% -*- mode: poly-noweb; -*- \documentclass[a4paper,11pt]{article} \usepackage[text={6.75in,10in},centering]{geometry} \usepackage{graphicx} \usepackage{microtype} \DisableLigatures[-]{family=tt*} \usepackage[T1]{fontenc} \usepackage[utf8]{inputenc} \usepackage[english]{babel} \usepackage{csquotes} \usepackage{etoolbox} \AtBeginEnvironment{quote}{\small} \AtBeginEnvironment{verbatim}{\small} %% \usepackage[backend=biber]{biblatex} %% \addbibresource{refs.bib} %% \renewcommand{\cite}{\parencite} % \usepackage[hyperref]{xcolor} \usepackage[colorlinks=true,citecolor=]{hyperref} % linkcolor=red \usepackage{amsmath,amsthm,amssymb} \allowdisplaybreaks \usepackage{lmodern} \usepackage{noweb} \noweboptions{longchunks,smallcode} \DeclareMathOperator{\mdc}{mdc} \DeclareMathOperator{\gcdext}{gcdext} \DeclareMathOperator{\remainder}{remainder} \DeclareMathOperator{\quotient}{quotient} \DeclareMathOperator{\diff}{diff} \def\nwendcode{\endtrivlist \endgroup} \let\nwdocspar=\par %% Popular words. \newcommand{\lxxp}{{\tt loop}} \newcommand{\Lp}{{\tt LOOP}} \newcommand{\lp}{\Lp} \newcommand{\bug}{{\em bug}} \newcommand{\symlink}{{\em symbolic link}} \newcommand{\symlinks}{\symlink s} \title{\Lp\\ {a circle out of fashion}} \date{January 2024} \begin{document} \pdfbookmark[1]{Introduction}{intro} \fontfamily{cmr}\selectfont \maketitle %\setlength{\parskip}{3pt} %\setlength{\parindent}{0pt} \Lp\ is an out-of-fashion program, used as medium of communication by antiquated people. \Lp\ members insist that technical communication be made in writing, anxiety-free. To give you an idea, they write \Lp\ in jurassic technology---you wouldn't pay them any attention. % \begin{verbatim} Drunk and dressed in their best brown baggies and their platform soles They don't give a damn about any trumpet-playing band It ain't what they call rock and roll -- Mark Knopfler, 1978. \end{verbatim} It's easy to make a conference on the Internet. E-mail suffices. When we write an e-mail to various friends to discuss a certain subject, we form a circle. When we decide to add another person to this quickly-made conference by e-mail, sometimes we use a well-known idiom---``adding John to the loop''. We add John's e-mail address to the list of destinataries. So long as everyone replies to everyone, John too will get all the messages. If anyone violates this rule of replying to everyone involved, the loop is broken. %% \begin{figure}[!htb] %% \centering \includegraphics[width=0.8\linewidth]{images/gnus-summary.png} %% \caption{Gnus, a news reader embedded in the GNU EMACS text editor.} %% \label{fg:gnus} %% \end{figure} \begin{figure}[!htb] \centering \includegraphics[width=0.8\linewidth]{images/gnus-article.png} \caption{Gnus, a news reader embedded in the GNU EMACS text editor.} \label{fg:gnus} \end{figure} %% \begin{figure}[!htb] %% \centering \includegraphics[width=0.8\linewidth]{images/tbird-summary.png} %% \caption{Thunderbird, a news reader produced by the Mozilla Foundation.} %% \label{fg:bird} %% \end{figure} %% \begin{figure}[!htb] %% \centering \includegraphics[width=0.8\linewidth]{images/sylpheed-summary.png} %% \caption{Sylpheed, a news reader produced by Hiroyuki Yamamoto.} %% \label{fg:sylpheed} %% \end{figure} \begin{figure}[!htb] \centering \includegraphics[width=0.8\linewidth]{images/sylpheed-reading.png} \caption{Sylpheed, a news reader produced by Hiroyuki Yamamoto.} \label{fg:sylpheed} \end{figure} There are surely inconveniences in using e-mail as conference medium. For example, after John has been added to the loop, he is not able to leave by his own account. He needs to ask everyone involved to stop writing to him. This is usually easy to do, but instead people tend to ask for more technology such as mailing lists. A mailing list is nothing but an automated version of this idea of writing to various people at once. When the mailing address is written to, a program resends the message to all subscribers of the mailing list and, this way, the conference takes place. But \lp\ has nothing to do with e-mail. \Lp\ uses a communication strategy---called a ``protocol''---that is even older than the web itself. The web started out around 1989--1990 and the protocol \lp\ uses begins its history in 1979, the year in which AT\&T released UNIX V7, which was UUCP-ready, an acronym that stands for ``UNIX to UNIX copy''. UUCP is both a protocol and a set of programs for copying files between UNIX systems. People began exchanging messages by copying files between machines. The idea of using UUCP for exchanging