| Age | Commit message (Collapse) | Author |
|---|
|
interrupts respectively.
Added the scheduling primatives sleep_on() and interruptible_sleep_on()
to sched.c. sleep_on() adds the current task to the specified wait queue
then puts it into an uninterruptible sleep. Wait queues are formed by
having each waiting task holding a pointer to the waiting task that
preceeded it, then placing a pointer to it's own task struct into the
queue's head pointer. When the first task in the queue is awoken, it
wakes the preceeding task prior to returning from sleep_on(). The
function interruptible_sleep_on() works in a similar manner, however
when a task is awoken from interruptible_sleep_on(), it wakes up the
first task in it's wait queue, then puts itself back to sleep. The
purpose of this, is to automatically dissolve and reform the wait queue.
As each task in the queue is a awakened, tasks that need to handle
signals will do so whilst tasks that have no pending signals and are
still waiting will place themselves back into the wait queue by calling
interruptible_sleep_on() again. Whilst this method isn't the most
effiecient due to waking every task in a wait queue because one received
a signal, it does avoid the mess and complications involved in
maintaining the queue as an array or linked list.
Added prototypes for the scheduling primatives wake_up(), sleep_on() and
interruptible_sleep_on() to kernel/sched.h so that these functions may
be used throughout the kernel.
Added code to temporaily re-enable interrupts in reschedule() so that
the idle loop may continue to function even if interrupts were disabled
prior to calling the function.
|
|
printk() to notify the user of the kernel panic. This resulted in a
system call being made to the kernel itself and the machine not fully
halting.
Fixed an issue with the serial driver in which the functions rsputs()
and rsread() will still attempt a data transfer even if serial_init()
failed to detect and initialize a serial port.
Added the ability for tasks to be interrupted whilst reading from the
serial port. This was done by putting the task into TSTATE_INTERRUPTIBLE
instead of TSTATE_UNINTERRUPTIBLE when waiting for data in the serial
buffer. Furthermore, a check was introduced after the task wakes up to
see if any data was put in the buffer, or if the task was awoken by
another source.
Changed the type pid_t from an unsigned 16-bit integer to a signed
16-bit integer. This was done to make passing PID's to certain functions
easier.
Added the new system call sys_kill which will allow one process to send
a signal to another.
Added the kill_proc() function to sched.c to kill a process. Currently,
this works by nullifying the PID field in the process' task structure,
marking all the pages mapped to it's address space as free for use, then
calling the scheduler to switch to another runnable task (or to idle).
Modified the default signal handler within the kernel to now handle the
SIGKILL signal by calling kill_proc().
|
|
Added the _syscall2 macro to unistd.h to facilitate system calls that
require two arguments to be passed.
Modified the ATA driver to simply abort initialisation if a drive is not
found, or cannot be configured. This will allow the kernel to function
on a diskless system without invoking panic() unnecessarily.
Added the functions irq_enable() and irq_disable() to asm/interrupt.h to
make it easier for C code to mask and unmask IRQ's on each PIC.
Moved the declaration for rsputs() from kernel/con.h to the new
kernel/serial.h file since this is a function provided by the serial
driver.
Implemented a basic I/O input framework. This involves the new system
call sys_read, which takes an I/O read request and directs it to the
appropriate kernel handler function depending on the calling process'
ctty value. This mechanism is identical to the sys_puts system call.
Added the rsread() function to service sys_read calls from processes
whose ctty value is equal to 1. This function will continually copy data
from the serial buffer to the location specified. If there is not a
sufficient amount of new data in the buffer to satisfy the request, the
process is put into the TSTATE_UNINTERRUPTIBLE state and the scheduler
is called to switch tasks. Prior to calling the scheduler, the function
will set the waiting_task pointer to the calling process. This pointer
will later be used by the interrupt handler to wake the process when new
data arrives.
Added an interrupt handler to service the IRQ4 (UART) interrupt. This
subroutine is a stub which will save the machine's state then transfer
control to rs_handler() in serial.c which will read bytes from the
serial port and place them in a buffer. Before returning, rs_handler()
checks the waiting_task pointer to see if a task is waiting for the
newly received data and if so, it sets the task's state to
TSTATE_RUNNING before resetting the pointer to NULL and returning.
