Subversion Repositories eco32

Project History

---------------

14-Nov-2002

Project launched.

Main makefile written.

Directory structure created.

15-Nov-2002

Kernel of UNIX 7th Edition copied.

Download and installation of lcc.

This included the following steps:

1) In 'src', copy 'mips.md' to 'eco32.md' without changes,

   except that the interface record is named 'eco32IR'.

   The code output remains for now as it is in 'mips.md'.

2) Add instructions to makefile for compiling 'eco32.md'.

3) In 'src', add the interface record definition 'eco32IR'

   to the file 'bind.c'.

4) Add a directory 'include/eco32/linux' and copy all header

   files from the corresponding directory for the mips processor.

5) In 'etc', copy 'irix.c' to 'eco32-linux.c' and change the

   contents of this file to properly call the preprocessor,

   the compiler, etc.

6) Add the option -DLCCDIR=\"$(BUILDDIR)/\" to the instruction

   in the makefile which compiles $(HOSTFILE).

7) In 'src', substitute 'lex.c' with a newer version, which

   computes the types of hexadecimal constants correctly.

Disk creator done.

Program to write the bootstrap sector done.

16-Nov-2002

Preliminary instruction formats defined.

Preliminary instruction set constructed.

Some ideas for an address translation unit noted.

17-Nov-2002

Final instruction formats defined.

Preliminary version of simulator finished.

19-Nov-2002

Instruction set cleanup done.

Work on assembler started.

23-Nov-2002

Four of the five ROMs translated to ECO32.

24-Nov-2002

Preliminary version of assembler/linker/loader finished.

25-Nov-2002

Bootstrap monitor ROM translated to ECO32.

Simulator is running.

Bootstrap sector on disk created.

Now it is possible to boot from the disk.

26-Nov-2002

Work on an ECO32 back-end for the C compiler started.

27-Nov-2002

Preliminary version of back-end done.

28-Nov-2002

'Load address' instruction invented as macro for a set

of short sequences of instructions which are difficult

to generate by the back-end.

29-Nov-2002

Startup code (c0.s, c1.s) revised.

Writing the bootstrap sector revised.

System ROMs revised.

Tests revised.

30-Nov-2002

Started to program the MMU.

Surprise: we don't need an MMU enable bit! The unit can be

switched on all the time, but we have to be careful not to

use the TLB before it is properly initialized.

01-Dec-2002

Interrupt system re-designed.

Added the last missing case in the 'lda' implementation.

Timer interrupt is working properly.

We now have a new instruction: 'nor'. With this it is easy

to implement the bitwise complement of an integer. Done.

The compiler option -Wo-kernel compiles a program so that

it can be run in kernel mode.

This is a big step forward: hello.c is running (in kernel

mode of course).

02-Dec-2002

The characteristic feature of kernel mode programs as well as

of the kernel itself is that they want to handle all sorts of

special conditions for themselves (e.g., interrupts, exceptions,

startup code). So if the flag -Wo-kernel is given, the startup

library code (c0.s and c1.s) is no longer linked in automatically.

03-Dec-2002

The assembler now contains a complete expression parser which can

handle additive, multiplicative and shift expressions as well as

unary and parenthesized expressions.

Error in back-end corrected: mul & div instructions did not

correctly specify the second source register.

04-Dec-2002

The low-level code generators in the assembler had completely to

be redone: they could not handle big constants (immediate values

and addresses) in instructions that allow only 16 bit constants.

05-Dec-2002

I now have a pair of functions which allow to get/set an

interrupt service routine in the interrupt service routine

table. The low-level register save/restore is handled in

assembler, but the interrupt service routine itself can be

written in C. An application of this principle shows how to

determine the total amount of physical memory by probing

the memory until a bus timeout exception occurs. This can

fully be formulated in C.

The back-end generates a 'jal' instruction even when the

argument is a register. Although it would perhaps be more

appropriate to generate a 'jalr' instruction, I changed the

assembler to accept an address as well as a register with

this instruction.

06-Dec-2002

TLB flush is the first instruction dealing with the TLB.

Implemented.

07-Dec-2002

Until now we only had one special processor register, the

processor status word. With the MMU present, this will change.

Renamed 'gpsw' (get processor status word) to 'mvfs' (move

from special register). Renamed 'ppsw' (put processor status

word) to 'mvts' (move to special register). These instructions

have two arguments: a single register and a number specifying

the special register.

