Wording fixes.

[pintos-anon] / doc / userprog.texi

@node Project 2--User Programs, Project 3--Virtual Memory, Project 1--Threads, Top
@chapter Project 2: User Programs

Now that you've worked with Pintos and are familiar with its
infrastructure and thread package, it's time to start working on the
parts of the system that will allow users to run programs on top of
your operating system.  The base code already supports loading and
running a single user program at a time with little interactivity
possible.  You will allow multiple programs to be loaded in at once,
and to interact with the OS via system calls.

You will be working out of the @file{userprog} directory for this
assignment.  However, you will also be interacting with almost every
other part of the code for this assignment. We will describe the
relevant parts below. If you are confident in your HW1 code, you can
build on top of it.  However, if you wish you can start with a fresh
copy of the code and re-implement @func{thread_join}, which is the
only part of project #1 required for this assignment.  Your submission
should define @code{THREAD_JOIN_IMPLEMENTED} in @file{constants.h}
(@pxref{Conditional Compilation}).

Up to now, all of the code you have written for Pintos has been part
of the operating system kernel.  This means, for example, that all the
test code from the last assignment ran as part of the kernel, with
full access to privileged parts of the system.  Once we start running
user programs on top of the operating system, this is no longer true.
This project deals with consequences of the change.

We allow more than one user program to run at a time.  Because user
programs are written and compiled to work under the illusion that they
have the entire machine, when you load into memory and run more than
one process at a time, you must manage things correctly to maintain
this illusion.

Before we delve into the details of the new code that you'll be
working with, you should probably undo the test cases from project 1.

@menu
* Project 2 Code::              
* Using the File System::       
* How User Programs Work::      
* Virtual Memory Layout::       
* Global Requirements::         
* Problem 2-1 Argument Passing::  
* Problem 2-2 System Calls::    
* User Programs FAQ::           
* 80x86 Calling Convention::    
* System Calls::                
@end menu

@node Project 2 Code
@section Code

The easiest way to get an overview of the programming you will be
doing is to simply go over each part you'll be working with.  In
@file{userprog}, you'll find a small number of files, but here is
where the bulk of your work will be:

@table @file
@item process.c
@itemx process.h
Loads ELF binaries and starts processes.

@item pagedir.c
@itemx pagedir.h
A simple manager for 80@var{x} page directories and page tables.
Although you probably won't want to modify this code for this project,
you may want to call some of its functions.  In particular,
@func{pagedir_get_page} may be helpful for accessing user memory.

@item syscall.c
@itemx syscall.h
Whenever a user process wants to access some kernel functionality, it
needs to do so via a system call.  This is a skeleton system call
handler.  Currently, it just prints a message and terminates the user
process.  In part 2 of this project you will add code to do everything
else needed by system calls.

@item exception.c
@itemx exception.h
When a user process performs a privileged or prohibited operation, it
traps into the kernel as an ``exception'' or ``fault.''@footnote{We
will treat these terms as synonymous.  There is no standard
distinction between them, although the Intel processor manuals define
them slightly differently on 80@var{x}86.}  These files handle
exceptions.  Currently all exceptions simply print a message and
terminate the process.  Some, but not all, solutions to project 2
require modifying @func{page_fault} in this file.

@item gdt.c
@itemx gdt.c
The 80@var{x}86 is a segmented architecture.  The Global Descriptor
Table (GDT) is a table that describes the segments in use.  These
files set up the GDT.  @strong{You should not need to modify these
files for any of the projects.}  However, you can read the code if
you're interested in how the GDT works.

@item tss.c
@itemx tss.c
The Task-State Segment (TSS) is used for 80@var{x}86 architectural
task switching.  Pintos uses the TSS only for switching stacks when a
user process enters an interrupt handler, as does Linux.  @strong{You
should not need to modify these files for any of the projects.}
However, you can read the code if you're interested in how the TSS
works.
@end table

Finally, in @file{lib/kernel}, you might want to use
@file{bitmap.[ch]}.  A bitmap is basically an array of bits, each of
which can be true or false.  Bitmaps are typically used to keep track
of the usage of a large array of (identical) resources: if resource
@var{n} is in use, then bit @var{n} of the bitmap is true.  You might
find it useful for tracking memory pages, for example.

@node Using the File System
@section Using the File System

You will need to use some file system code for this project.  First,
user programs are loaded from the file system.  Second, many of the
system calls you must implement deal with the file system.  However,
the focus of this project is not on the file system code, so we have
provided a simple file system in the @file{filesys} directory.  You
will want to look over the @file{filesys.h} and @file{file.h}
interfaces to understand how to use the file system, and especially
its many limitations.  @strong{You should not modify the file system
code for this project}.  Proper use of the file system routines now
will make life much easier for project 4, when you improve the file
system implementation.  Until then, you will have to put up with the
following limitations:

@itemize @bullet
@item
No synchronization.  Concurrent accesses will interfere with one
another, so external synchronization is needed.  @xref{Synchronizing
File Access}, for more details.

