Dear seL4 community, I'm announcing the release of v0.2 of Neptune OS, an operating system that I have been developing that aims to create a Windows NT personality on top of the seL4 microkernel, by implementing the upper layer of the Windows kernel called the NT Executive, as well as Windows kernel drivers, as userspace processes under seL4. For those that remember the release of v0.1 two years ago and have been wondering, yes, the project is still alive. I have been a bit busy with postdoc work (academia is hard!!) last year but finally got a bit of free time lately to get enough work done that I think is interesting enough to warrant a v0.2 release. The project can be found on github (https://github.com/cl91/NeptuneOS). The biggest feature of the v0.2 release is a (I hope) reasonably complete file system stack supporting read-ahead and write-back caching, that includes a FAT12/16/32 file system driver and a floppy controller driver, both of which are ported over from ReactOS (but running in userspace instead of kernel space). I'll explain the technical points below but you can watch a short video (https://www.youtube.com/watch?v=o3FLRnkh0ic) that demonstrates the basic file system commands, including cd, dir, copy, move, del, md, and mount/umount. Arguably the biggest challenge for a userspace file system stack is minimizing the performance penalties of the additional context switches. There are radical solutions to this problem along the lines of Unikernels that require linking the FS code with the applications. These solutions are not very practical for everyday desktop or server OSes. On traditional OSes there is FUSE on Linux/Unix which is often regarded as having inferior performance compared to equivalent kernel-mode drivers. To some extent this problem is inevitable, as context switches are always more expensive than function calls, no matter how much you optimize the former. However, that doesn't mean we can't optimize our OS design so that the performance penalties of running file systems in userspace are small enough that the benefit of userspace file systems outweighs their performance cons. The key to this performance goal I believe is an efficient cache design. Indeed, for FUSE it is possible to tune the cache parameters to drastically increase its performance (see [1] and the citation therein). On Neptune OS we cache aggressively: not only do we cache the underlying block device pages, we also maintain the relationship between file regions and their underlying disk regions. In other words, the primary role of the file system driver is to translate an offset of a file object into an offset of the underlying storage device, and this relationship is cached since they are not expected to change, at least for currently open files. Furthermore, we cache file attributes as well as the file hierarchy itself, so that if a process opened a file and the file system driver returned a valid handle, we do not need to query the FS driver again (at least not before the first process closes the handle) when another process opens the same file with the same attributes (because we know it exists and know about its attributes). Of course, we need to be careful when things get closed/deleted/moved, etc, so we need a mechanism to synchronize information with the file system driver process. This is done by sending messages using the seL4 IPC. What is amazing is the fact that despite these changes to the inner workings of the NT cache manager, not to mention the fact that all drivers are now running in userspace, the Windows kernel driver API remains largely unchanged, so that it is indeed possible to port the relevant ReactOS drivers to our architecture without doing too much work (I believe I spent three to four man-weeks worth of work porting the FAT file system itself). The vast majority of work is in fact removing code rather than writing new code, because of the simplifications of locking and other synchronization issues that simply disappear when drivers run in userspace. For those who are interested in knowing more about our architecture and are perhaps interested in porting drivers from ReactOS, I am in the process of writing a book (found under the docs of the github repo) called "Neptune OS Developer Guide" that explains all these things. Note this is very much a work-in-progress. In addition to the cache manager work, in order for the file system stack to function, several other system components must be implemented, most notably DMA and Win32 SEH (structured exception handling). Our DMA architecture supports both PCI bus mastering and the weird ISA DMA that must go through the ISA DMA controller. The latter requires managing the so-called "map registers" that must be shared across different ISA devices. Again we find that the Windows kernel API can be adapted to a seL4-based userspace driver model almost straightforwardly. This is perhaps not so surprising, because NT was allegedly originally designed as a microkernel OS (this is inevitably an urban legend, but early NT design documents refer to the code beneath the NT Executive as the "microkernel", so there is reason to believe). To summarize, what I hope is that with the design outlined above, we can in fact have a practical, reasonably performant userspace file system stack without resorting to radical departures from traditional desktop/server computing paradigms. Of course, it's ridiculous to talk about performance for floppy disk controllers, so the primary goal of the next release is to port the ATA/AHCI hard disk drivers from ReactOS to Neptune OS, in order to produce a valid benchmark against equivalent kernel mode designs. Anyway, if you read so far, I hope you find the work interesting (I personally do!). If you have any comments, or even better if you ran the code and found bugs/issues, please do open an issue in the github repo. Thank you! [1] https://medium.com/@xiaolongjiang/linux-fuse-file-system-performance-learnin... --- Dr. Chang Liu, PhD. github.com/cl91/NeptuneOS