It is perhaps better to use an established set of instructions

dealing with the TLB (i.e., the way MIPS is doing this) instead

of an even more minimalistic one, which is not guaranteed to

cover all cases. So I will abandon 'TLB flush' in favor of the

four instructions found in the MIPS instruction set ('tbs',

'tbwr', 'tbri', and 'tbwi').

As is documented in the MIPS reference book, the valid bit V is

not involved in the TLB matching process. Implementation changed.

Perhaps we don't need the V bit at all? I'll leave it out for now.

08-Dec-2002

Today I came up with the idea to test the MMU implementation by

running a somewhat more realistic test: a task which is compiled

for virtual address 0 and has a minimal I/O library which traps

to the kernel to output a character.

09-Dec-2002

It is not necessary to prepare real page tables for the task, as I

thought initially. Since we have only one task which uses exactly

two pages (code and stack), it is possible to preset the TLB with

two entries for the pages and then leave it alone.

The test is running.

10-Dec-2002

Now for the next step: two tasks running concurrently, preempted

by a timer interrupt. This is really a lot more complicated than

a single task because now the entries in the TLB must be flushed

and re-programmed when a task switch occurs.

Another complication is the need for two kernel-mode stacks, one

for each task, in addition to the task-specific user-mode stacks.

This is necessary because of two facts:

  - the kernel cannot rely on the integrity of the user-mode stacks

  - the state of the suspended task must be preserved somewhere,

    even if the other task is running and trapping to the kernel

    for a system call.

11-Dec-2002

This was indeed much harder than expected, but now it works.

12-Dec-2002

It is evident that only a very short sequence of instructions is

able to handle a TLB miss exception efficiently. MIPS is doing

this by giving the TLB miss exception a separate interrupt vector.

We don't want to go so far, but of course we cannot tolerate to

execute all the lengthy actions at the beginning of the standard

trap handler. So we will sort out the TLB miss exception at the

very beginning of the standard exception handling.

13-Dec-2002

It is easy to catch the TLB miss - but how should it be handled?

Some sort of page table has to supply the necessary information

so that a new entry can be filled into the TLB. How should the

page tables be organized? In the past I saw cascaded page tables

as the only solution. But it might be better to have a single

page table which itself is paged and lies in virtual memory.

The big advantages of this scheme are fast access (no pointers

to chase) and (almost) no waste of memory because gaps in the

table do not need any real memory. But then of course we have

to handle TLB misses while accessing the page table...

Very confusing!

14-Dec-2002

What information is needed in which format if a TLB miss (or any

other TLB exception) occurs? I really have no stringent answer to

this question, so I will follow again a minimalistic design path:

Implement a register in the MMU which is set to the offending

virtual address in case the translation didn't work. Done.

As I realize now, this hardware equipment can support a variety of

different organizations of the page tables handled by software.

Even inverted page tables seem feasible!

It would be very convenient if the kernel could use another

register without having to save its contents first. I think

we could let the kernel have also register 28 at its disposal.

17-Dec-2002

Yesterday I realized that the 'lda' macro instruction is no longer

needed because 'add' does the very same job. Deleted.

18-Dec-2002

It is very desirable to have an interrupt mask in the PSW.

If, e.g., the timer interrupts and the timer ISR decides to

disable interrupts from all devices but the timer itself,

then this is not easily achievable without a central mask.

19-Dec-2002

One's complement, represented by '~', is another operator

implemented in the assembler.

20-Dec-2002

As well as bitwise and, bitwise xor, and bitwise or,

represented by '&', '^', and '|', respectively. These

operators are left associative and have the same priority

as they have in C, i.e., & > ^ > |, in order of decreasing

priority.

21-Dec-2002

Because of the new structure of the PSW all tests must be

updated and carried out once more. A lot of work...

22-Dec-2002

I spotted a minor bug in the simulator: a breakpoint which

was set to address C0000004 in order to catch any interrupt

or exception will not work for exceptions. This is due to

the handling of exceptions by setjmp/longjmp. Corrected.

23-Dec-2002

ECO32 Version 0.0 released.

26-Dec-2002

Today I started to convert the original sources of

UNIX 7th Edition to ANSI C.

29-Dec-2002

I found a real programming error in the text.c module:

  if (ip->i_flag&ITEXT==0) return;

My compiler warned: 'expression with no effect elided'.

'==' has higher precedence than '&', but ITEXT does not

equal zero, so the whole test expression becomes zero

and the return is never executed. The line should read:

  if ((ip->i_flag&ITEXT)==0) return;

31-Dec-2002

The masters made a typoe: in the file kl.c, the function

named 'ttioccomm' is spelled 'ttioccom'. I think this

did go undetected because of the very restricted length

of symbols with external linkage.