@item
File size is fixed at creation time.  Because the root directory is
represented as a file, the number of files that may be created is also
limited.

@item
File data is allocated as a single extent, that is, data in a single
file must occupy a contiguous range of sectors on disk.  External
fragmentation can therefore become a serious problem as a file system is
used over time.

@item
No subdirectories.

@item
File names are limited to 14 characters.

@item
A system crash mid-operation may corrupt the disk in a way
that cannot be repaired automatically.  No `fsck' tool is
provided in any case.
@end itemize

However one important feature is included:

@itemize @bullet
@item
Unix-like semantics for filesys_remove() are implemented.
That is, if a file is open when it is removed, its blocks
are not deallocated and it may still be accessed by the
threads that have it open until the last one closes it.  @xref{Removing
an Open File}, for more information.
@end itemize

You need to be able to create and format simulated disks.  The
@command{pintos} program provides this functionality with its
@option{make-disk} command.  From the @file{userprog/build} directory,
execute @code{pintos make-disk fs.dsk 2}.  This command creates a 2 MB
simulated disk named @file{fs.dsk}.  (It does not actually start
Pintos.)  Then format the disk by passing the @option{-f} option to
Pintos on the kernel's command line: @code{pintos run -f}.

You'll need a way to get files in and out of the simulated file
system.  The @code{pintos} @option{put} and @option{get} commands are
designed for this.  To copy @file{@var{file}} into the Pintos file
system, use the command @file{pintos put @var{file}}.  To copy it to
the Pintos file system under the name @file{@var{newname}}, add the
new name to the end of the command: @file{pintos put @var{file}
@var{newname}}.  The commands for copying files out of a VM are
similar, but substitute @option{get} for @option{get}.

Incidentally, these commands work by passing special options
@option{-ci} and @option{-co} on the kernel's command line and copying
to and from a special simulated disk named @file{scratch.dsk}.  If
you're very curious, you can look at the @command{pintos} program as
well as @file{filesys/fsutil.c} to learn the implementation details,
but it's really not relevant for this project.

Here's a summary of how you would create and format a disk, copy the
@command{echo} program into the new disk, and then run @command{echo}.
It assumes that you've already built the tests in
@file{tests/userprog} and that the current directory is
@file{userprog/build}:

@example
pintos make-disk fs.dsk 2
pintos run -f
pintos put ../../tests/userprog/echo echo
pintos run -ex echo
@end example

You can delete a file from the Pintos file system using the @option{-r
@var{file}} kernel option, e.g.@: @code{pintos run -r @var{file}}.
Also, @option{-ls} lists the files in the file system and @option{-p
@var{file}} prints a file's contents to the display.

@node How User Programs Work
@section How User Programs Work

Pintos can run normal C programs.  In fact, it can run any program you
want, provided it's compiled into the proper file format, and uses
only the system calls you implement.  (For example, @func{malloc}
makes use of functionality that isn't provided by any of the syscalls
we require you to support.)  The only other limitation is that Pintos
can't run programs using floating point operations, since it doesn't
include the necessary kernel functionality to save and restore the
processor's floating-point unit when switching threads.  You can look
in @file{tests/userprog} directory for some examples.

Pintos loads ELF executables, where ELF is an executable format used
by Linux, Solaris, and many other Unix and Unix-like systems.
Therefore, you can use any compiler and linker that produce
80@var{x}86 ELF executables to produce programs for Pintos.  We
recommend using the tools we provide in the @file{tests/userprog}
directory.  By default, the @file{Makefile} in this directory will
compile the test programs we provide.  You can edit the
@file{Makefile} to compile your own test programs as well.

One thing you should realize immediately is that, until you copy a
test program to the emulated disk, Pintos will be unable to do very
much useful work.  You will also find that you won't be able to do
interesting things until you copy a variety of programs to the disk.
A useful technique is to create a clean reference disk and copy that
over whenever you trash your @file{fs.dsk} beyond a useful state,
which may happen occasionally while debugging.

@node Virtual Memory Layout
@section Virtual Memory Layout

Virtual memory in Pintos is divided into two regions: user virtual
memory and kernel virtual memory.  User virtual memory ranges from
virtual address 0 up to @code{PHYS_BASE}, which is defined in
@file{threads/mmu.h} and defaults to @t{0xc0000000} (3 GB).  Kernel
virtual memory occupies the rest of the virtual address space, from
@code{PHYS_BASE} up to 4 GB.

User virtual memory is per-process.  Conceptually, each process is
free to use the entire space of user virtual memory however it
chooses.  When the kernel switches from one process to another, it
also switches user virtual address spaces by switching the processor's
page directory base register (see @func{pagedir_activate in
@file{userprog/pagedir.c}}.  @struct{thread} contains a pointer to a
process's page directory.

