xv6-65oo2/labs/syscall.html

<html>
<head>
<title>Lab: Alarm and uthread</title>
<link rel="stylesheet" href="homework.css" type="text/css" />
</head>
<body>

<h1>Lab: Alarm and uthread</h1>

This lab makes you familiar with the implementation of system calls
and switching between threads of execution.  In particular, you will
implement new system calls (<tt>sigalarm</tt> and <tt>sigreturn</tt>)
and switching between threads of a user-level thread package.

<h2>Warmup: RISC-V assembly</h2>

<p>For this lab it will be important to understand a bit of RISC-V assembly.

<p>Add a file user/call.c with the following content, modify the
  Makefile to add the program to the user programs, and compile (make
  fs.img).  The Makefile also produces a binary and a readable
  assembly a version of the program in the file user/call.asm.
<pre>
#include "kernel/param.h"
#include "kernel/types.h"
#include "kernel/stat.h"
#include "user/user.h"

int g(int x) {
  return x+3;
}

int f(int x) {
  return g(x);
}

void main(void) {
  printf(1, "%d %d\n", f(8)+1, 13);
  exit();
}
</pre>

<p>Read through user/call.asm and understand it.  The instruction manual
  for RISC-V is in the doc directory (doc/riscv-spec-v2.2.pdf).  Here
  are some questions that you should answer for yourself:

  <ul>
    <li>Which registers contain arguments to functions?  Which
    register holds 13 in the call to <tt>printf</tt>?  Which register
    holds the second argument? Which register holds the third one?  Etc.

    <li>Where is the function call to <tt>f</tt> from main? Where
        is the call to <tt>g</tt>?
        (Hint: the compiler may inline functions.)

    <li>At what address is the function <tt>printf</tt> located?

    <li>What value is in the register <tt>ra</tt> just after the <tt>jalr</tt>
    to <tt>printf</tt> in <tt>main</tt>?
  </ul>

<h2>Warmup: system call tracing</h2>

<p>In this exercise you will modify the xv6 kernel to print out a line
for each system call invocation. It is enough to print the name of the
system call and the return value; you don't need to print the system
call arguments.

<p>
When you're done, you should see output like this when booting
xv6:

<pre>
...
fork -> 2
exec -> 0
open -> 3
close -> 0
$write -> 1
 write -> 1
</pre>

<p>
That's init forking and execing sh, sh making sure only two file descriptors are
open, and sh writing the $ prompt.  (Note: the output of the shell and the
system call trace are intermixed, because the shell uses the write syscall to
print its output.)

<p> Hint: modify the syscall() function in kernel/syscall.c.

<p>Run the xv6 programs you wrote in earlier labs and inspect the system call
  trace.  Are there many system calls?  Which system calls correspond
  to code in the applications you wrote?
    
<p>Optional: print the system call arguments.

  
<h2>Alarm</h2>

<p>
In this exercise you'll add a feature to xv6 that periodically alerts
a process as it uses CPU time. This might be useful for compute-bound
processes that want to limit how much CPU time they chew up, or for
processes that want to compute but also want to take some periodic
action. More generally, you'll be implementing a primitive form of
user-level interrupt/fault handlers; you could use something similar
to handle page faults in the application, for example.

<p>
You should add a new <tt>sigalarm(interval, handler)</tt> system call.
If an application calls <tt>sigalarm(n, fn)</tt>, then after every
<tt>n</tt> "ticks" of CPU time that the program consumes, the kernel
should cause application function
<tt>fn</tt> to be called. When <tt>fn</tt> returns, the application
should resume where it left off. A tick is a fairly arbitrary unit of
time in xv6, determined by how often a hardware timer generates
interrupts.