01-Jan-2003

Today I began investigating the startup procedure. I will

place the per-process kernel stack in the top half of the

u area, which itself is one page in size and is located

directly underneath the kernel at 0xC0000000 - 0x1000.

This part of the virtual address space is mapped and so

the u area of the running process is always accessible.

02-Jan-2003

How do we do the memory management? This is THE central

question of our porting effort. One possibility which

offers the chance of only minor changes to the existing

sources would be to keep the management by the coremap

array, but with the granularity changed to 4 k, the page

size. This route of course would not barely touch the

power of demand paging, which is possible by our hardware.

Nevertheless, I will follow this path for the time being.

I changed the coremap/swapmap entries into structs of two

integers instead of two shorts. Otherwise the range of

addresses might get too small (although this is unlikely:

64 k pages of 4 k bytes each equals 256 M bytes - quite a

lot of memory).

04-Jan-2003

This last change is nonsense - much of the other stored

information about processes would have to be changed also.

Reverted.

Because ECO32 does not have two different stack pointers

(as the PDP-11 has) we must switch to the kernel stack

explicitly. But we must not do the switch if the interrupt

hit while the machine was running in kernel mode already -

then the sp points deep into a properly setup kernel stack.

Switching stacks would then destroy the stack completely.

11-Jan-2003

I discovered a mysterious overwriting of some bytes of

the loaded UNIX kernel. This took place even before any

of its code was executed! A little bit of searching, and

the guilty one was found: the bootmon ROM code and its

stack. Now there is a problem: where should we put the

stack of the bootmon ROM?

We have two ways to get some program into the bare machine:

a) The 'realistic hardware' scenario: a ROM holds an initial

program which e.g. may boot the operating system from disk.

b) The 'convenient debugging scenario': the program, e.g. the

operating system, is somewhat magically loaded into RAM.

The error described above is caused by the unwanted interaction

of a) and b) when used together. So I think that the correct

solution of the problem is to forbid the specification of both

simulator options -r and -l at the same time.

13-Jan-2003

The simulator uses all CPU power which is available, even

when only waiting for a command line. This can easily be

changed by inserting a usleep(100) call into the input

waiting loop of the curses interface. Done.

Two new pseudo-ops (.syn and .nosyn) added to the assembler.

They allow or disallow synthesizing instructions with 'big'

constants, respectively. This can be used e.g. in the TLB miss

handler, where much finer control over the usage of register

$1 is needed.

15-Jan-2003

The constant BSLOP in param.h must be set to 0. Otherwise there

would exist blocks which don't start at an address which is a

multiple of 4. This would result in a Bus Address Error if e.g.

an integer is fetched from the block.

16-Jan-2003

The type mapping from C to machine types is as follows:

         PDP-11                   ECO32

char     byte (1 byte)            byte (1 byte)

short    word (2 bytes)           half (2 bytes)

int      word (2 bytes)           word (4 bytes)

long     double word (4 bytes)    word (4 bytes)

T *      word (2 bytes)           word (4 bytes)

As can be seen, only ints and pointers are differently mapped.

This is of not much concern with in-memory data structures but

can be desastrous with on-disk data structures (super block,

inodes). Therefore we have to do the following:

a) define ino_t as unsigned short

b) rearrange the components in struct filsys (and in principle in

struct dinode and struct direct too) so that no padding occurs.

I thought that I could simply write a disk with the PDP-11 simulator

and then use it with the ECO32 simulator; this is not possible

due to different endianness.

17-Jan-2003

Here are the changes I actually made:

1) in param.h:

   a) type 'ino_t' is 'unsigned short' instead of 'unsigned int'

   b) type 'dev_t' is 'short' instead of 'int'

2) in filsys.h:

   a) delete all components of struct filsys below the comment

      'remainder not maintained by this version of the system'

   b) 's_isize' is of type 'daddr_t', not of type 'unsigned short'

   c) 's_nfree' is of type 'int', not of type 'short'

   d) 's_ninode' is of type 'int', not of type 'short'

   e) exchange components 's_isize' and 's_fsize'

3) in inode.h:

   a) 'i_flag' is of type 'short', not of type 'char'

   b) 'i_count' is of type 'short', not of type 'char'

The size of the superblock structure is thus increased by 6 bytes

to 424 bytes (well below the 512 bytes block boundary), and all

padding or misalignment is avoided.

