There are a number of factors which go into this consideration of Linux on the mainframe. I must admit it was really cool when I first learned of it, having had an MP3000 to myself at an IBM training facilty to learn how to bring up VM/ESA and Linux/390(2001). Then I realized a few things like:
1. Linux cannot take advantage of the advantages of channel-based disk i/o, because it uses Unix i/o approaches which can never be as efficient as the traditional mainframe-based approaches. No one has shown me any evidence that Linux does anything particularly intelligent in its channel program construction and management. Linux assures that IBM can happily sell lots of IFL or general purpose CPUs which are necessary to compensate for this inefficient use of
mainframe resources.
2. Managing workloads under zVM can be great and is extremely well refined, but this requires zVM-specific skills which supposedly no one wants to pay for.
3. For transaction-based work, it's hard to beat TPF/zTPF, but unfortunately that requires some real mainframe skill to implement. And regrettably, zTPF requires Linux and zOS because IBM refuses to convert the programs running on zOS to run on Linux instead. Since TPF/zTPF and zOS both involve onerous monthly licensing charges based on capacity, it's no wonder that TPF/zTPF languish in relative obscurity.