Ideally, the scheduler should be invoked at this point to select another
task but since our basic round-robin scheduler currently has no concept
of task priorities (and for the sake of simplicity), we will avoid
invoking the scheduler in response to interrupts for now.
|
|
Implemented the sprintf() library function in lib/stdio.c which uses the
vsprintf() function.
Implemented a very primative controlling TTY for each process. This is
achieved by a switch in the sys_puts system call which uses the 'ctty'
element of the process' task structure to determine an appropriate I/O
channel. A negative ctty value doesn't equate to any I/O channel
effectively disabling the process' output.
Added the sys_ctty system call which allows a process to set it's own
ctty value.
Removed the sys_rsputs system call. Output to serial is now performed by
the process first setting it's ctty value to 1, then invoking sys_puts.
|
|
runnable task is found. This is a very basic method of putting the CPU
into an idle state to reduce power consumption and heat production. This
method is far from perfect however since when the CPU is woken by a
timer interrupt, the scheduler runs through the entire process table
again regardless of whether any task has become runnable, before putting
the CPU back to sleep. In practice, this basic sleep mechanism reduced
idle CPU usage of the VM from 100% to ~6%, a very effective amount.
Updated the sleep() library function to use the new sys_alarm and
sys_pause system calls. This method works by first registering a dummy
signal handler, setting an alarm and finally calling pause() to put the
process to sleep. When the alarm expires, the dummy signal handler is
called (which returns immediately) and finally, the sleep() call
returns. Note however, that this is a temporary function implemented
poorly since it overwrites any pending alarms as well as the SIGALRM
handler.
|
|
CPU is not an i386 before executing the invlpg instruction with the
provided address. This will once again make the kernel compatible with
the i386 processor as executing the invlpg instruction without these
checks would have resulted in an invalid opcode exception.
|
|
a pointer to the saved state information to check_signals(). This bug
would have only manifested itself if multiple signals were to be
processed (sigret) or if a signal had been set during the handling of a
timer interrupt (tick_handler) and *ONLY* if switching to the user's own
handler since the state information is not needed to invoke the kernel's
default signal handler.
Implemented alarms for userspace processes. This required significant
modification of the scheduler algorithm. When idling waiting for a
process that can run, the scheduler now continually checks the alarms
and signals of each process and updates their state accordingly.
Implemented the sys_alarm system call to set the new 'alarm' field for
the calling process.
Created the sys_pause system call which changes the state of the calling
process to TSTATE_INTERRUPTIBLE, effectively removing it from the
scheduler's run queue, putting it to sleep until a signal arrives.
|
|
the IDT. This function takes the same parameters as register_isr() which
creates interrupt gate entries in the IDT.
The register_isr() function now sets the gate type to 0x0E regardless of
what was already in the descriptor. This is to break reliance on the IDT
already being initialized to a known state as well as avoiding conflicts
with the new register_trap() function.
Added declaration for the 'ticks' variable in kernel/sched.h so that
it's value may be used throughout the kernel.
Changed the system call gate to a trap gate. This means that interrupts
will not be disabled prior to entry into the system call handler. This
will allow hardware functions such as the timer to operate continuously
even if the user makes a system call.
Added checks to the timer interrupt handler. These checks prevent the
scheduler from being called if the interrupt occurred during kernel mode
execution. The idea here, is that the timer interrupt handler only
services the hardware (increments the tick count and sends an EOI to the
PIC's) if a system call was already running in the kernel. The system
call handler has also been expanded to check if the timer fired prior to
returning to userspace. If the timer did fire, the syscall handler will
invoke the scheduler (as the timer handler would have), so that it can
decide if it's time to switch tasks.
|
|
Implemented a basic serial interface using COM0 which can be accessed
with the system call sys_puts as well as the library functions rsputs()
and rsprintf().
Renamed puts() in con.c to con_puts() and made the function static to
avoid interference with the library function puts().
|
|
prior to return. This meant that switching to the same task did not
abort properly as the incorrect return address was popped off the stack.