18-Jan-2003

I replaced the "Hello, world!" message in the default master boot

record by a somewhat more appropriate message.

19-Jan-2003

In order to simplify things a bit, I removed the '-n' option from

the assembler. Now assembler output files always do have headers.

There are of course situations in which the header must be stripped

off, e.g. with the boot record, or with the ROM in the simulator.

We have therefore now a 'loader', a little program which converts

an assembler output file into a memory image without any header.

This has the additional benefit that the simulator does no longer

need to know the executable format, which is especially good if we

change the format or add new formats.

20-Jan-2003

Started to port the 'mkfs' utility.

21-Jan-2003

Started to program a 'show file system' utility.

22-Jan-2003

Today I decided to spare a small number of TLB entries (4) from

random replacement. This helps to keep such vital information as

the current kernel stack permanently mapped.

25-Jan-2003

I changed the disk so that it is now similar to the 'rp' device

in the PDP-11. This has the advantage that the usr file system

is located on a different device (specifically: a different minor

device, i.e., a different partition) so that mounting can be tested.

The partitions and their intended use is as follows:

   device      size in blocks     use as

  -------------------------------------------

    rp0           14000           whole disk

    rp1            2000           root

    rp2            4000           swap

    rp3            8000           usr

28-Jan-2003

This was quite a lot of work, but finally 'mkfs' is running.

The crucial tool to achieve this was 'shfs' - show file system,

which allows displaying arbitrary blocks of a disk in several

formats (raw data, super block, inode block, directory block,

free list block, or indirect block). So everything on the disk

can be examined in detail.

By the way, the main hurdles to overcome had been different

endianness of ECO32 and x86, different padding of structures,

and alignment restrictions with ECO32, which are not present

on the x86 architecture.

29-Jan-2003

Here is a short report on byte-ordering on x86, ECO32, and PDP-11.

The C data type 'long' is stored in 4 bytes on all thre machines,

as already stated above (16-Jan-2003). The order of the bytes is

different on all three machines. We number the bytes from MSB

(most signigicant byte) to LSB (least significant byte): then

the number (b1 b2 b3 b4) has the value b1*2^24+b2*2^16+b3*2^8+b4.

If 'A' is the lowest (byte) address where the number is to be stored,

then the memory layout of the three architectures look like this:

x86       A+0:b4  A+1:b3  A+2:b2  A+3:b1    (i.e., little endian)

ECO32     A+0:b1  A+1:b2  A+2:b3  A+3:b4    (i.e., big endian)

PDP-11    A+0:b2  A+0:b1  A+0:b4  A+0:b3    (i.e., mixed)

An additional complication results from the fact that disk addresses

in inodes on a disk are stored in a three byte format, which is

different on ECO32 and PDP-11 (there is no such format on the x86,

because we don't want to run a file system on a bare PC, only within

the ECO32 simulator). In this format only the three least significant

bytes (b2 b3 b4) are stored; if b1 were non-zero an error would be

triggered. So then, here are the layouts of this special format:

ECO32     A+0:b2  A+1:b3  A+2:b4

PDP-11    A+0:b2  A+1:b4  A+2:b3

As can be seen, this format is constructed on both architectures by

omitting b1 (the MSB) from the four byte format.

I changed the routines that convert the three byte format from and to

the four byte format (iexpand() and iupdat(), both in module iget.c)

in order to cope with the changed byte-ordering.

Reduced the RAM size to 1M; this fits better to the disk swap size.

Added a counter to the simulator which counts the total number of

instructions executed since the last reset. This helps a lot if one

wants to re-visit a specific location in the program, if a lot of

single-stepping has been done to find the location for the first time.

31-Jan-2003

There has been an inconsistency with the object file format since

its invention: the header data is stored in little endian format,

which is convenient for the tools working in the cross development

setup. The tools running on the target architecture require the

data to be in ECO32 format of course. This is especially true for

the kernel exec file loader. Changed.

03-Feb-2003

ECO32 Version 0.1 released.

04-Feb-2003

Added to 'shfs' a command which translates an inode number into the

block number in which this inode is located, together with an inode

number relative to that block.

The simulator's 't' command now displays also the special registers

of the MMU (Index, EntryHi, EntryLo, BadAddress).

Memory commands of the simulator completed.

05-Feb-2003

ATTENTION: I changed the output of the disassembler to get it more

consistent with the rest of the simulator, and especially with its

inline assembler. There are no more '0x' prefixes in front of numbers,

and we don't have any decimal output. All numbers are given as hex

numbers without a prefix (the only exception to this rule are the

numbers of registers proper, of course, which are given in decimal).

