$24
Assignment 4: stsh stanford shell Solution
Kudos to Randal Bryant and Dave O’Hallaron of Carnegie Mellon for assignment inspiration and for parts of this handout. Huge thanks to Truman Cranor for writing the command line parser using tools you'll learn in CS143, which you should all take someday because it's amazing. You’ve all been using shells to drive your conversations with UNIX systems since the first day you logged into a myth. It’s high time we uncover the shell’s magic by leveraging
the simpleshs we built in lecture together and extending the final version to support process control, job lists, signals, pipelines, and I/O redirection—all while managing the interprocess concurrency problems that make a shell’s implementation a genuinely advanced systems programming project. There’s lots of neat code to write, and with your smarts and my love to guide you, I’m confident you can pull it off.
All coding should be done on a myth cluster machine, as that’s where we’ll be testing all assign4 submissions. You should clone the master git repository we’ve set up for you by typing:
myth62$ git clone /usr/class/cs110/repos/assign4/$USER assign4
Doing so will create an assign4 directory within your AFS space, and you can descend into that assign4 directory and code there. There's a working sanitycheck for this assignment,
and your repo includes a soft link to a fully functional solution. When in doubt about how something should work, just run my solution (which can be found at ./samples/stsh_soln) to see what it does and imitate the solution.
Builtin stsh Commands
Your Assignment 4 shell needs to support a collection of builtin commands that should execute without creating any new processes. These builtins are:
quit, which exits the shell and abandons any jobs that were still running. If there are any extra arguments after the quit, just can just ignore them.
exit, which does the same thing as quit. Extraneous arguments? Just ignore them.
fg, which prompts a stopped job to continue in the foreground or brings a running background job into the foreground. fg takes a single job number (e.g. fg 3). If the provided argument isn't a number, there's no such job with that number, or the number of arguments isn't correct, then throw an STSHException around an actionable error message and allow your shell to carry on as if you never typed anything.
bg, which prompts a stopped job to continue in the background. bg takes a single job number (e.g. bg 3). If the provided argument isn't a number, there's no such job with that number, or the number of arguments isn't correct, then throw
an STSHException around an actionable error message and allow your shell to carry on as if you never typed anything.
slay, which is used to terminate a single process (which may have many sibling processes as part of a larger pipeline). slay takes either one or two numeric arguments. If only one number is provided, it's assumed to be the pid of the process to be killed. If two numbers are provided, the first number is assumed to be the job number, and the second is assumed to be a process index within the job. So, slay 12345 would terminate (in the SIGKILL sense of terminate) the process with pid 12345. slay 2 0 would terminate the first process in the pipeline making up job 2. slay 13 8 would terminate the 9th process in the pipeline of processes making up job 13. If there are any issues (e.g., the arguments aren't numbers, or they are numbers but identify some nonexistent process, or the argument count is incorrect), then throw
an STSHException around an actionable error message and allow your shell to carry on as if you never typed anything.
halt, which has nearly the same story as slay, except that its one or two arguments identify a single process that should be halted (but not terminated) if it isn't already stopped. If it's already stopped, then don't do anything and just return.
cont, which has the same story as slay and halt, except that its one or two arguments identify a single process that should continue if it isn't already running. If it's already running, then don't do anything and just return. When you prompt a single process to continue, you're asking that it do so in the background.
jobs, which prints the job list to the console. If there are any additional arguments, then just ignore them.
quit, exit, and jobs are already implemented for you. You're responsible for implementing the others and ensuring the global job list is appropriately updated.