Kernel virtual memory is global.  It is always mapped the same way,
regardless of what user process or kernel thread is running.  In
Pintos, kernel virtual memory is mapped one-to-one to physical
memory.  That is, virtual address @code{PHYS_ADDR} accesses physical
address 0, virtual address @code{PHYS_ADDR} + @t{0x1234} access
physical address @t{0x1234}, and so on up to the size of the machine's
physical memory.

User programs can only access user virtual memory.  An attempt to
access kernel virtual memory will cause a page fault, handled by
@func{page_fault} in @file{userprog/exception.c}, and the process
will be terminated.  Kernel threads can access both kernel virtual
memory and, if a user process is running, the user virtual memory of
the running process.  However, even in the kernel, an attempt to
access memory at a user virtual address that doesn't have a page
mapped into it will cause a page fault.

You must handle memory fragmentation gracefully, that is, a process
that needs @var{N} pages of memory must not require that all @var{N}
be contiguous.  In fact, it must not require that any of the pages be
contiguous.

@node Global Requirements
@section Global Requirements

For testing and grading purposes, we have some simple overall
requirements:

@itemize @bullet
@item
The kernel should print out the program's name and exit status whenever
a process terminates, whether termination is caused by the @code{exit}
system call or for another reason.

@itemize @minus
@item
The message must be formatted exactly as if it was printed with
@code{printf ("%s: exit(%d)\n", @dots{});} given appropriate arguments.

@item
The name printed should be the full name passed to
@func{process_execute}, except that it is acceptable to truncate it to
15 characters to allow for the limited space in @struct{thread}.  The
name printed need not include arguments.

@item
Do not print a message when a kernel thread that is not a process
terminates.

@item
Do not print messages about process termination for the @code{halt}
system call.

@item
No message need be printed when a process fails to load.
@end itemize

@item
Aside from this, the kernel should print out no other messages that
Pintos as provided doesn't already print.  You
may understand all those debug messages, but we won't, and it just
clutters our ability to see the stuff we care about.

@item
Additionally, while it may be useful to hard-code which process will
run at startup while debugging, before you submit your code you must
make sure that it takes the start-up process name and arguments from
the @samp{-ex} argument.  For example, running @code{pintos run -ex
"testprogram 1 2 3 4"} will spawn @samp{testprogram 1 2 3 4} as the
first process.
@end itemize

@node Problem 2-1 Argument Passing
@section Problem 2-1: Argument Passing

Currently, @func{process_execute} does not support passing arguments
to new processes.  UNIX and other operating systems do allow passing
command line arguments to a program, which accesses them via the argc,
argv arguments to main.  You must implement this functionality by
extending @func{process_execute} so that instead of simply taking a
program file name as its argument, it divides it into words at spaces.
The first word is the program name, the second word is the first
argument, and so on.  That is, @code{process_execute("grep foo bar")}
should run @command{grep} passing two arguments @code{foo} and
@file{bar}.  A few details:

@itemize
@item
Multiple spaces are considered the same as a single space, so that
@code{process_execute("grep foo bar")} would be equivalent to our
original example.

@item
You can impose a reasonable limit on the length of the command line
arguments.  For example, you could limit the arguments to those that
will fit in a single page (4 kB).

@item
You can parse the argument strings any way you like.  If you're lost,
look at @func{strtok_r}, prototyped in @file{lib/string.h} and
implemented with thorough comments in @file{lib/string.c}.  You can
find more about it by looking at the man page (run @code{man strtok_r}
at the prompt).

@item
@xref{80x86 Calling Convention}, for information on exactly how you
need to set up the stack.
@end itemize

@strong{This functionality is extremely important.}  Almost all our
test cases rely on being able to pass arguments, so if you don't get
this right, a lot of things will not appear to work correctly with our
tests.  If the tests fail, so do you.  Fortunately, this part
shouldn't be too hard.

@node Problem 2-2 System Calls
@section Problem 2-2: System Calls

Implement the system call handler in @file{userprog/syscall.c} to
properly deal with all the system calls described below.  Currently,
it ``handles'' system calls by terminating the process.  You will need
to decipher system call arguments and take the appropriate action for
each.

You are required to support the following system calls, whose syscall
numbers are defined in @file{lib/syscall-nr.h} and whose C functions
called by user programs are prototyped in @file{lib/user/syscall.h}:

@table @code
@item SYS_halt
@itemx void halt (void)
Stops Pintos by calling @func{power_off} (declared in
@file{threads/init.h}).  Note that this should be seldom used, since
then you lose some information about possible deadlock situations,
etc.

@item SYS_exit
@itemx void exit (int @var{status})
Terminates the current user program, returning @var{status} to the
kernel.  If the process's parent @func{join}s it, this is the status
that will be returned.  Conventionally, a @var{status} of 0 indicates
a successful exit.  Other values may be used to indicate user-defined
conditions (usually errors).