messages eventually evolved into a protocol called NNTP---Network News Transfer Protocol---, which is the protocol used by \lp. (The protocol used to be called Netnews Transfer Protocol.) Since e-mail was already daily practice of members of the Internet back then, many things from e-mail were taken by the NNTP designers. So, an NNTP message looks a lot like an e-mail message and the two---NNTP and SMTP (the protocol used by e-mail)---can often mingle seamlessly. The impression we get from using NNTP is that we're sending e-mail to a certain group of people. It's as though the message goes into a collective mailbox and anyone interested in that mailbox reads the messages there. If anyone would like to reply to a message, they do so and, this way, communication flows among the interested crowd. If anyone would like to leave the group, nothing is needed---the person just doesn't go back to read any more messages. Unlike mailing lists, there is no need to formally commit to reading one of these collective mailboxes and no need to formally notify anyone or any system that you're not interested in that group any longer. These collective mailboxes are called ``news groups'' and are often written as ``newsgroups''. And the messages posted to these newsgroups are called ``articles''. Just like e-mail and the web, network news is an open protocol. Anyone could write a program capable of speaking NNTP. There are many NNTP-aware programs. You could write your own. Figures \ref{fg:gnus}--\ref{fg:sylpheed} show a few programs for reading network news via NNTP. \section*{Principles for a discussion group}\label{principles} \pdfbookmark[1]{Principles for a discussion group}{principles} We believe a discussion group should be small and grow slowly. By ``slowly'' we mean that each member comes in through an invitation. This way, the group being closed by definition, we keep spam out and give members a certain sense of privilege. A discussion group should be formed by interested people. If a participant doesn't log-in for a certain period of time, \lp\ locks the participant's account---see Section \ref{sec:inactive-users}. The account can be reactivated, but it will take asking another participant (with an active account) to do so. In other words, there's an encouragement for an uninterested member not to come back to the loop. The idea is to keep a certain cohesion in the discussion groups. When an account is locked or unlocked, an article is posted to the group {\tt local.control.news}, so everyone knows who is leaving and arriving. This way, participants get to have an idea of who is reading, who is watching them. Each invitation comes with a certain responsibility: it's possible to see who invited who. If {\tt BOB} misbehaves, everyone gets to see that {\tt ALICE} doesn't have nice friends. The {\tt USERS} command shows the relationship graph: % \begin{verbatim} USERS 200 List of current users: ANONYMOUS, last seen on Fri Mar 8 19:01:56 2024, invited (ALICE) ALICE, last seen on Sun Mar 10 21:25:45 2024, invited (BOB CARLA) BOB, last seen on Sun Mar 10 21:17:30 2024, invited nobody CARLA, last seen on Sun Mar 10 18:30:48 2024, invited nobody \end{verbatim} We conjecture that a discussion group tends to prosper when each member feels as important as any other member. So we think each member should have as much managerial power as any other. In an attempt to realize this ideal, each member is able to not only invite other people---see the [[CREATE-ACCOUNT]] command---, but also to create new groups---see the [[CREATE-GROUP]] command. Despite this introduction, this is not a tutorial on the history of the Internet or how NNTP works. This is the source code of \lp. Hereafter, our conversation continues in Lisp. Understanding how \lp\ is made is only necessary if you intend to modify it. If you just want to use the system, you probably should stop right here. \section*{How to install} See \href{https://git.antartida.xyz/loop/srv/raw/branch/main/README}{[[README]]} in \lp's source code. \section{Implementation strategy}\label{sec:design} Anything a user sends to the \lp\ is wrapped in a [[request]] and any command processing must produce a [[response]]: % \begin{verbatim} request ---> process ---> response. \end{verbatim} % The request arrives, \lp\ interprets it, finds out which command should look at the request and then dispatches the the request to the right command. The command process the request and returns a response. Then \lp\ takes this response and sends it out back to the user through the network using [[send-response!]]