<p>
You should put the following test program in <tt>user/alarmtest.c</tt>:

<b>XXX Insert the final program here; maybe just give the code in the repo</b>
<pre>
#include "kernel/param.h"
#include "kernel/types.h"
#include "kernel/stat.h"
#include "kernel/riscv.h"
#include "user/user.h"

void test0();
void test1();
void periodic();

int
main(int argc, char *argv[])
{
  test0();
  test1();
  exit();
}

void test0()
{
  int i;
  printf(1, "test0 start\n");
  sigalarm(2, periodic);
  for(i = 0; i < 1000*500000; i++){
    if((i % 250000) == 0)
      write(2, ".", 1);
  }
  sigalarm(0, 0);
  printf(1, "test0 done\n");
}

void
periodic()
{
  printf(1, "alarm!\n");
}

void __attribute__ ((noinline)) foo(int i, int *j) {
  if((i % 2500000) == 0) {
    write(2, ".", 1);
  }
  *j += 1;
}

void test1() {
  int i;
  int j;

  printf(1, "test1 start\n");
  j = 0;
  sigalarm(2, periodic);
  for(i = 0; i < 1000*500000; i++){
    foo(i, &j);
  }
  if(i != j) {
    printf(2, "i %d should = j %d\n", i, j);
    exit();
  }
  printf(1, "test1 done\n");
}
</pre>

The program calls <tt>sigalarm(2, periodic1)</tt> in <tt>test0</tt> to
ask the kernel to force a call to <tt>periodic()</tt> every 2 ticks,
and then spins for a while.  After you have implemented
the <tt>sigalarm()</tt> system call in the kernel,
<tt>alarmtest</tt> should produce output like this for <tt>test0</tt>:

<b>Update output for final usertests.c</b>
<pre>
$ alarmtest
alarmtest starting
.....alarm!
....alarm!
.....alarm!
......alarm!
.....alarm!
....alarm!
....alarm!
......alarm!
.....alarm!
...alarm!
...$ 
</pre>
<p>

<p>
(If you only see one "alarm!", try increasing the number of iterations in
<tt>alarmtest.c</tt> by 10x.)

<p>The main challenge will be to arrange that the handler is invoked
  when the process's alarm interval expires.  You'll need to modify
  usertrap() in kernel/trap.c so that when a
  process's alarm interval expires, the process executes
  the handler. How can you do that?  You will need to understand in
  detail how system calls work (i.e., the code in kernel/trampoline.S
  and kernel/trap.c). Which register contains the address where
  system calls return to?

<p>Your solution will be few lines of code, but it will be tricky to
  write the right lines of code.  The most common failure scenario is that the
  user program crashes or doesn't terminate.  You can see the assembly
  code for the alarmtest program in alarmtest.asm, which will be handy
  for debugging.

<h3>Test0: invoke handler</h3>

<p>To get started, the best strategy is to first pass test0, which
  will force you to handle the main challenge above. Here are some
  hints how to pass test0:
  
<ul>

<li>You'll need to modify the Makefile to cause <tt>alarmtest.c</tt>
to be compiled as an xv6 user program.

<li>The right declarations to put in <tt>user/user.h</tt> is:
<pre>
    int sigalarm(int ticks, void (*handler)());
</pre>

<li>Update user/sys.pl (which generates user/usys.S),
    kernel/syscall.h, and kernel/syscall.c 
   to allow <tt>alarmtest</tt> to invoke the sigalarm system
  call.

<li>Your <tt>sys_sigalarm()</tt> should store the alarm interval and
the pointer to the handler function in new fields in the <tt>proc</tt>
structure; see <tt>kernel/proc.h</tt>.

<li>You'll need to keep track of how many ticks have passed since the
last call (or are left until the next call) to a process's alarm
handler; you'll need a new field in <tt>struct&nbsp;proc</tt> for this
too.  You can initialize <tt>proc</tt> fields in <tt>allocproc()</tt>
in <tt>proc.c</tt>.

<li>Every tick, the hardware clock forces an interrupt, which is handled
in <tt>usertrap()</tt>; you should add some code here.

<li>You only want to manipulate a process's alarm ticks if there's a a
  timer interrupt; you want something like
<pre>
    if(which_dev == 2) ...
</pre>

<li>Only invoke the process's alarm function, if the process has a
  timer outstanding.  Note that the address of the user's alarm
  function might be 0 (e.g., in alarmtest.asm, <tt>periodic</tt> is at
  address 0).