The instructions which do sign extension show their arguments as

signed hex numbers. Labels are also no longer prefixed.

Inline assembler completed.

17-Feb-2003

ECO32 Version 0.2 released.

20-Aug-2003

ECO32 Version 0.3 released.

21-Aug-2003

1) I changed the 'ldhi' opcode from 0x1E to 0x1F.

   Reason: The instruction decoding gets more regular.

2) The ALU has now an 'xnor' instruction instead of a 'nor':

   (a xnor b) = ~(a ^ b) = (~a ^ b) = (a ^ ~b) = (a == b),

   bitwise parallel for all 32 bits.

   Reason: The ALU is easier to implement, and the instruction

   set becomes more regular.

3) The PC relative jumps (j, jal, all conditional branches) now

   have PC+4 as their reference location (and no longer PC).

   Reason: PC+4 is computed early in the instruction cycle; all

   other RISC machines do this alike.

4) The PC relative jumps (j, jal, all conditional branches) now

   encode a word offset, and no longer a byte offset.

   Reason: The target ranges for the jumps are quadrupled without

   any additional effort.

07-Oct-2003

Explored the possibilities to implement a graphics controller for

ECO32. Easy from the simulator's point of view, not quite so easy

from the implementer's point of view (due to X complications,

threads needed) but certainly feasible.

13-Oct-2003

Graphics controller implemented.

20-Oct-2003

Curses feeds certain keypad codes directly into the simulator's

terminal, e.g. KEY_BACKSPACE. There the MSBs get stripped off and

a wrong keycode is supplied to the simulated program. Now these

special codes are mapped to regular control characters, e.g. BS.

25-Oct-2003

The assembler now understands the option '-h' which suppresses the

output of the executable's header and appends the BSS's size bytes

of zeroes at the end of the executable. This in turn renders the

'load' tool redundant, which can be deleted. Done.

26-Oct-2003

A group of students is working on a ROM-based monitor program. They

need an instruction to implement breakpoints. Because the instruction

map is already crowded they have to use 'trap'. But they want of

course a breakpoint trap be easily distinguishable from any other

kind of trap. So the assembler now allows a constant operand for

'trap'. This operand is assembled into the otherwise unused bits

of the instruction. These bits are in turn ignored by the CPU (as

was the case all the time) but can be inspected by software which

must read the instruction from memory in order to do this. Now a

breakpoint trap can get a unique bit pattern distinct from the bit

patterns of all other traps. Remark: not only the regular assembler

must understand this but also the simulator's inline assembler as

well as its disassembler. Done.

09-Nov-2003

The curses library correctly sets the terminal mode to "raw" so that

neither SIGINT nor SIGQUIT signals can be generated by the keyboard.

With a graphics controller attached it is however possible that the

user selects "kill" from the window menu of the monitor window and

thus shuts down the connection to the X server. This in turn leads

to an uncontrolled instant termination of the whole simulator which

cannot be tolerated (e.g. the terminal would stay in raw mode: very

annoying for an unsuspecting user). Solution: install an I/O error

handler which gets called from the X system in case the connection

to the server is lost. Done.

16-Nov-2003

I added a check to the initialization of the memory simulation ensuring

that neither ROM image files nor program files loaded into main memory

can overflow the available space (and thus crash the simulator).

18-Nov-2003

The terminal simulation now handles the following control characters:

(input)             0x0D

                 0x08

                0x0E

                  0x10

                0x02

               0x06

(output)   0x07             beep

           0x0D             return cursor to begin of current line

           0x0A             scroll if cursor is in last line,

                            set cursor to next line in same column

           0x02             move cursor left (stop at leftmost column)

           0x06             move cursor right (stop at rightmost column)

           0x10             move cursor up (stop at topmost line)

           0x0E             move cursor down (stop at bottom line)

           0x08             delete character to the left of cursor,

                            move cursor left

           0x09             output spaces (minimum is one space),

                            until column mod 8 = 0

19-Nov-2003

Today I took the first few steps towards a bootable system disk.

