• Jonas Termansen's avatar
    Multithreaded kernel and improvement of signal handling. · 51e3de97
    Jonas Termansen authored
    Pardon the big ass-commit, this took months to develop and debug and the
    refactoring got so far that a clean merge became impossible. The good news
    is that this commit does quite a bit of cleaning up and generally improves
    the kernel quality.
    
    This makes the kernel fully pre-emptive and multithreaded. This was done
    by rewriting the interrupt code, the scheduler, introducing new threading
    primitives, and rewriting large parts of the kernel. During the past few
    commits the kernel has had its device drivers thread secured; this commit
    thread secures large parts of the core kernel. There still remains some
    parts of the kernel that is _not_ thread secured, but this is not a problem
    at this point. Each user-space thread has an associated kernel stack that
    it uses when it goes into kernel mode. This stack is by default 8 KiB since
    that value works for me and is also used by Linux. Strange things tends to
    happen on x86 in case of a stack overflow - there is no ideal way to catch
    such a situation right now.
    
    The system call conventions were changed, too. The %edx register is now
    used to provide the errno value of the call, instead of the kernel writing
    it into a registered global variable. The system call code has also been
    updated to better reflect the native calling conventions: not all registers
    have to be preserved. This makes system calls faster and simplifies the
    assembly. In the kernel, there is no longer the event.h header or the hacky
    method of 'resuming system calls' that closely resembles cooperative
    multitasking. If a system call wants to block, it should just block.
    
    The signal handling was also improved significantly. At this point, signals
    cannot interrupt kernel threads (but can always interrupt user-space threads
    if enabled), which introduces some problems with how a SIGINT could
    interrupt a blocking read, for instance. This commit introduces and uses a
    number of new primitives such as kthread_lock_mutex_signal() that attempts
    to get the lock but fails if a signal is pending. In this manner, the kernel
    is safer as kernel threads cannot be shut down inconveniently, but in return
    for complexity as blocking operations must check they if they should fail.
    
    Process exiting has also been refactored significantly. The _exit(2) system
    call sets the exit code and sends SIGKILL to all the threads in the process.
    Once all the threads have cleaned themselves up and exited, a worker thread
    calls the process's LastPrayer() method that unmaps memory, deletes the
    address space, notifies the parent, etc. This provides a very robust way to
    terminate processes as even half-constructed processes (during a failing fork
    for instance) can be gracefully terminated.
    
    I have introduced a number of kernel threads to help avoid threading problems
    and simplify kernel design. For instance, there is now a functional generic
    kernel worker thread that any kernel thread can schedule jobs for. Interrupt
    handlers run with interrupts off (hence they cannot call kthread_ functions
    as it may deadlock the system if another thread holds the lock) therefore
    they cannot use the standard kernel worker threads. Instead, they use a
    special purpose interrupt worker thread that works much like the generic one
    expect that interrupt handlers can safely queue work with interrupts off.
    Note that this also means that interrupt handlers cannot allocate memory or
    print to the kernel log/screen as such mechanisms uses locks. I'll introduce
    a lock free algorithm for such cases later on.
    
    The boot process has also changed. The original kernel init thread in
    kernel.cpp creates a new bootstrap thread and becomes the system idle thread.
    Note that pid=0 now means the kernel, as there is no longer a system idle
    process. The bootstrap thread launches all the kernel worker threads and then
    creates a new process and loads /bin/init into it and then creates a thread
    in pid=1, which starts the system. The bootstrap thread then quietly waits
    for pid=1 to exit after which it shuts down/reboots/panics the system.
    
    In general, the introduction of race conditions and dead locks have forced me
    to revise a lot of the design and make sure it was thread secure. Since early
    parts of the kernel was quite hacky, I had to refactor such code. So it seems
    that the risk of dead locks forces me to write better code.
    
    Note that a real preemptive multithreaded kernel simplifies the construction
    of blocking system calls. My hope is that this will trigger a clean up of
    the filesystem code that current is almost beyond repair.
    
    Almost all of the kernel was modified during this refactoring. To the extent
    possible, these changes have been backported to older non-multithreaded
    kernel, but many changes were tightly coupled and went into this commit.
    
    Of interest is the implementation of the kthread_ api based on the design
    of pthreads; this library allows easy synchronization mechanisms and
    includes C++-style scoped locks. This commit also introduces new worker
    threads and tested mechanisms for interrupt handlers to schedule work in a
    kernel worker thread.
    
    A lot of code have been rewritten from scratch and has become a lot more
    stable and correct.
    
    Share and enjoy!
    51e3de97
init.cpp 1.7 KB