Getting Started
Inspect the stsh.cc file we give you. This is the only file you should need to modify. The core main function you're provided looks like this:
int main(int argc, char *argv[]) {
pid_t stshpid = getpid();
installSignalHandlers();
rlinit(argc, argv); // configures stsh-readline library so readline works properly
while (true) {
string line;
if (!readline(line)) break;
if (line.empty()) continue;
try {
pipeline p(line);
bool builtin = handleBuiltin(p);
if (!builtin) createJob(p); // createJob is initially defined as a wrapper around cout << p;
} catch (const STSHException& e) { cerr << e.what() << endl;
if (getpid() != stshpid) exit(0); // if exception is thrown from child
process, kill it
}
}
return 0;
}
The general idiom here is the same used in the simplesh examples from lecture, except that this version is in C++, and it'll support many more features. The readline function prompts the user to enter a command and a pipeline record is constructed around it. readline and pipeline (which is a different from the pipeline function you implemented for Assignment 3) are implemented via a suite of files in the stsh-parser subdirectory, and for the most part you can ignore those implementations. You should, however, be familiar with the type definitions of the command and pipeline types, though, and they are right here:
const size_t kMaxCommandLength = 32;
const size_t kMaxArguments = 32;
struct command {
char command[kMaxCommandLength + 1]; // '\0' terminated
char *tokens[kMaxArguments + 1]; // NULL-terminated array, C strings are
all '\0' terminated
};
struct pipeline {
std::string input; // empty if no input redirection file to first command
std::string output; // empty if no output redirection file from last command
std::vector<command commands;
bool background;
pipeline(const std::string& str);
~pipeline();
};
Check out what the initial version of stsh is capable of before you add any new code.
Setting Milestones
The best approach to implementing anything this complex is to invent a collection of milestones that advance you toward your final goal. Never introduce more than a few lines of code before compiling and confirming that the lines you added do what you expect. I repeat: Never introduce more than a few lines of code before compiling, testing, and confirming that the additional lines do what you expect. View everything you add as a slight perturbation to a working system that slowly evolves into the final product. Try to understand every single line you add, why it's needed, and why it belongs where you put it.
Here is a sequence of milestones I'd like you to work through in order to get started:
Descend into the stsh-parser directory, read through the stsh-readline.h and stsh-parse.h header files for data type definitions and function/method prototypes, type make, and play with the stsh-parse-test to gain a sense of what readline and the pipeline constructor do for you. In general, the readline function is like getline, except that you can use your up and down arrows to scroll through your history of inputs (neat!). The pipeline record defines a bunch of fields that store all of the various commands that chain together to form a pipeline. For example, the text cat < /usr/include/stdio.h | wc output.txt would be split into two commands—one for the cat and a second for the wc—and populate the vector<command in the pipeline with information about each of them. The input and output fields would each be nonempty, and the background field would be false.
Get a pipeline of just one command (e.g. stsh sleep 5) to run in the foreground until it's finished. Rely on a call to waitpid to stall stsh until the foreground job finishes. Ignore the job list, ignore the SIGCHLD handler, don't worry about background jobs, pipelining, or redirection. Don't worry about programs like emacs just yet. Focus on these executables instead: ls, date, sleep, as their execution is well understood and predictable.
Establish the process group ID of the job to be the PID of the process by investigating the setpgid system call. When you run stsh from the standard Unix shell, note that stsh is running in the foreground. If your shell then creates a child process, by default that child will also be a member of the foreground process group as well, and you don’t want that. Since typing ctrl-c sends a SIGINT to every process in the foreground group, typing ctrl-c will send a SIGINT to your shell, as well as to every process that your shell created, and you don't want that. The notion of a group ID isn't all that important yet, because at the moment, a process group consists of only one process. But we'll eventually be dealing with jobs comprised of many processes, and we'll want a single process group id to represent all of them.
Read through stsh-job-list.h, stsh-job.h, and stsh-process.h to learn how to add a new foreground job to the job list, and how to add a process to that job. Add code that does exactly that to the stsh.cc file, right after you successfully fork off a new process. After your waitpid call returns, remove the job from the job list. If it helps, inline cout << joblist; lines in strategically chosen locations to confirm your new job is being added after fork and being removed after waitpid.
Implement the SIGCHLD handler to reap the resources of a foreground process after it exits, and suspend stsh's main thread of execution using a sigset_t, sigprocmask, sigsuspend, and related functions until an examination of the job list proves that the
foreground job is no longer the foreground job. Your call to waitpid should be moved into the SIGCHLD handler, and that should be the only place in your entire solution— even the final one you submit—with a waitpid call.