Fixed a bug where the task register was not initialized before the
scheduler. This meant that on the first task switch, the CPU would dump
it's current state to a random address (0 most likely), potentially
corrupting important data. This has been corrected by introducing a
'garbage TSS' (and associated descriptor in the GDT) which is selected
before the scheduler is initialized as a safe place for the data to be
written.
Modified the scheduler so that it now waits indefinitely until a task
becomes ready to run. This fixes the possible bug where the scheduler
won't re-schedule the currently running task if it is the only task on
the system.
Add signal handling capabilities to the kernel. The bulk of this is
present in the subroutine check_signals() defined in traps.s. This
function is called on every timer tick and system call prior to
userspace return. The subroutine operates by pushing fake state
information onto the kernel's stack, then using it to return to
userspace. Prior to this, the subroutine pushes the return address
0xFFFFE000 onto the user's stack. This address corresponds to the
unmapped page located between the top of the user's stack (lower) and
the kernel's stack page (upper). When the user's signal handler tries to
return, it will cause a page fault that will be used as a notification
mechanism to inform the kernel that the signal handler is done. The
kernel will then switch to the originally pushed state information and
use it to return the task to the original execution state. Due to it's
nature, check_signals() must only be called prior to exiting the kernel
since it may not return.
Added the header file 'signal.h' which defines (most) of the POSIX
signals as well as the prototype for the signal() function.
Added the 'signal' element to the task structure. This field is a bitmap
of all currently pending signals.
Added the 'sig_handlers' element to the task structure. This is an array
of all user-defined signal handlers. Currently, a value of 0 indicates
the default handler should be used whilst any other value is considered
to be the address of a userspace signal handler. The ability to ignore a
signal is not yet present but will be added sometime soon.
Added the sys_signal system call to register a signal.
Added the stub function sighandler_default() to sched.c which handles
all signals not caught by the user.
|
|
be used freely throughout the kernel.
|
|
rather than in kmain() as some subsystems may now require early console
I/O.
Added 16-bit read/write I/O functions to asm/io.h. These functions are
inw() and outw() respectively.
Added the file kernel/fs.h which will contain definitions relating to
filesystem functions.
Defined the type off_t as a signed 32-bit value in sys/types.h. This
type will be required for filesystem functionality.
Added the directory 'kernel/fs' to the project's source tree. The
kernel's makefile has been updated accordingly. This directory will
contain any source files relating to filesystem functionality (both
assembly and C files).
Added the file 'fs/hd.c' to the kernel's source tree. This file
currently contains three main functions (which are defined in
kernel/hd.h). These functions are as follows; hd_init() to enumerate and
initialize the hard disks, hd_read() to read sectors from the disk and
hd_write() to write sectors to the disk. Currently, all transfers are
done in ATA PIO mode using polling, however this will change in future.
The function hd_init() is called during the kernel's boot sequence in
kboot().
Added the file hd.img to the project's root directory. This is a 20MB
raw image file that will be used by Qemu as a 20MB hard disk. The main
makefile has been updated to tell Qemu to use this file on launch.
|
|
Now, reschedule() requires a task to be in the TSTATE_RUNNING state for
it to run.
Renamed the TSS structure within the task structure from 'state' to
'tss' to avoid confusion with the task's run state.
Removed the task_state structure declaration from sched.c as it is no
longer needed due to context switches now being performed entirely in
hardware.
Removed the cstate pointer from sched.c. Interrupt handlers no longer
set a pointer to saved state information on the stack since this was
only needed for software task switching. NOTE: This may be re-introduced
should it become necessary to access state information on the stack.
Added a basic implementation of the wake_up() scheduling primative. This
function is currently not used and may be completely re-written in
future.
|
|
Since the kernel now has it's own stack unique to each address space, we
can now rely on hardware task switching to *also* save the task states.
To accomplish this, most of the code in switch_to() has been elimated.
This includes; the clearing of the busy flag in the old TSS on each
switch, setting the TR register to null prior to each switch and calling
save_state() (which has also been removed entirely), to copy the state
information.
Modified the for loop in reschedule() to account for the fact that the
switch_to() function may now return (which it *always* does when
returning to the task). For the same reason, switch_to() must also be
careful to preserve registers such as EBX and actually make a return
following the far jump.
