUSE device OVS with and without MQ, and live > migrating between two hosts back and forth while maintaining ping > alive in all the stages. All tested with and without lockdep.