Changes in bootstrap monitor:

  - 'g' command deleted

  - monitor stack sits just below 1M

  - master boot record loaded at physical address 0x00000000

  - master boot record started at virtual address 0xC0000000

Changes in constructing the default master boot record:

  - relocate code to 0xC0000000

20-Nov-2003

Now here is the big picture of the boot process:

a) ROM bootstrap

   The ROM loads the first sector of the disk and starts whatever it

   has loaded (provided that the signature is present). Until now this

   was always the default master boot record. In the future this will

   be the first stage bootstrap. This part is already finished.

b) First stage bootstrap

   The first stage looks for a file named "boot" in the root directory

   of the file system on the boot disk. This part must fit into a single

   sector and thus is restricted in its capabilities:

     - boot must lie in the root directoy

     - it must be one of the first 320 directory entries in the root

       directory (only the 10 direct blocks of the directory file are

       searched for it)

     - it must not be bigger than 5120 Bytes (only the 10 direct blocks

       of the executable file are loaded)

c) Second stage bootstrap

   This stage loads the kernel, perhaps asking the user for its name.

22-Nov-2003

First stage bootstrap completed.

27-Nov-2003

Second stage bootstrap completed.

It was indeed possible to ask the user for the path to the kernel he

wants to get started. The only kernel available for now is one of the

tests from tst/os.

29-Nov-2003

Makefiles tweaked.

06-Dec-2003

Today I started to develop the tools shfs and mkfs for file systems

with 4K blocksize. With this blocksize it should be possible to do

the bootstrap without a named second stage. Conceptually there are

again two stages (first load a single sector, then the rest), but

both stages should fit within the boot block.

20-Dec-2003

I discovered a small bug (?) in the simulator: if a write to $0 is

requested by an instruction, not only the write is suppressed (this

is expected) but also the source operand does not get computed.

Thus the instruction ldw $0,$0,1 which is incorrect if no TLB entry

for page 0 exists (and is so even if it exists because of alignment

violation) does not lead to an exception (which is quite unexpected).

Solution: Always compute the source operands, even if $0 is the target.

30-Dec-2003

In order to use swap space on the same disk where the root file system

is located we need some form of disk partitioning, preferrably NOT

deeply coded into the disk driver (as the original UNIX has done it).

In other words, we need a real master boot record, a partition table

and a tool ("fdisk" or "mkpart") to set up such a table.

01-Jan-2004

"mkpart" finished. It constructs a partition table from a textual

configuration file and writes it to sector 1. It also copies an MBR

(which has not been done yet) to sector 0, and a bootstrap manager

(which also has yet to be done) to sectors 2-7.

MBR coded.

02-Jan-2004

Bootstrap manager done.

03-Jan-2004

Tool "mkpart" is working.

New tool "shpart" displays a partition table, read from a disk image.

04-Jan-2004

I made the file systen viewer aware of partitions.

05-Jan-2004

Now "mkfs" also knows something about partitions.

08-Jan-2004

If the operating system is booted from a partition of a disk (in

contrast to being booted from the start of the whole disk) its disk

driver must know where logical sector 0 of the partition is located

on the disk. In other words, the master bootstrap must transfer a

little bit of information to the boot loader of the OS. This could

be done as follows: $26 holds the start sector of the partition from

where the OS is booted and $27 holds its size. In order to use the

very same bootstrap also in the case that the disk has no partitions

(and thus the file system occupies the whole disk), this information

must also be supplied by the ROM bootstrap sequence. Done.

Surprise: this scheme would allow the sub-partitioning of a partition

without any modification!

10-Jan-2004

Now the whole bootstrap chain is working. It goes like this:

a) The ROM loads the first sector of the disk and starts whatever it

   has loaded (provided that the 0x55AA signature is present). This

   may be either the master boot record if the disk is partitioned,

   or an ordinary boot record if it is not (see remark below).

b) In case the master boot record is executing, it loads sectors 2..7

   into memory (sector 1, the second sector of the disk, is skipped

   because it holds the partition table). These sectors contain a small

   program, the boot manager, in executable format. The boot manager is

   started by the master boot record.

c) The boot manager loads the partition table from sector 1 of the disk,

   displays the partitions of the disk, and asks the user to enter a

   partition number to boot. An attempt to boot a partition outside the

   range 0..15 is rejected, as well as requesting to boot a partition

   which does not contain a file system of any kind, or does not have

   the "bootable" flag set. If all is well, the boot record of the

   selected partition is loaded. If it has a valid 0x55AA signature,

   it is started.

   Remark: This is the point in the bootstrap chain which is reached

   immediately from the ROM code if the disk is not partitioned.

d) The boot record code loads partition relative sectors 1..7, i.e.

   the rest of the boot block. These sectors contain a small program,

   the bootstrap loader, in executable format. The bootstrap loader is

   started by the boot record code.

e) The bootstrap loader asks the user for a path to a file which must

   exist, must be a regular file, and must be executable by its owner.