Install functions to activate ctrl-z and ctrl-c on your keyboard to stop and kill the foreground process instead of the shell itself. If, for instance, a sleep 500 is running as the foreground, you may want to kill the process by pressing ctrl-c. When you press ctrl-c, the OS sends a SIGINT to your shell, which unless handled will terminate the shell itself. If, however, you install a function to handle SIGINT, then that handler can (and should) forward the SIGINT on to the foreground job, should one exist. The story for ctrl-z is similar, except ctrl-z prompts the OS to send your shell a SIGTSTP. If you install a custom handler to intercept a SIGTSTP, you can forward the SIGTSTP on to the foreground job, should one exist. In these cases, the kill function is your friend and will contribute to your implementation of both of these handlers.
Implement the fg builtin, which takes a stopped process—stopped presumably because it was running in the foreground at the time you pressed ctrl-z—and prompts it to continue, or it takes a process running in the background and brings it into the foreground. The fg builtin takes job number, translates that job number to a process group ID, and, if necessary, forwards a SIGCONT on to the process group via a call to kill(-groupID, SIGCONT). Right now, process groups consist of just one process, but once you start to support pipelines (e.g. cat words.txt | sort | uniq | wc -l), you'll want fg to bring the entire job into the foreground, and if all relevant processes are part of the same process group, you can achieve this with a single kill call. Of course, if the argument passed to fg isn't a number, or it is but it doesn't identify a real job, then you should throw an STSHException that's wrapped around a clear error message saying so.
Update the SIGCHLD handler to detect state changes in all processes under stsh's jurisdiction. Processes can exit gracefully, exit abnormally because of some signal, or be terminated by ctrl-c. Processes can halt because of a signal or a ctrl-z. And processes can continue because of a signal or an fg builtin. The job list needs to remain in sync with the state of your shell's world, and waitpid is the perfect function to tell you about changes. You're already familiar with WNOHANG, WUNTRACED, WIFEXITED, WEXITSTATUS, etc. Read through waitpid's man page to get word on some Linux-specific flags and macros that tell you about processes that have started executing again. Buzzwords include WCONTINUED and WIFCONTINUED, so read up on those.
Add support for background jobs. The pipeline constructor already searches for trailing &'s and records whether or not the pipeline should be run in the background. If it does, then you still add information about the job to the job list, but you immediately come back to the shell prompt without waiting.
Add support for slay, halt, cont (the process-oriented commands) and bg (which prompts all processes in a single job to continue in the background), and use some of the sample user programs I include in your repo (int, fpe, tstp, spin, split) to test all of these.
The following are additional milestones you need to hit on your way to a fully functional stsh.
Each of these bullet points represents something larger.
Add support for foreground jobs whose leading process (e.g. cat, more, emacs, vi, and other executables requiring elaborate control of the console) requires control of the terminal. You should investigate the tcsetpgrp function as a way of transferring terminal control to a process group, and update your solution to always call it, even if it fails. If tcsetpgrp(STDIN_FILENO, pgid) succeeds, then it'll return 0. If it fails with a return value of -1 but it sets errno to ENOTTY, that just means that your stsh instance didn't have control of the terminal or the authority to donate it to another process group. (This will happen if your shell is being driven by stsh-driver) If it fails with a different errno value, then that's a more serious problem that should be identified via an STSHException. If stsh succeeds to transferring control to the foreground process group, then stsh should take control back when that group falls out of the foreground (perhaps because it exits, or it stops, or something else).
Add support for pipelines consisting of two processes (i.e. binary pipelines, e.g cat / usr/include/stdio.h | wc). Make sure that the standard output of the first is piped to the standard input of the second, and that each of the two processes are part of the same process group, using a process group ID that is identical to the pid of the leading process. The pipeline function from Assignment 3 (which, again, is different from the pipeline type here) was a paying-it-forward nod to the basic pipeline functionality needed right here. You needn't go nuts on the error checking: You can assume that all system calls succeed, with the exception of execvp, which may fail because of user error (misspelled executable name, file isn't an executable, lack of permissions, etc.). You might want to include more error checking if it helps you triage bugs, assert the truth of certain expectations during execution, and arrive at a working product more quickly, but do all that because it's good for you and not because you're trying to make us happy. (Hint: the conduit user program I dropped in your repo starts to become useful as soon as you deal with nontrivial pipelines. Try typing echo 12345 | ./conduit --delay 1 in the standard shell to see what happens, and try to replicate the behavior in stsh.)