@item SYS_exec
@itemx pid_t exec (const char *@var{cmd_line})
Runs the executable whose name is given in @var{cmd_line}, passing any
given arguments, and returns the new process's program id (pid).  If
there is an error loading this program, may return pid -1, which
otherwise should not be a valid id number.

@item SYS_join
@itemx int join (pid_t @var{pid})
Joins the process @var{pid}, using the join rules from the last
assignment, and returns the process's exit status.  If the process was
terminated by the kernel (i.e.@: killed due to an exception), the exit
status should be -1.  If the process was not a child of the calling
process, the return value is undefined (but kernel operation must not
be disrupted).

@item SYS_create
@itemx bool create (const char *@var{file}, unsigned @var{initial_size})
Create a new file called @var{file} initially @var{initial_size} bytes
in size.  Returns true if successful, false otherwise.

@item SYS_remove
@itemx bool remove (const char *@var{file})
Delete the file called @var{file}.  Returns true if successful, false
otherwise.

@item SYS_open
@itemx int open (const char *@var{file})
Open the file called @var{file}.  Returns a nonnegative integer handle
called a ``file descriptor'' (fd), or -1 if the file could not be
opened.  All open files associated with a process should be closed
when the process exits or is terminated.

File descriptors numbered 0 and 1 are reserved for the console: fd 0
is standard input (@code{stdin}), fd 1 is standard output
(@code{stdout}).  These special file descriptors are valid as system
call arguments only as explicitly described below.

@item SYS_filesize
@itemx int filesize (int @var{fd})
Returns the size, in bytes, of the file open as @var{fd}.

@item SYS_read
@itemx int read (int @var{fd}, void *@var{buffer}, unsigned @var{size})
Read @var{size} bytes from the file open as @var{fd} into
@var{buffer}.  Returns the number of bytes actually read (0 at end of
file), or -1 if the file could not be read (due to a condition other
than end of file).  Fd 0 reads from the keyboard using
@func{kbd_getc}.

@item SYS_write
@itemx int write (int @var{fd}, const void *@var{buffer}, unsigned @var{size})
Write @var{size} bytes from @var{buffer} to the open file @var{fd}.
Returns the number of bytes actually written, or -1 if the file could
not be written.   Fd 1 writes to the console.

@item SYS_seek
@itemx void seek (int @var{fd}, unsigned @var{position})
Changes the next byte to be read or written in open file @var{fd} to
@var{position}, expressed in bytes from the beginning of the file.
(Thus, a @var{position} of 0 is the file's start.)

A seek past the current end of a file is not an error.  A later read
obtains 0 bytes, indicating end of file.  A later write extends the
file, filling any unwritten gap with zeros.  (However, in Pintos files
have a fixed length until project 4 is complete, so writes past end of
file will return an error.)  These semantics are implemented in the
file system and do not require any special effort in system call
implementation.

@item SYS_tell
@itemx unsigned tell (int @var{fd})
Returns the position of the next byte to be read or written in open
file @var{fd}, expressed in bytes from the beginning of the file.

@item SYS_close
@itemx void close (int @var{fd})
Close file descriptor @var{fd}.
@end table

The file defines other syscalls.  Ignore them for now.  You will
implement some of them in project 3 and the rest in project 4, so be
sure to design your system with extensibility in mind.

To implement syscalls, you will need to provide a way of copying data
from the user's virtual address space into the kernel and vice versa.
This can be a bit tricky: what if the user provides an invalid
pointer, a pointer into kernel memory, or points to a block that is
partially in one of those regions?  You should handle these cases by
terminating the user process.  You will need this code before you can
even obtain the system call number, because the system call number is
on the user's stack in the user's virtual address space.  We recommend
writing and testing this code before implementing any other system
call functionality.

@anchor{Synchronizing File Access}
You must make sure that system calls are properly synchronized so that
any number of user processes can make them at once.  In particular, it
is not safe to call into the filesystem code provided in the
@file{filesys} directory from multiple threads at once.  For now, we
recommend adding a single lock that controls access to the filesystem
code.  You should acquire this lock before calling any functions in
the @file{filesys} directory, and release it afterward.  Don't forget
that @func{process_execute} also accesses files.  @strong{For now, we
recommend against modifying code in the @file{filesys} directory.}

We have provided you a user-level function for each system call in
@file{lib/user/syscall.c}.  These provide a way for user processes to
invoke each system call from a C program.  Each of them calls an
assembly language routine in @file{lib/user/syscall-stub.S}, which in
turn invokes the system call interrupt and returns.

When you're done with this part, and forevermore, Pintos should be
bulletproof.  Nothing that a user program can do should ever cause the
OS to crash, halt, assert fail, or otherwise stop running.  It is
important to emphasize this point: our tests will try to break your
system calls in many, many ways.  You need to think of all the corner
cases and handle them.  The sole way a user program should be able to
cause the OS to halt is by invoking the @code{halt} system call.