   This will of course be normally an operating system kernel. The

   default value is /boot/unix. The file is loaded and started.

Throughout the whole process registers $26 and $27 hold the start sector

and size of the currently selected range of the disk, respectively.

12-Jan-2004

The register conventions shown above must be changed somewhat because

with more than one disk present it is also necessary to tell the disk

driver which disk to use:

$26 base address of disk controller

$27 start sector of partition (or disk)

$28 size of partition (or disk) in sectors

28-Jan-2004

After rewriting parts of the make file system utility (so that double

indirect blocks can be used, which is necessary for a file system with

512 byte blocks to hold files of more than 69K) it is now possible to

boot our port of UNIX 7th Edition from the disk. Alas, it is not yet

running very stable.

10-Mar-2004

Finally I added the generation of copying loops to the compiler's

back-end. This has long been missing.

17-Apr-2004

Translating virtual addresses using the TLB costs time because the

parallel associative hardware must be simulated by serial software.

This is very noticeable with programs running in user mode; they run

at roughly one third of the speed kernel mode programs do. In order

to alleviate this I added a function which acts almost the same as

the TLB associative lookup does, but without any error checks. Its

sole pupose is to use up time and it is called whenever an unmapped

address is detected by the virtual-to-physical address translation.

21-Apr-2004

Up till now the assembler located the data segment directly above

the code segment. This is not a viable strategy once paging is in

effect. Due to code sharing and protection reasons, the data segment

should start on the next page boundary above the code. The resulting

gap may also be present in the executable file: there is no need to

store any bytes for it in the executable (at least as long as it is

an executable with a header - for headerless executables the gap

bytes must be stored within the file).

It would have been possible to define two different executable formats,

one with the data segment directly following the code segment, and

another one with the data segment aligned on a page boundary. The

two formats then would have to be distinguished by different magic

numbers in the header. I did not do this, however, because I saw

no reason to keep the "directly following" format. If one wants to

compress code and data as much as possible, one can in any case put

both into the code segment.

24-Apr-2004

The simulated graphics device did not detect memory accesses which

lie within the device address bounds but outside the screen memory.

These accesses were propagated to the X Window System and in turn an

X Window data structure was accessed outside of its allocated memory.

At least on our Sun computer this translated to an immediate death

of the simulator process. Now the graphics device detects such an

attempt to access simulated memory outside the screen, and produces

a 'bus timeout' exception.

06-May-2004

ECO32 Version 0.4 released.

08-May-2004

I intend to give the ECO32 simulator the ability to run with more than

one simulated terminal. This will work best if there is no difference

between the first terminal and the other ones, so I will open a separate

window for each terminal (and use the first window in which the simulator

was started only for controlling the simulator, and NOT for a simulated

terminal). Then using curses is no longer necessary and can be removed.

But I want to retain command line editing, so I will use the 'getline'

library.

Curses interface removed.

Getline library installed.

Interrupt priorities changed:

  15    not used

  14    timer

  13    not used

  12    not used

  11    not used

  10    not used

  09    not used

  08    disk

  07    terminal 3 receiver

  06    terminal 3 transmitter

  05    terminal 2 receiver

  04    terminal 2 transmitter

  03    terminal 1 receiver

  02    terminal 1 transmitter

  01    terminal 0 receiver

  00    terminal 0 transmitter

10-May-2004

SIGINT handler installed (for use within the control window to regain

control in case of a runaway program).

In order to use a single timer callback routine for all terminals, the

callback routine must get an argument, telling which terminal is doing

the callback. So ALL timer callbacks now have such an argument, which

can be used arbitrarily by the callback initiator to communicate an

integer to the callback routine.

17-May-2004

The new terminal implementation must be tested as well as the changed

interrupt priority scheme. And even more important: allocation of the

data segment on a page boundary (see 21-Apr-2004) must be verified!

Done for the 'tst' directory.

18-May-2004

Done for the 'eos' directory.

27-Feb-2005

ECO32 Version 0.5 released.

28-Feb-2005

Despite the long time since the release of the previous version this

was very premature: I did not read the comments above and forgot to

conduct many of the tests. And indeed, the bootstrap does not work

any longer.

01-Mar-2005

The bootstrap manager should not be compiled to any standard executable

format, because this part of the boot process is independent of any

operating system. It should be compiled to plain binary format, as is

the master boot record. Now we are facing a dilemma: We want to write

the boot manager (mostly) in C and therefore will have a separate data

segment, aligned on a page boundary. But we also want a headerless

binary output, so there will be no loading information. All gaps in

the loaded program and data will therefore be filled with zeroes.