Added basic definitions for task states in sched.h. These are; running,
interruptible, uninterruptible, zombie and stopped. These states will
(possibly) be used in the future to implement blocking system calls.
|
|
during process creation. This new kernel stack is placed at 0xFFFFFFFF,
as the last page in the 4GB address space. The base of the user's stack
has been moved down to 0xFFFFDFFF, exactly two pages below. The extra
page below the kernel's stack page will be left unmapped so as to
trigger a page fault if the kernel's stack overflows.
The initial values of the user's TSS have also been updated to reflect
the new bases of the stacks, and to ensure that the kernel's unique
stack is loaded when an interrupt occurs.
|
|
longer calls reschedule() directly. It now calls sched_tick() so that
the scheduler can make it's own decisions relating to handling the timer
interrupt.
The reschedule() function now *actually* implements a basic round-robin
scheduling algorithm that iterates through the list of tasks. This is
still a temporary algorithm but at least it now includes *all* runnable
processes instead of switching back and forth between process 1 and 2.
|
|
each structure has a TSS inside, as well as several descriptor entries
in the GDT pointing to each task's TSS. Task switching is performed in
hardware by means of far jumping to the new task's TSS descriptor in the
GDT. Task states are saved manually by copying the saved state from the
interrupt handler's stack, back into the task's TSS.
Added the function save_state() to sched.c which can be called from
assembly to copy the state information of the suspended task from the
current stack, to it's respective task structure. NOTE: this assumes
that the pointer cstate has been set ahead of time to point to the state
information on the stack, and that the pointer ctask has not yet been
modified.
Added the function reschedule() which is currently called from the
timer's tick handler. Currently, this function merely alternates between
the two tasks created in sched_init(), but in future, will be expanded
to decided independently which task will run next.
The function userspace_init() was renamed to the more appropriate
sched_init().
Moved most of the code from the original userspace_init() to the new
function create_proc() which sets up a virtual address space and state
information, copies the specified binary into the new address space,
then returns a pointer to the newly created tasks entry in the task
table.
Defined the type pid_t in unistd.h as a 16-bit unsigned integer.
Added two new debugging system calls for checking which task is
currently active. The getpid() system calls returns the caller's PID
whilst the getpdir() call returns the physical location of the caller's
page directory (the value of CR3). The getpdir() call can be used as a
fallback to determine which task is *actually* running if the
information in ctask or the getpid() call is faulty.
Added two new subroutines; set_tss() and clear_tss() (whose prototypes
are defined in asm/system.h) for managing the task-state descriptor
entries in the GDT. set_tss() initializes a TSS descriptor and sets it's
base pointer to the supplied TSS pointer whilst clear_tss() simply marks
a TSS descriptor as 'not present' without clearing it.
|
|
Added the prototype for userspace_init() to sched.h removing the need to
the extern declaration in kmain.c.
|
|
vsprintf() function to render formatted strings and then the puts system
call to output them.
Moved the vsprintf() function from the kernel to the library.
Furthermore, the prototype for the function has been moved from the
kernel's headers, to the new header file stdio.h.
Renamed the kernel's internal printf() function to printk() in order to
avoid confusion with the library provided function.
Renamed the sys_print system call to the more appropriate name,
sys_puts.
Added a new system call sys_time, which returns the system's uptime in
seconds. This is mainly for testing the userspace binary and will not be
permanent.
Added the file time.c to the library which contains the caller for
sys_time and a helper routine sleep() which delays execution for the
specified number of seconds. The new header file time.h contains
prototypes for both these functions as well as the definition for the
type time_t.
Fixed a bug in which the value of EAX was not properly passed to the
system call handler, resulting in the wrong system call being executed.
This was caused by the code in the SAVE macro not properly preserving
the value.
Fixed a bug in which the value of EAX was not preserved during a return
from system call, but rather restored with the original EAX value prior
to the call. As a result, system call return codes were not properly
passed. This has been corrected by introducing a new macro RESTORE_SYS
which carries out the same restore operations, but maintains EAX prior
to the return.
|
|
is produced. This may change later.