If a system call is passed an invalid argument, acceptable options
include returning an error value (for those calls that return a
value), returning an undefined value, or terminating the process.

@xref{System Calls}, for more information on how syscalls work.

@node User Programs FAQ
@section FAQ

@enumerate 1
@item
@b{Do we need a working project 1 to implement project 2?}

You may find the code for @func{thread_join} to be useful in
implementing the join syscall, but besides that, you can use
the original code provided for project 1.

@item
@b{@samp{pintos put} always panics.}

Here are the most common causes:

@itemize @bullet
@item
The disk hasn't yet been formatted (with @samp{pintos run -f}).

@item
The file name specified is too long.  The file system limits file names
to 14 characters.  If you're using a command like @samp{pintos put
../../tests/userprog/echo}, that overflows the limit.  Use
@samp{pintos put ../../tests/userprog/echo echo} to put the file under
the name @file{echo} instead.

@item
The file system is full.

@item
The file system already contains 10 files.  (There's a 10-file limit for
the base Pintos file system.)

@item
The file system is so fragmented that there's not enough contiguous
space for your file.
@end itemize

@item
@b{All my user programs die with page faults.}

This will generally happen if you haven't implemented problem 2-1
yet.  The reason is that the basic C library for user programs tries
to read @var{argc} and @var{argv} off the stack.  Because the stack
isn't properly set up yet, this causes a page fault.

@item
@b{I implemented 2-1 and now all my user programs die with
@samp{system call!}.}

Every reasonable program tries to make at least one system call
(@func{exit}) and most programs make more than that.  Notably,
@func{printf} invokes the @code{write} system call.  The default
system call handler just prints @samp{system call!} and terminates the
program.  You'll have to implement 2-2 before you see anything more
interesting.  Until then, you can use @func{hex_dump} to convince
yourself that 2-1 is implemented correctly (@pxref{Argument Passing to
main}).

@item
@b{Is there a way I can disassemble user programs?}

The @command{i386-elf-objdump} utility can disassemble entire user
programs or object files.  Invoke it as @code{i386-elf-objdump -d
@var{file}}.  You can also use @code{i386-elf-gdb}'s
@command{disassemble} command to disassemble individual functions in
object files compiled with debug information.

@item
@b{Why can't I use many C include files in my Pintos programs?}

The C library we provide is very limited.  It does not include many of
the features that are expected of a real operating system's C library.
The C library must be built specifically for the operating system (and
architecture), since it must make system calls for I/O and memory
allocation.  (Not all functions do, of course, but usually the library
is compiled as a unit.)

@item
@b{Can I use lib@var{foo} in my Pintos programs?}

The chances are good that lib@var{foo} uses parts of the C library
that Pintos doesn't implement.  It will probably take at least some
porting effort to make it work under Pintos.  Notably, the Pintos
userland C library does not have a @func{malloc} implementation.

@item
@b{How do I compile new user programs?}

You need to modify @file{tests/Makefile}.

@item
@b{What's the difference between @code{tid_t} and @code{pid_t}?}

A @code{tid_t} identifies a kernel thread, which may have a user
process running in it (if created with @func{process_execute}) or not
(if created with @func{thread_create}).  It is a data type used only
in the kernel.

A @code{pid_t} identifies a user process.  It is used by user
processes and the kernel in the @code{exec} and @code{join} system
calls.

You can choose whatever suitable types you like for @code{tid_t} and
@code{pid_t}.  By default, they're both @code{int}.  You can make them
a one-to-one mapping, so that the same values in both identify the
same process, or you can use a more complex mapping.  It's up to you.

@item
@b{I can't seem to figure out how to read from and write to user
memory. What should I do?}

The kernel must treat user memory delicately.  As part of a system
call, the user can pass to the kernel a null pointer, a pointer to
unmapped virtual memory, or a pointer to kernel virtual address space
(above @code{PHYS_BASE}).  All of these types of invalid pointers must
be rejected without harm to the kernel or other running processes.  At
your option, the kernel may handle invalid pointers by terminating the
process or returning from the system call with an error.

There are at least two reasonable ways to do this correctly.  The
first method is to ``verify then access'':@footnote{These terms are
made up for this document.  They are not standard terminology.} verify
the validity of a user-provided pointer, then dereference it.  If you
choose this route, you'll want to look at the functions in
@file{userprog/pagedir.c} and in @file{threads/mmu.h}.  This is the
simplest way to handle user memory access.

The second method is to ``assume and react'': directly dereference
user pointers, after checking that they point below @code{PHYS_BASE}.
Invalid user pointers will then cause a ``page fault'' that you can
handle by modifying the code for @func{page_fault} in
@file{userprog/exception.cc}.  This technique is normally faster
because it takes advantage of the processor's MMU, so it tends to be
used in real kernels (including Linux).

