On 11.06.21 19:05, Jason Gunthorpe wrote:
On Wed, Jun 09, 2021 at 06:46:33PM -0400, Tony Krowiak wrote:
This patch introduces two new r/w locks to replace the wait_queue_head_t
that was introduced to fix a lockdep splat reported when testing
pass-through of AP queues to a Secure Execution guest. This was the
abbreviated dependency chain reported by lockdep that was fixed using
a wait queue:
kvm_arch_crypto_set_masks+0x4a/0x2b8 [kvm] kvm->lock
vfio_ap_mdev_group_notifier+0x154/0x170 [vfio_ap] matrix_dev->lock
handle_pqap+0x56/0x1d0 [vfio_ap] matrix_dev->lock
kvm_vcpu_ioctl+0x2cc/0x898 [kvm] vcpu->mutex
kvm_s390_cpus_to_pv+0x4e/0xf8 [kvm] vcpu->mutex
kvm_arch_vm_ioctl+0x3ec/0x550 [kvm] kvm->lock
Is the problem larger than kvm_arch_crypto_set_masks()? If not it
looks easy enough to fix, just pull the kvm->lock out of
kvm_arch_crypto_set_masks() and obtain it in vfio_ap_mdev_set_kvm()
before the rwsem. Now your locks are in the right order and all should
be well?
+static int vfio_ap_mdev_matrix_store_lock(struct ap_matrix_mdev *matrix_mdev)
+{
+ if (!down_write_trylock(&matrix_mdev->rwsem))
+ return -EBUSY;
+
+ if (matrix_mdev->kvm) {
+ up_write(&matrix_mdev->rwsem);
+ return -EBUSY;
+ }
+
+ if (!down_write_trylock(&matrix_mdev->matrix.rwsem)) {
+ up_write(&matrix_mdev->rwsem);
+ return -EBUSY;
+ }
+
+ return 0;
+}
This double locking is quite strange, at least it deserves a detailed
comment? The comments suggest these locks protect distinct data so..
+
+ ret = vfio_ap_mdev_matrix_store_lock(matrix_mdev);
+ if (ret)
+ return ret;
clear_bit_inv((unsigned long)apqi, matrix_mdev->matrix.aqm);
here it obtained both locks but only touched matrix.aqm which is only
protected by the inner lock - what was the point of obtaining the
outer lock?
Also, not convinced down_write_trylock() is appropriate from a sysfs
callback, it should block and wait, surely? Otherwise userspace gets
random racy failures depending on what the kernel is doing??
It might we worth exploring lock_device_hotplug_sysfs() which does a
"return restart_syscall()" with some delay.
--
Thanks,
David / dhildenb