<li>It will be easier to look at traps with gdb if you tell qemu to
use only one CPU, which you can do by running
<pre>
    make CPUS=1 qemu
</pre>

</ul>

<h3>test1(): resume interrupted code</h3>

<p>Test0 doesn't tests whether the handler returns correctly to
  interrupted instruction in test0.  If you didn't get this right, it
  is likely that test1 will fail (the program crashes or the program
  goes into an infinite loop).

<p>A main challenge is to arrange that when the handler returns, it
  returns to the instruction where the program was interrupted.  Which
  register contains the return address of a function?  When the kernel
  receives an interrupt, which register contains the address of the
  interrupted instruction?

<p>Your solution is likely to require you to save and restore
  registers---what registers do you need to save and restore to resume
  the interrupted code correctly? (Hint: it will be many).  There are
  several ways to do this, but one convenient way is to add another
  system call <tt>sigreturn</tt> that the handler calls when it is
  done. Your job is to arrange that <tt>sigreturn</tt> returns to the
  interrupted code.

  Some hints:
  <ul>
    <li>Add the <tt>sigreturn</tt> system call, following the changes
      you made to support <tt>sigalarm</tt>.
      
    <li>Save enough state when the timer goes in the <tt>struct
      proc</tt> so that <tt>sigreturn</tt> can return to the
      interrupted code.

    <li>Prevent re-entrant calls to the handler----if a handler hasn't
      returned yet, don't call it again.
  </ul>
  
<p>Once you pass <tt>test0</tt> and <tt>test1</tt>, run usertests to
  make sure you didn't break any other parts of the kernel.

<h2>Uthread: switching between threads</h2>
  
<p>Download <a href="uthread.c">uthread.c</a> and <a
 href="uthread_switch.S">uthread_switch.S</a> into your xv6 directory.
Make sure <tt>uthread_switch.S</tt> ends with <tt>.S</tt>, not
<tt>.s</tt>.  Add the
following rule to the xv6 Makefile after the _forktest rule:

<pre>
$U/_uthread: $U/uthread.o $U/uthread_switch.o
	$(LD) $(LDFLAGS) -N -e main -Ttext 0 -o $U/_uthread $U/uthread.o $U/uthread_switch.o $(ULIB)
	$(OBJDUMP) -S $U/_uthread > $U/uthread.asm
</pre>
Make sure that the blank space at the start of each line is a tab,
not spaces.

<p>
Add <tt>_uthread</tt> in the Makefile to the list of user programs defined by UPROGS.

<p>Run xv6, then run <tt>uthread</tt> from the xv6 shell. The xv6 kernel will print an error message about <tt>uthread</tt> encountering a page fault.

<p>Your job is to complete <tt>uthread_switch.S</tt>, so that you see output similar to
this (make sure to run with CPUS=1):
<pre>
~/classes/6828/xv6$ make CPUS=1 qemu
...
$ uthread
my thread running
my thread 0x0000000000002A30
my thread running
my thread 0x0000000000004A40
my thread 0x0000000000002A30
my thread 0x0000000000004A40
my thread 0x0000000000002A30
my thread 0x0000000000004A40
my thread 0x0000000000002A30
my thread 0x0000000000004A40
my thread 0x0000000000002A30
...
my thread 0x0000000000002A88
my thread 0x0000000000004A98
my thread: exit
my thread: exit
thread_schedule: no runnable threads
$
</pre>

<p><tt>uthread</tt> creates two threads and switches back and forth between
them. Each thread prints "my thread ..." and then yields to give the other
thread a chance to run. 

<p>To observe the above output, you need to complete <tt>uthread_switch.S</tt>, but before
jumping into <tt>uthread_switch.S</tt>, first understand how <tt>uthread.c</tt>
uses <tt>uthread_switch</tt>.  <tt>uthread.c</tt> has two global variables
<tt>current_thread</tt> and <tt>next_thread</tt>.  Each is a pointer to a
<tt>thread</tt> structure.  The thread structure has a stack for a thread and a
saved stack pointer (<tt>sp</tt>, which points into the thread's stack).  The
job of <tt>uthread_switch</tt> is to save the current thread state into the
structure pointed to by <tt>current_thread</tt>, restore <tt>next_thread</tt>'s
state, and make <tt>current_thread</tt> point to where <tt>next_thread</tt> was
pointing to, so that when <tt>uthread_switch</tt> returns <tt>next_thread</tt>
is running and is the <tt>current_thread</tt>.