Then either we must define another executable format in which the data

segment follows directly after the code segment (without padding to the

next page boundary, see discussion above) or we have to spend somewhat

more space on disk for the bootstrap managers's program/data image.

I think I will do the latter.

04-Mar-2005

The solution above is not very elegant: The same reasoning applies also

to the bootstrap within a partition. Then we are forced to reserve two

blocks instead of the traditional single one for the bootstrap. This is

not totally out of question, but not very appealing either.

There is another solution, if we accept the boot manager to be compiled

to standard executable format: write a very compact loader for this format

and squeeze it into the master boot record. Difficult, but it can be

done, as I proved today.

Changes from above reverted.

Disk partitioning, master boot record, and boot manager done.

05-Mar-2005

Boot record and bootstrap loader done.

08-Mar-2005

I think that today I completed the tests and changes which I begun

on 17-May-2004.

ECO32 Version 0.6 released.

27-Oct-2005

ECO32 Version 0.7 released.

This was done in order to have a stable release on which to build

a simplified version of the ECO32, which is named ECO32e (the 'e'

stands for 'embedded'). The ECO32e will get implemented in an FPGA.

26-Jun-2006

We do have an error in the implementation of the SAR instructions,

really! The C standard says that the result of shifts with shift

amounts greater than or equal to (!) the number of bits in the left

expression's type is undefined. Consequently, in the CPU implementation

cpu.c we have to replace

smsk = (MSB ? 0xFFFFFFFF << (32 - scnt) : 0);

in the SAR and SARI instructions with

smsk = (MSB ? ~(((Word) 0xFFFFFFFF) >> scnt) : 0);

because if scnt == 0 then 0xFFFFFFFF << (32 - scnt) is undefined.

Thanks to Rolf Viehmann for pointing this out.

27-Jun-2006

The rfx instruction must not be executed in user mode, because it

writes the previous mode and interrupt enable bits of the PSW over

the current ones. Therefore rfx becomes a privileged instruction

which triggers EXC_PRV_INSTRCT if executed in user mode.

Again thanks to Rolf Viehmann for recognizing the potential danger

if this instruction is executed in user mode.

09-Jul-2006

Pending interrupts are now displayed when executing the 'register'

command. It is not meaningful to display pending exceptions, however,

because these are immediately handled within the longjump framework

and thus would never show anything else but 'off'.

10-Jul-2006

We have a new command to display (and possibly change) the value of

the PSW. This comes in very handy if debugging in user mode and thus

not having access to privileged addresses.

12-Jul-2006

There is a new control bit in the PSW, the 'vector bit' V, which directs

interrupts to ROM_BASE + 4 if it is off and to RAM_BASE + 4 if it is on.

This bit is cleared when the CPU is reset. In this way the software can

decide whether the interrupt service routine is located in ROM or in RAM.

15-Jul-2006

I decided to give "user TLB misses" an extra interrupt service routine

entry point. These are TLB misses triggered by an access to a virtual

address with MSB = 0, regardless of execution mode (user or kernel).

22-Jul-2006

The physical address space (2^30 bytes) is partitioned as follows

(in order to be compatible with the ECO32e):

RAM    virt start 0xC0000000    phys start 0x00000000    size 0x20000000

ROM    virt start 0xE0000000    phys start 0x20000000    size 0x10000000

I/O    virt start 0xF0000000    phys start 0x30000000    size 0x10000000

The I/O address space is evenly divided into 256 devices which may

occupy up to 1MByte each.

30-Jul-2006

We now have two control bits in the outgoing part of every entry of

the TLB: a "valid" bit (which triggers a "TLB invalid exception" if

this entry is selected by a memory translation and the bit is 0) and

a "write" bit (which triggers a "TLB write exception" if this entry

is used to map a write access to a page and the bit is 0). The "TLB

double hit exception" does not exist any longer.

The reason for existence of the "write" bit is obvious (possibility

to protect pages from writing, implement a copy-on-write policy).

The reason for existence of the "valid" bit is speed: the TLB refill

routine should run as fast as possible and therefore should avoid

tests for special cases as much as possible. If some processing must

be done for a page to be usable, the check for this condition should

not be done in the TLB miss handler. If the valid bit is cleared in

the page table, the refill routine will enter the translation that it

gets from the page table into the TLB without checking, but the next

use of this entry will lead to another exception, where the processing

then can be done.

31-Jul-2006

The implementation of the ECO32e suggested a different coding of the