In either case, you need to make sure not to ``leak'' resources.  For
example, suppose that your system call has acquired a lock or
allocated a page of memory.  If you encounter an invalid user pointer
afterward, you must still be sure to release the lock or free the page
of memory.  If you choose to ``verify then access,'' then this should
be straightforward, but for ``assume and react'' it's more difficult,
because there's no way to return an error code from a memory access.
Therefore, for those who want to try the latter technique, we'll
provide a little bit of helpful code:

@verbatim
/* Tries to copy a byte from user address USRC to kernel address DST.
   Returns true if successful, false if USRC is invalid. */
static inline bool get_user (uint8_t *dst, const uint8_t *usrc) {
  int eax;
  asm ("movl $1f, %%eax; movb %2, %%al; movb %%al, %0; 1:"
       : "=m" (*dst), "=&a" (eax) : "m" (*usrc));
  return eax != 0;
}

/* Tries write BYTE to user address UDST.
   Returns true if successful, false if UDST is invalid. */
static inline bool put_user (uint8_t *udst, uint8_t byte) {
  int eax;
  asm ("movl $1f, %%eax; movb %b2, %0; 1:"
       : "=m" (*udst), "=&a" (eax) : "r" (byte));
  return eax != 0;
}
@end verbatim

Each of these functions assumes that the user address has already been
verified to be below @code{PHYS_BASE}.  They also assume that you've
modified @func{page_fault} so that a page fault in the kernel causes
@code{eax} to be set to 0 and its former value copied into @code{eip}.

@item
@b{I'm also confused about reading from and writing to the stack. Can
you help?}

@itemize @bullet
@item
Only non-@samp{char} values will have issues when writing them to
memory.  If a digit is in a string, it is considered a character.
However, the value of @code{argc} would be a non-char.

@item
You will need to write characters and non-characters into main memory.

@item
When you add items to the stack, you will be decrementing the stack
pointer.  You'll need to decrement the stack pointer before writing to
the location.

@item
Each character is 1 byte.
@end itemize

@item
@b{Why doesn't keyboard input work with @samp{pintos -v}?}

Serial input isn't implemented.  Don't use @samp{pintos -v} if you
want to use the shell or otherwise provide keyboard input.
@end enumerate

@menu
* Problem 2-1 Argument Passing FAQ::  
* Problem 2-2 System Calls FAQ::  
@end menu

@node Problem 2-1 Argument Passing FAQ
@subsection Problem 2-1: Argument Passing FAQ

@enumerate 1
@item
@b{Why is the top of the stack at @t{0xc0000000}?  Isn't that off the
top of user virtual memory?  Shouldn't it be @t{0xbfffffff}?}

When the processor pushes data on the stack, it decrements the stack
pointer first.  Thus, the first (4-byte) value pushed on the stack
will be at address @t{0xbffffffc}.

Also, the stack should always be aligned to a 4-byte boundary, but
@t{0xbfffffff} isn't.

@item
@b{Is @code{PHYS_BASE} fixed?}

No.  You should be able to support @code{PHYS_BASE} values that are
any multiple of @t{0x10000000} from @t{0x80000000} to @t{0xc0000000},
simply via recompilation.
@end enumerate

@node Problem 2-2 System Calls FAQ
@subsection Problem 2-2: System Calls FAQ

@enumerate 1
@item
@b{Can I just cast a pointer to a @struct{file} object to get a
unique file descriptor?  Can I just cast a @code{struct thread *} to a
@code{pid_t}?  It's so much simpler that way!}

This is a design decision you will have to make for yourself.
However, note that most operating systems do distinguish between file
descriptors (or pids) and the addresses of their kernel data
structures.  You might want to give some thought as to why they do so
before committing yourself.

@item
@b{Can I set a maximum number of open files per process?}

From a design standpoint, it would be better not to set an arbitrary
maximum.  That said, if your design calls for it, you may impose a
limit of 128 open files per process (as the Solaris machines here do).

@item
@anchor{Removing an Open File}
@b{What happens when two (or more) processes have a file open and one of
them removes it?}

You should copy the standard Unix semantics for files.  That is, when
a file is removed an process which has a file descriptor for that file
may continue to do operations on that descriptor.  This means that
they can read and write from the file.  The file will not have a name,
and no other processes will be able to open it, but it will continue
to exist until all file descriptors referring to the file are closed
or the machine shuts down.

@item
@b{I've discovered that some of my user programs need more than one 4
kB page of stack space.  What should I do?}

You may modify the stack setup code to allocate more than one page of
stack space for each process.
@end enumerate

@node 80x86 Calling Convention
@section 80@var{x}86 Calling Convention

What follows is a quick and dirty discussion of the 80@var{x}86
calling convention.  Some of the basics should be familiar from CS
107, and if you've already taken CS 143 or EE 182, then you should
have seen even more of it.  I've omitted some of the complexity, since
this isn't a class in how function calls work, so don't expect this to
be exactly correct in full, gory detail.  If you do want all the
details, you can refer to @bibref{SysV-i386}.