. That's all there is to \lp. Everything else is just the details of this strategy. Before you investigate the source code of \lp, you should have {\tt SBCL} installed---the Common Lisp compiler we've been using---and the package [[:loop]]. This way you can interact with the system: % \begin{verbatim} CL-USER> (ql:quickload :loop) To load "loop": Load 1 ASDF system: loop ; Loading "loop" .................. (:LOOP) CL-USER> (in-package :loop) # LOOP> (make-request "HELP") #S(REQUEST :VERB "HELP" :ARGS NIL :SAID NIL) \end{verbatim} % The procedure [[make-request]] constructs a [[request]]. When a user connects to the \lp, each command sent by the user is packaged inside a [[request]] structure like this one above. To see how \lp\ answers this request for help, we can invoke the procedure [[dispatch]], which is responsible to find the correct procedure to reply to the request. \begin{verbatim} LOOP> (dispatch (make-request :verb "HELP")) #S(RESPONSE :CODE 400 :DATA "unrecognized command" :REQUEST #S(REQUEST :VERB "HELP" :ARGS NIL :SAID NIL) :MULTI-LINE NIL) \end{verbatim} It turns that \lp\ doesn't recognize the command. That's because we only loaded the [[:loop]] package. The reason it doesn't recognize the command is that the command table has not been built. Let's build it and try again: \begin{verbatim} LOOP> (set-up-tables!) (("ARTICLE" . #S(COMMAND :FN # :VERB "ARTICLE" :DESCRIPTION "fetches full articles")) [...] ("HELP" . #S(COMMAND :FN # :VERB "HELP" :DESCRIPTION "displays this menu")) [...]) LOOP> (dispatch (make-request :verb "HELP")) #S(RESPONSE :CODE 200 :DATA (87 104 97 [...] 112) :REQUEST #S(REQUEST :VERB "HELP" :ARGS NIL :SAID NIL) :MULTI-LINE YES) \end{verbatim} That's better. These numbers we see in the response are the bytes in the response. You can get a string version of these numbers: \begin{verbatim} LOOP> (bytes->string (response-data (dispatch (make-request :verb "HELP")))) "What's on the menu today? ARTICLE fetches full articles AUTHINFO makes me trust you BODY fetches an article body GROUP sets the current group HEAD fetches article headers HELP displays this menu LIST lists all groups MODE handles the mode request from clients NEXT increments the article pointer POST posts your article QUIT politely says good-bye XDD displays your state of affairs XOVER fetches the overview database of a group" \end{verbatim} This is the text that the user sees when they ask for {\tt HELP}. In other words, the field {\tt data} in a [[response]] stores the data to be delivered back to the user. The program {\tt nc}---for ``netcat''---that we use below is capable of opening a TCP connection and handling that connection to our keyboard so that we can interact with the \lp. You can effectively achieve the same thing using a program such as {\tt telnet}. \begin{verbatim} C:\>nc antartida.xyz 119 200 Welcome! Say ``help'' for a menu. help 200 What's on the menu today? ARTICLE fetches full articles AUTHINFO makes me trust you BODY fetches an article body GROUP sets the current group HEAD fetches article headers HELP displays this menu LIST lists all groups MODE handles the mode request from clients NEXT increments the article pointer POST posts your article QUIT politely says good-bye XDD displays your state of affairs XOVER fetches the overview database of a group . quit 205 Good-bye. \end{verbatim} Despite \lp\ being made to talk to computer programs such as the ones illustrated in Figures~\ref{fg:gnus}--\ref{fg:sylpheed}, it's perfectly possible for a user to interact with \lp\ directly using a keyboard and a command-line tool such as {\tt nc} or {\tt telnet}. In fact, \lp\ takes advantage of that to be {\em hackable}---see also Section~\ref{sec:repl}. Commands such as [[CREATE-ACCOUNT]], [[CREATE-GROUP]], [[PASSWD]] are not part of the NNTP protocol, so users need to know how to use {\tt nc} or {\tt telnet} to take advantage of all of \lp's capabilities. \section{NNTP protocol} An Internet protocol is usually defined by a document whose tradition calls RFC---for ``request for comments''---, an acronym suggested by Stephen D.