Added the new directory 'lib' to the source tree which build lib.a, an
archive containing common library routines for both the kernel and
userspace code to use.
Added the file string.c to the lib directory (as well as the appropriate
headers in /include) which provides some basic functions from the
standard C string library.
Added a physical memory manager which is now located in memory.c. This
memory manager tracks free pages from 1MB-8MB with a simple table and
allocates memory in blocks of 4KB pages. Multiple pages can be allocated
in which they are returned as a linked list.
Added a 'page window' in memory.c which allows the temporary mapping of
a single page at a time into the current address space.
Moved all paging routines that were previously located in page.s over to
memory.c where they have been re-implemented as a mixture of C and
inline assembly.
Moved the primative userspace routines from usrspace.s over to the new
sched.c. The only remaining routine, usrcall is now located in asm.s as
'switch_to' which takes two arguments, pointers to the task structure
and task state structure of the new task which is being switched to.
Pages for userspace are now allocated dynamically. The user binary is
loaded in at 1GB upwards. The user stack is located at the end of the
4GB address space with the lower 1GB being reserved for the kernel.
Updated the link.ld file for the userspace binary to include the new
starting address 0x40000000 (1GB).
Renamed the symbols for the user binary blob to make them shorter.
|
|
assembly function register_isr making it usable within the C portions of
the source.
Added a new file panic.s with the function panic that will print a panic
message, disable interrupts and halt the system.
Created the skeleton framework for paging in the new file page.s. The
new function paging_init (called in kboot) will setup a simple page
directory with two tables covering all addresses 0-8MB. It will also
mark pages from 0-1MB as 'supervisor-only' to protect the kernel. NOTE:
The function paging_init must be called before initialising the IDT as
it does not disable interrupts!
Modified the page fault handler to print the offending linear address
along with the supplied error code. Following that, the handler will
initiate a kernel panic. This function (along with panic) assumes
console I/O to be operational.
Modified the userspace test code to deliberately intiate a page fault by
accessing an unmapped page.
|
|
Hopefully further separation will help to keep the code readable and
understandable.
|
|
invoking the tool recursively.
Disabled GCC's position-independent-code generation in makefiles.
Modified makefile for kernel/usrbin so that it now compiles and links C
code into the userspace test.
Created errno.h and populated it with standard error definitions.
Replaced the va_list based system call handlers with a system call table
defined in the header kernel/sys.h. NOTE: This header is included in
kmain.c and should ONLY be included there! Do NOT include this header in
sys.c.
Rather than fetching the user's stack pointer and using it to initialize
a va_list, parameters are now passed to the call handlers via the
general purpose registers EAX, EBX, ECX and EDX where EAX contains the
requested call number and conveys the return value.
Setup macros in unistd.h to aid to making system calls from userspace.
Implemented two basic system calls; sys_print and sys_dummy. The former
takes a single char* argument and displays it on screen whilst the
latter is used to populate otherwise empty entries of the system call
table. sys_dummy returns the error ENOSYS whenever it is called.
|
|
32-255. A separate routine in the new file traps.s initializes the first
32 entries with addresses pointing to exception handlers within said
file.
Modified the register_isr function to now accept a descriptor privilege
level which it will assign to the modified IDT entry.
Added a task state segment and corresponding entry to the GDT. The TSS
will store the kernel's stack pointer and stack segment when switching
to userspace. NOTE: The stack pointer MUST be saved manually before
switching to userspace!
Added the framework for a system call interface at interrupt vector 0x80
(128).
|
|
a zero were not printed (affecting %x).
|
|
time_t.
Moved the kernel's loading point down to address 0 in
conventional memory and updated linker scripts accordingly.
Began to experiment with loading a binary blob into extended memory
(0x100000), switching to userspace mode, and executing it.
|
|
counter.
|
|
escape sequence printing a literal '%'.
Added the 'X' conversion specifier which differs from the 'x' specifier
in that resulting letters are converted to uppercase.
Added interpretation of the '#' alt flag which now causes the '0x'
prefix to be added to every hex conversion.
|
|
|
|
|
|
|
|
trailing 's' characters were appended to subsequent strings.
|
|
|
|
|