<p>You should study <tt>thread_create</tt>, which sets up the initial stack for
a new thread. It provides hints about what <tt>uthread_switch</tt> should do.
Note that <tt>thread_create</tt> simulates saving all callee-save registers
on a new thread's stack.

<p>To write the assembly in <tt>thread_switch</tt>, you need to know how the C
compiler lays out <tt>struct thread</tt> in memory, which is as
follows:

<pre>
    --------------------
    | 4 bytes for state|
    --------------------
    | stack size bytes |
    | for stack        |
    --------------------
    | 8 bytes for sp   |
    --------------------  <--- current_thread
         ......

         ......
    --------------------
    | 4 bytes for state|
    --------------------
    | stack size bytes |
    | for stack        |
    --------------------
    | 8 bytes for sp   |
    --------------------  <--- next_thread
</pre>

The variables <tt>&next_thread</tt> and <tt>&current_thread</tt> each
contain the address of a pointer to <tt>struct thread</tt>, and are
passed to <tt>thread_switch</tt>.  The following fragment of assembly
will be useful:

<pre>
   ld t0, 0(a0)
   sd sp, 0(t0)
</pre>

This saves <tt>sp</tt> in <tt>current_thread->sp</tt>.  This works because
<tt>sp</tt> is at
offset 0 in the struct.
You can study the assembly the compiler generates for
<tt>uthread.c</tt> by looking at <tt>uthread.asm</tt>.

<p>To test your code it might be helpful to single step through your
<tt>uthread_switch</tt> using <tt>riscv64-linux-gnu-gdb</tt>.  You can get started in this way:

<pre>
(gdb) file user/_uthread
Reading symbols from user/_uthread...
(gdb) b *0x230

</pre>
0x230 is the address of uthread_switch (see uthread.asm). When you
compile it may be at a different address, so check uthread_asm.
You may also be able to type "b uthread_switch".  <b>XXX This doesn't work
  for me; why?</b>

<p>The breakpoint may (or may not) be triggered before you even run
<tt>uthread</tt>. How could that happen?

<p>Once your xv6 shell runs, type "uthread", and gdb will break at
<tt>thread_switch</tt>.  Now you can type commands like the following to inspect
the state of <tt>uthread</tt>:

<pre>
  (gdb) p/x *next_thread
  $1 = {sp = 0x4a28, stack = {0x0 (repeats 8088 times),
      0x68, 0x1, 0x0 <repeats 102 times>}, state = 0x1}
</pre>
What address is <tt>0x168</tt>, which sits on the bottom of the stack
of <tt>next_thread</tt>?

With "x", you can examine the content of a memory location
<pre>
  (gdb) x/x next_thread->sp
  0x4a28 <all_thread+16304>:      0x00000168
</pre>
Why does that print <tt>0x168</tt>?

<h3>Optional challenges</h3>

<p>The user-level thread package interacts badly with the operating system in
several ways.  For example, if one user-level thread blocks in a system call,
another user-level thread won't run, because the user-level threads scheduler
doesn't know that one of its threads has been descheduled by the xv6 scheduler.  As
another example, two user-level threads will not run concurrently on different
cores, because the xv6 scheduler isn't aware that there are multiple
threads that could run in parallel.  Note that if two user-level threads were to
run truly in parallel, this implementation won't work because of several races
(e.g., two threads on different processors could call <tt>thread_schedule</tt>
concurrently, select the same runnable thread, and both run it on different
processors.)

<p>There are several ways of addressing these problems.  One is
 using <a href="http://en.wikipedia.org/wiki/Scheduler_activations">scheduler
 activations</a> and another is to use one kernel thread per
 user-level thread (as Linux kernels do).  Implement one of these ways
 in xv6.  This is not easy to get right; for example, you will need to
 implement TLB shootdown when updating a page table for a
 multithreaded user process.

<p>Add locks, condition variables, barriers,
etc. to your thread package.
    
</body>
</html>