Whenever a function call happens, you need to put the arguments on the
call stack for that function, before the code for that function
executes, so that the callee has access to those values.  The caller
has to be responsible for this (be sure you understand why).
Therefore, when you compile a program, the assembly code emitted will
have in it, before every function call, a bunch of instructions that
prepares for the call in whatever manner is conventional for the
machine you're working on.  This includes saving registers as needed,
putting stuff on the stack, saving the location to return to somewhere
(so that when the callee finishes, it knows where the caller code is),
and some other bookkeeping stuff.  Then you do the jump to the
callee's code, and it goes along, assuming that the stack and
registers are prepared in the appropriate manner.  When the callee is
done, it looks at the return location as saved earlier, and jumps back
to that location.  The caller may then have to do some cleanup:
clearing arguments and the return value off the stack, restoring
registers that were saved before the call, and so on.

If you think about it, some of these things should remind you of
context switching.

As an aside, in general, function calls are not cheap.  You have to do
a bunch of memory writes to prepare the stack, you need to save and
restore registers before and after a function call, you need to write
the stack pointer, you have a couple of jumps which probably wrecks
some of your caches.  This is why inlining code can be much faster.

@menu
* Argument Passing to main::    
@end menu

@node Argument Passing to main
@subsection Argument Passing to @code{main()}

In @func{main}'s case, there is no caller to prepare the stack
before it runs.  Therefore, the kernel needs to do it.  Fortunately,
since there's no caller, there are no registers to save, no return
address to deal with, etc.  The only difficult detail to take care of,
after loading the code, is putting the arguments to @func{main} on
the stack.

(The above is a small lie: most compilers will emit code where main
isn't strictly speaking the first function.  This isn't an important
detail.  If you want to look into it more, try disassembling a program
and looking around a bit.  However, you can just act as if
@func{main} is the very first function called.)

Pintos is written for the 80@var{x}86 architecture.  Therefore, we
need to adhere to the 80@var{x}86 calling convention.  Basically, you
put all the arguments on the stack and move the stack pointer
appropriately.  You also need to insert space for the function's
``return address'': even though the initial function doesn't really
have a caller, its stack frame must have the same layout as any other
function's.  The program will assume that the stack has been laid out
this way when it begins running.

So, what are the arguments to @func{main}? Just two: an @samp{int}
(@code{argc}) and a @samp{char **} (@code{argv}).  @code{argv} is an
array of strings, and @code{argc} is the number of strings in that
array.  However, the hard part isn't these two things.  The hard part
is getting all the individual strings in the right place.  As we go
through the procedure, let us consider the following example command:
@samp{/bin/ls -l foo bar}.

The first thing to do is to break the command line into individual
strings: @samp{/bin/ls}, @samp{-l}, @samp{foo}, and @samp{bar}.  These
constitute the arguments of the command, including the program name
itself (which belongs in @code{argv[0]}).

These individual, null-terminated strings should be placed on the user
stack.  They may be placed in any order, as you'll see shortly,
without affecting how main works, but for simplicity let's assume they
are in reverse order (keeping in mind that the stack grows downward on
an 80@var{x}86 machine).  As we copy the strings onto the stack, we
record their (virtual) stack addresses.  These addresses will become
important when we write the argument vector (two paragraphs down).

After we push all of the strings onto the stack, we adjust the stack
pointer so that it is word-aligned: that is, we move it down to the
next 4-byte boundary.  This is required because we will next be
placing several words of data on the stack, and they must be aligned
to be read correctly.  In our example, as you'll see below,
the strings start at address @t{0xffed}.  One word below that would be
at @t{0xffe9}, so we could in theory put the next word on the stack
there.  However, since the stack pointer should always be
word-aligned, we instead leave the stack pointer at @t{0xffe8}.

Once we align the stack pointer, we then push the elements of the
argument vector, that is, a null pointer, then the addresses of the
strings @samp{/bin/ls}, @samp{-l}, @samp{foo}, and @samp{bar}) onto
the stack.  This must be done in reverse order, such that
@code{argv[0]} is at the lowest virtual address, again because the
stack is growing downward.  (The null pointer pushed first is because
@code{argv[argc]} must be a null pointer.)  This is because we are now
writing the actual array of strings; if we write them in the wrong
order, then the strings will be in the wrong order in the array.  This
is also why, strictly speaking, it doesn't matter what order the
strings themselves are placed on the stack: as long as the pointers
are in the right order, the strings themselves can really be anywhere.
After we finish, we note the stack address of the first element of the
argument vector, which is @code{argv} itself.

Then we push @code{argv} (that is, the address of the first element of
the @code{argv} array) onto the stack, along with the length of the
argument vector (@code{argc}, 4 in this example).  This must also be
done in this order, since @code{argc} is the first argument to
@func{main} and therefore is on first (smaller address) on the
stack.  Finally, we push a fake ``return address'' and leave the stack
pointer to point to its location.

