The performance problem has only gotten worse with time.

Modern CPUs have a rough time switching tasks. There are all those mitigations for meltdown and spectre. Even before that though, the situation with the TLB was ugly.

Limiting your data structures to ones that are friendly to message passing will really hurt you. Hacks to minimize the performance loss add complexity.

Fundamentally, glue isn't free. The parts may be simple, but the interactions are not.

Hardware IOMMU's have certainly made microkernels a lot more viable than they used to be in the past. You can actually place meaningful protection boundaries around core hardware drivers; they're no longer part of your inherent trust base, as they would be in a system where IO transfers can involve arbitrary memory.

Most embedded OSes for high integrity computing are microkernel based, then there is QNX, Minix running in legions of Intel CPUs, Android's workaround to have drivers as userspace processes with Android IPC as means to increase Android security and provide a proper driver ABI with Treble, and the ongoing push from Apple to move all kernel extensions into userspace, and then all those type 1 hypervisors.

Hardly forgotten.