iommu/amd: serialize sequence allocation under concurrent TLB invalidations

In the Linux kernel, the following vulnerability has been resolved: iommu/amd: serialize sequence allocation under concurrent TLB invalidations With concurrent TLB invalidations, completion wait randomly gets timed out because cmd_sem_val was incremented outside the IOMMU spinlock, allowing CMD_COMPL_WAIT commands to be queued out of sequence and breaking the ordering assumption in wait_on_sem(). Move the cmd_sem_val increment under iommu->lock so completion sequence allocation is serialized with command queuing. And remove the unnecessary return.