Anon (anon.delete@this.anon.com) on January 10, 2020 5:51 pm wrote:
> Adrian (a.delete@this.acm.org) on January 9, 2020 11:49 pm wrote:
> > Nevertheless, I am actually working at that.
>
> Are you sure mmaped files are so seldom used that a Linux without virtual memory would be useful at all?
>


You also use the term "virtual memory" with another meaning than I do.


I use "virtual memory" only with its original meaning, i.e. creating the illusion of a virtual memory larger than the real memory, by using a swap partition or a swap file on an HDD or SSD.


The paging system can provide various other things besides virtual memory, e.g. memory protection, dynamic memory relocation, address space extension.


Mapping files in memory is a feature that can be implemented in various ways, it is not intrinsically tied to paging.

Of course, on modern CPUs the best way to implement mmap is using the paging system, because this is what the hardware supports.


For efficient implementation of mmap, you need exceptions to be generated, e.g. when accessing a page that has not been loaded from the HDD/SSD.

For this you may use the "present" bit in the page tables, whose original purpose was the implementation of virtual memory.

Nevertheless, the "present" bit is not really necessary. On a CPU that would not provide it you could use the memory protection bits and achieve the same.



Some of the features currently provided by the paging system are essential, e.g. memory protection, which is enough to be able to also implement mmap.


I have said that I believe that 2 of the others are not really useful, the virtual memory, which I have never used on hundreds of computers, from large servers to small laptops, during 20 years, and the dynamic memory relocation, i.e. the address translation.

For the second feature, I could be wrong, because while there is no doubt that one can implement a Linux compatible kernel without it, it is not clear how the performance of this approach would compare with the traditional memory management.


Because I am curious about this, I work at a kernel with various experimental features, one of which is this.