Once you get your head around pipelines of one and two processes, work on getting arbitrarily long pipeline chains to do the right thing. So, if the user types in echo 12345 | ./conduit --delay 1 | ./conduit | ./conduit, four processes are created, each with their own pid, and all in a process group whose process group id is that of the leading process (in this case, the one running echo). echo's standard out feeds the standard in of the first conduit, whose standard out feeds into the standard in of the second conduit, which pipes its output to the standard input of the last conduit, which at last prints to the console. I'm calling this out separately from the binary pipeline discussed in the previous bullet point, because the code needed to realize pipelines of three or more doesn't have as much in common with the binary pipeline code as you might think. Yes, there are pipe, setpgid, dup2, close, and execvp calls, but figuring out how to get one of the inner processes to redefine what STDIN_FILENO and STDOUT_FILENO are connected to is very tricky, and this trickiness doesn't present itself in the binary pipeline.
Finally, add support for input and output redirection via < and (e.g. cat < /usr/ include/stdio.h | wc output.txt). The names of input and output redirection files are surfaced by the pipeline constructor, and if there is a nonempty
string in the input and/or output fields of the pipeline record, that's your signal that input redirection, output redirection, or both are needed. If the file you're writing to doesn't exist, create it, and go with 0644 (with the leading zero) as the octal constant to establish the rw-r—r-- permission. If the output file you're redirecting to already exists, then truncate it using the O_TRUNC flag. Note that input redirection always impacts where the leading process draws its input from and that output redirection influences where the caboose process publishes its output. Sometimes those two processes are the same, and sometimes they are different. Type man 2 open for the full skinny on the open system call and a reminder of what flags can be bitwise-OR'ed together for the second argument.
Shell Driver
The stsh-driver program (there's a copy of it in your repo) executes stsh as a child process, sends it commands and signals as directed by a trace file, and allows the shell to print to standard output and error as it normally would. The stsh process is driven by the stsh-driver, which is why we call stsh-driver a driver.
Go ahead and type ./stsh-driver -h to learn how to use it:
Usage: ./stsh-driver [-hv] -t <trace -s <shell [-a <args]
Options:
-h Print this message
-v Output more information
-t <trace Trace file
-s <shell Version of stsh to test
-a <args Arguments to pass through to stsh implementation
myth62$
We’ve also provided several trace files that you can feed to the driver to test your stsh. If you look drill into your repo's samples symlink, you'll arrive at /usr/class/cs110/samples/ assign4, which includes not only a copy of my own stsh solution, but also a directory of shared trace files called scripts. Within scripts , you'll see simple, intermediate, and advanced subdirectories, each of which contains one or more trace files you can use for testing.
Run the shell driver on your own shell using trace file bg-spin.txt by typing this:
myth62$ ./stsh-driver -t ./samples/scripts/simple/bg-spin.txt -s ./ stsh -a "--suppress-prompt --no-history”
(the -a "--suppress-prompt --no-history" argument tells stsh to not emit a prompt or to use the fancy readline history stuff, since it confused the sanitycheck and autograder stuff).
Similarly, to compare your results with those generated by my own solution, you can run the driver on ./stsh_soln shell by typing:
myth62$ ./stsh-driver -t ./samples/scripts/simple/bg-spin.txt -s ./ samples/stsh_soln -a "--suppress-prompt —no-history"
The neat thing about the trace files is that they generate the same output you would have gotten had you run your shell interactively (except for an initial comment identifying the output as something generated via stsh-driver). For example:
myth62$ ./stsh-driver -t ./samples/scripts/advanced/simple-pipeline-1.txt -s ./samples/stsh_soln -a "--suppress-prompt --no-history"
Trace: simple-pipeline-1
------------------------
Exercises support for pipes via a foreground pipeline with
just two processes.
stsh /bin/echo abc | ./conduit --count 3 aaabbbcccdddeeefffggghhhiiijjj
The process ID’s listed as part of a trace's output will be different from run to run, but otherwise your output should be exactly the same as that generated by my solution.
Tips and Tidbits
Read every word of Chapters 1 and 2 in your B&O reader (or Chapters 8 and 10 in the full textbook).
Your implementation should be in C++ unless there's no way to avoid it. By that, I mean you should use C++ strings unless you interface with a system call that requires C strings, use cout instead of printf, and so forth. Understand that when you redefine where STDOUT_FILENO directs its text, that impacts where printf (which you're not using) and cout (which you are) place that text.
