If performance is second, it seems to be a very, very distant second for most uses. So much software is just absurdly slow.
This patch shaved a minuscule number of cycles off a hot code path, and was accepted on the basis that microbenchmarks showed that it made extremely hot loops run a few times faster:
https://github.com/openzfs/zfs/commit/59493b63c18ea223857066...
This patch eliminated a redundant calculation that used 0.8% of CPU time:
https://github.com/openzfs/zfs/commit/3236c0b891d0a09475bef8...
The impact of the following patch was not measured, but it is believed to have made checksum calculations run several times faster by eliminating unnecessary overhead:
https://github.com/openzfs/zfs/commit/59493b63c18ea223857066...
The reason why the impact was not measured is that the code runs with interrupts disabled, which makes it invisible to kernel profilers. Another way of measuring was needed to quantify the improvement, but everyone agreed that it was a good improvement, so there was no need to evaluate the before/after improvement.
Few people are willing to accept the performance impact of moving their filesystem driver into userland and practically no one wants the performance impact of moving almost everything into userland.
https://github.com/openzfs/zfs/commit/677c6f8457943fe5b56d7a...
It seems to me that the microkernel idea came from observing that virtual memory protection made life easier and then saying “what would life be like if we applied virtual memory protection to as much of the kernel as we can”. Interestingly, the original email thread shows that they even tried doing microkernels without virtual memory protection in the name of portability, even though there was no real benefit to the idea without virtual memory protection, as you end up with everything being able to write to each other’s memory anyway such that there is no point,
Flash was a security nightmare. Not supporting it was a feature.