This involves renaming the descriptor_from_fd function to the more appropriately named
open_descriptor_from_fd (since we check if the descriptor was opened and error out otherwise),
and creating a new function that does not verify that the file descriptor was opened.
That was a very old one from back in the old days. Now that the framebuffer is finally a device file,
and it can be memory-mapped by user programs for more performance,
this syscall is MORE than obsolete.
Some pages, such as framebuffer pages, are not physical memory frames reserved for the current process.
Some, such as the framebuffer, may be shared between all processes.
Yet, on exit() or on exec(), a process frees all frames mapped into its address spaces.
And on fork(), it copies all data between frames. So how could we map framebuffers.
Simple: we use one of the bits in page table entries which are available to the OS, and mark whether that page is owned by the current process.
If it is owned, it will be:
- Freed on address space destruction
- Its data will be copied to a new page owned by the child process on fork()
If it is not owned, it will be:
- Left alone on address space destruction
- On fork(), the child's virtual page will be mapped to the same physical frame as the parent
This still needs a bit more work, such as keeping a reference of how many processes use a page to free it when all processes using it exit/exec.
This should be done for MAP_SHARED mappings, for example, since they are not permanent forever,
unlike the framebuffer for example.
Not part of C or POSIX, but since there is no procfs right now, I thought it would be nice to have an interface to query process information.
It works like this: you pass the process ID and a pointer to a struct pstat (can be null).
If the process ID is -1, the kernel picks the process with the highest PID.
Then, if the pointer to a pstat struct is not null, the kernel fills it in with the process's information, and returns the process's PID.
This function is a Luna alternative to fork() and exec().
Why? Simply because I can't figure out for the life of me how to implement a working fork().
So meanwhile, we have spawn() as a replacement. exec() still exists, though.
That's why we now have a VFS and a /dev pseudo-filesystem. To provide that kind of things.
Remember, everything is a file!!
The new way to ask the kernel for random numbers is to read from /dev/random.
The only thing doing weird stuff is exec(), so that's commented out and throws ENOSYS right now.
But we have two user tasks running in parallel, isolated from each other!
That will probably happen a lot. We want userspace to tell us IF THE ERROR IS RELEVANT.
So, these unnecessary warnings are just noise.
Userspace may also use these functions to check for file descriptors.
For example, libc does this at program initialization, it checks whether fd 0 and 1 exist (by calling lseek() and seeing if it fails with errno=EBADF).
Very bare-bones for now. Doesn't support arguments or environment (we don't have that stuff right now), and the executable is not a valid ELF, it terminates the task.
But it's a start!
If a section needs to be mapped at 0x50f50 and its size is 0x200, then that address space exceeds one page.
But since 0x200 is less than one page, we only map one page.
If we count the offset, 0xf50 + 0x200 need two pages. So we can map the right amount of memory.
We should start to drop the old InitRD API, which only allows for files to be loaded from the initrd, and which forces pathnames to be relative (bin/init)
With VFS, we can load any kind of file from any kind of filesystem, and using paths that make sense (/bin/init)
Kernel: Implement a descriptor struct which stores the opened node and read offset, and give each task 8 of those.
Implement three syscalls: sys_read, sys_open and sys_close (sys_write still writes to the console instead of using a fd, for now)
Implement three new errors: ENOENT, EBADF and EMFILE.
libc: Implement the new errors, and the new syscalls in syscall().
Also fix _RETURN_WITH_ERRNO() to set errno correctly, which was making strerror() return null, thus crashing perror().
userspace: make init demonstrate the new file API.
The exit() libc function already accepted an integer, but didn't pass it on to the kernel since we had no mechanism for it to do that.
Now, the kernel stores a task's exit status to display it later (and in the future, return it to userspace via wait()/waitpid())
Kernel: Add an errno.h header with definitions for each header,
and return those, negated, from syscalls when there is an error.
mmap() returns an invalid address with errno encoded, instead of
returning a negated errno; this address is encoded as ffffffffffffffEE
where EE is errno in hex.
libc: make syscall() return -1 and set errno on error, instead of
returning the raw return value of the system call. Also, add mmap()
and munmap() wrappers in sys/mman.h :).
userspace: make the memeater program show the value of errno
when allocating memory fails.
Things to improve: add perror() and strerror() to make the errno
experience even better! >.<
This struct allows us to keep track of what memory is used by the loaded executable. For some reason, freeing this memory when the task exits triggers a kernel page fault, so I'm not doing that right now.