We have reasonably low expectations on error checking for this assignment, but we do have some. I want you to focus on how to leverage system directives to get a fully functional shell working, but I don't want you to examine the return value of every system call when it's more or less impossible for them to fail. In general, your error checking should guard against user error—attempts to invoke a nonexistent executable, providing out-of-range arguments to command-line built-ins, and so forth. In some cases—and I point out those cases in this handout—you do need to check the return value of a system call or two, because sometimes system call “failure” (the air quotes are intentional) isn't a true failure. You've seen situations where waitpid returns -1 even though everything was fine, and that happens with a few other system calls as well.
All unused file descriptors should be closed.
When creating a pipeline—whether it consists of a single process, two processes, or many, many processes—you need to ensure that all of the pipeline processes are in the same process group, but in a different process group than the stsh instance is. To support this, you should investigate setpgid to see how all of the sibling processes in a pipeline can be added to a new process group whose id is the same as the pid of the leading process. So, if a pipeline consists of four processes with pids 4004, 4005, 4007, and 4008, they would all be placed in a process group with an ID of 4004. By doing this, you can send a signal to every process in a group using the kill function, where the first argument is the negative of the process group id (e.g. kill(-4004, SIGTSTP)) .
In order to avoid some race conditions, you should call setpgid in the parent and in each of the children. Why does the parent need to call it? Because it needs to ensure the process group exists before it advances on to add other processes in the pipeline to the same group. Why do child processes need to call it? Because if the child relies on the parent to do it, the child may execvp (and invalidate its own pid as a valid setpgid argument) before the parent gets around to it. Race conditions: deep stuff.
You do not need to support pipelines or redirection for any of the builtins. In principle, you should be able to, say, redirect the output of jobs to a file, or pipe the output to another process, but you don't need to worry about this for Assignment 4. Our tests will never mix builtins with pipes and redirection.
Every time your SIGCHLD handler learns of a stopped or terminated process, it's possible the surrounding job is impacted. Investigate the
STSHJobList::synchronize(STSHJob& job) method, which scrutinizes the supplied job to see if it should be removed from the job list or if it should demote itself to the background.
We've defined one global variable for you (the STSHJobList called joblist). You may not declare any others unless they are constants.
Make sure pipelines of multiple commands work even when the processes spawned on their behalf ignore standard in and standard out, e.g. sleep 10 | sleep 10 | sleep 10 | sleep 10 empty-output.txt should run just fine—the entire pipeline lasts for about 10 seconds, and empty-output.txt is created or truncated and ends up being 0 bytes.
When an entire pipeline is run as a background job, make sure you print out a job summary that's consistent with the following output:
stsh sleep 10 | sleep 10 | sleep 10 & [1] 27684 27685 27686
We will be testing your submission against many more traces than the ones we provide, so be sure to define lots and lots of additional traces to verify your shell is unbreakable!
Submitting your work
Once you’re done, you should test all of your work as you normally would and then run the infamous submissions script by typing ./tools/submit.
Grading
Your assignments will be rigorously tested using the tests we expose via sanitycheck plus a whole set of others. I reserve the right to add tests and change point values if during grading I find some features aren't being exercised, but I'm fairly certain the breakdown presented below will be a very good approximation regardless.
Basic Tests (47 points)
Clean Build: 2 points
Ensure that most basic of builtins (quit, exit, jobs) work properly: 15 points
Ensure that basic foreground and background processes work: 10 points
Ensure that SIGINT and SIGTSTP handlers work on single process pipelines: 20 points
Intermediate Tests (50 points)
Ensure that bg and fg builtins are properly supported: 15 points
Ensure that SIGINT and SIGTSTP signals play well with stopped processes, restarted processes, and fg and bg builtins: 30 points
Confirm error checking to guard against user error (e.g. mistyped commands, mistyped builtins, etc): 5 points (the one test is exposed via sanity)
Advanced Tests (70 points)
Ensure that pipelines with two processes work well: 15 points
Ensure that pipelines with three or more processes work well: 25 points
Ensure redirection works for pipelines of just 1 process: 15 points
Ensure redirection works for pipelines with two or more processes: 15 points