All this may sound very confusing, so here's a picture which will
hopefully clarify what's going on. This represents the state of the
stack and the relevant registers right before the beginning of the
user program (assuming for this example that the stack bottom is
@t{0xc0000000}):

@html
<CENTER>
@end html
@multitable {@t{0xbfffffff}} {``return address''} {@t{/bin/ls\0}}
@item Address @tab Name @tab Data
@item @t{0xbffffffc} @tab @code{*argv[3]} @tab @samp{bar\0}
@item @t{0xbffffff8} @tab @code{*argv[2]} @tab @samp{foo\0}
@item @t{0xbffffff5} @tab @code{*argv[1]} @tab @samp{-l\0}
@item @t{0xbfffffed} @tab @code{*argv[0]} @tab @samp{/bin/ls\0}
@item @t{0xbfffffec} @tab word-align @tab @samp{\0}
@item @t{0xbfffffe8} @tab @code{argv[4]} @tab @t{0}
@item @t{0xbfffffe4} @tab @code{argv[3]} @tab @t{0xbffffffc}
@item @t{0xbfffffe0} @tab @code{argv[2]} @tab @t{0xbffffff8}
@item @t{0xbfffffdc} @tab @code{argv[1]} @tab @t{0xbffffff5}
@item @t{0xbfffffd8} @tab @code{argv[0]} @tab @t{0xbfffffed}
@item @t{0xbfffffd4} @tab @code{argv} @tab @t{0xbfffffd8}
@item @t{0xbfffffd0} @tab @code{argc} @tab 4
@item @t{0xbfffffcc} @tab ``return address'' @tab 0
@end multitable
@html
</CENTER>
@end html

In this example, the stack pointer would be initialized to
@t{0xbfffffcc}.

As shown above, your code should start the stack at the very top of
the user virtual address space, in the page just below virtual address
@code{PHYS_BASE} (defined in @file{threads/mmu.h}).

You may find the non-standard @func{hex_dump} function, declared in
@file{<stdio.h>}, useful for debugging your argument passing code.
Here's what it would show in the above example, given that
@code{PHYS_BASE} is @t{0xc0000000}:

@verbatim
bfffffc0                                      00 00 00 00 |            ....|
bfffffd0  04 00 00 00 d8 ff ff bf-ed ff ff bf f5 ff ff bf |................|
bfffffe0  f8 ff ff bf fc ff ff bf-00 00 00 00 00 2f 62 69 |............./bi|
bffffff0  6e 2f 6c 73 00 2d 6c 00-66 6f 6f 00 62 61 72 00 |n/ls.-l.foo.bar.|
@end verbatim

@node System Calls
@section System Calls

We have already been dealing with one way that the operating system
can regain control from a user program: interrupts from timers and I/O
devices.  These are ``external'' interrupts, because they are caused
by entities outside the CPU.

The operating system is also called to deal with software exceptions,
which are events generated in response to the code.  These can be
errors such as a page fault or division by zero.  However, exceptions
are also the means by which a user program can request services
(``system calls'') from the operating system.

In the 80@var{x}86 architecture, the @samp{int} instruction is the
most commonly used means for invoking system calls.  This instruction
is handled in the same way as other software exceptions.  In Pintos,
user programs invoke @samp{int $0x30} to make a system call.  The
system call number and any additional arguments are expected to be
pushed on the stack in the normal fashion before invoking the
interrupt.

The normal calling convention pushes function arguments on the stack
from right to left and the stack grows downward.  Thus, when the
system call handler @func{syscall_handler} gets control, the system
call number is in the 32-bit word at the caller's stack pointer, the
first argument is in the 32-bit word at the next higher address, and
so on.  The caller's stack pointer is accessible to
@func{syscall_handler} as the @samp{esp} member of the @code{struct
intr_frame} passed to it.

Here's an example stack frame for calling a system call numbered 10
with three arguments passed as 1, 2, and 3.  The stack addresses are
arbitrary:

@html
<CENTER>
@end html
@multitable {@t{0xbffffe7c}} {Value}
@item Address @tab Value
@item @t{0xbffffe7c} @tab 3
@item @t{0xbffffe78} @tab 2
@item @t{0xbffffe74} @tab 1
@item @t{0xbffffe70} @tab 10
@end multitable
@html
</CENTER>
@end html

In this example, the caller's stack pointer would be at
@t{0xbffffe70}.

The 80@var{x}86 convention for function return values is to place them
in the @samp{EAX} register.  System calls that return a value can do
so by modifying the @samp{eax} member of @struct{intr_frame}.

You should try to avoid writing large amounts of repetitive code for
implementing system calls.  Each system call argument, whether an
integer or a pointer, takes up 4 bytes on the stack.  You should be able
to take advantage of this to avoid writing much near-identical code for
retrieving each system call's arguments from the stack.