~Crocker, who wrote RFC 1 back in 1969, when ARPANET connected four computers. NNTP is defined by RFCs 977, 2980, 3977, 4643 and 5536. RFC 977 was the first and replaced by 3977. Still, reading RFC 977 is interesting precisely because it gives us a historical account of the protocol, making it easier to understand the evolution of the system. The objective of RFC 2980 was to implement new ideas to the NNTP protocol---to extend the protocol. RFC 3977 adopts some of these extensions. RFC 4643 also extends RFC 2980---addressing concerns with authentication. \section{It's a network server} The \lp\ program is a network server, that is, it serves connections on the network. However, instead of implementing the handling of TCP connections, we use what is called a ``TCP superserver'' or ``TCP wrappers'' or perhaps just ``TCP server''. The idea is---a program called ``TCP server'' waits for connections from the network. When a client arrives, the TCP server handles the network connection to \lp. From \lp's perspective, the client is just another keyboard talking to it directly. This strategy simplifies the implementation of \lp. Both programs---the TCP server and \lp---have nothing to do with one another, but their work together makes the service. This implementation strategy is typically found in UNIX programs. % \begin{quote}\small This is the Unix philosophy. Write programs that do one thing and do it well. Write programs to work together. Write programs that handle text streams, because that is a universal interface.\\ --- Doug McIlroy, 1989, interviewed by Michael S.~Mahoney. \end{quote} % The TCP server just just one thing---listens for new connections and handles them to the interested program. It does the handling and does it well. \Lp, on the other hand, concerns itself with the NNTP protocol and does not worry about handling network connections. This way, they work together. And \lp\ handles only a text stream, which is why it's so easy to connect a keyboard to it and interact with it through the command line as illustrated in Section~\ref{sec:design}. \section{REPL for \lp}\label{sec:repl} There's a REPL to \lp's procedures---always useful. \begin{verbatim} %./loop --repl LOOP> (+ 1 1) 2 LOOP> (list-groups) (local.control.news local.test) \end{verbatim} The REPL can also be use in the middle of an NNTP conversation, which is useful for debugging. All authenticated users can use it, making passwords only a barrier to spammers, not to your circle of friends, since everyone can find a way to log in with anyone's account, change their passwords or just do whatever they want to \lxxp\ process or databases. <>= (defun repl (r) (in-package :loop) (loop (print/finish "LOOP> ") (handler-case (let ((expr (read))) (if (eq 'quit expr) (progn (return (make-response :code 200 :request r :data "Okay, that's enough hacking for today."))) (println "~a" (eval expr)))) (end-of-file () (print/finish "^D~%") (uiop:quit 0)) (interactive-interrupt () (print/finish "^C~%") (uiop:quit 0)) (t (c) (print/finish "Oops: ~a~%" (str:collapse-whitespaces (str:replace-all (string #\linefeed) " " (fmt "~a" c)))))))) @ %def repl \section{Representation of a client} How do we represent a client? A client is typically reading a group and an article; it's has authenticated itself or not yet. So we need a global structure to annonate the client's state. <>= (defstruct client group (article 1) (username "ANONYMOUS") (auth? 'no)) (defparameter *client* (make-client)) @ %def client *client* \section{Representation of a command} What does a client typically tell \lp? A client typically sends commands. Commands typically need arguments. Each command is dispatched to a procedure that answers it---it's the purpose of \verb|fn| in the command. Together, all commands make up a table of commands, which is essentially what the user sees when asks for \verb|HELP|. (Be aware that some clients use the output of \verb|HELP|. For example, Gnus v5.13.) <>= (defstruct command fn verb description) (defparameter *commands-assoc* nil) @ <>= (defun table-of-commands () `(("GROUP" ,#'cmd-group "sets the current group") ("NEXT" ,#'cmd-next "increments the article pointer") ("HELP" ,#'cmd-help "displays this menu") ("LIST" ,#'cmd-list "lists all groups") ("AUTHINFO" ,#'cmd-authinfo "makes me trust you") ("LOGIN" ,#'cmd-login "shorter interface to AUTHINFO") ("HEAD" ,#'cmd-head "fetches article headers") ("MODE" ,#'cmd-mode "handles the mode request from clients") ("BODY" ,#'cmd-body "fetches an article body") ("POST" ,#'cmd-post "posts your article") ("ARTICLE" ,#'cmd-article "fetches full articles") ("XOVER" ,#'cmd-xover "fetches the overview database of a group") ("CREATE-GROUP" ,#'cmd-create-group "creates a new group so you can discuss your favorite topic") ("CREATE-ACCOUNT",#'cmd-create-account "creates an account so you can invite a friend") ("PASSWD" ,#'cmd-passwd "changes your password") ("USERS" ,#'cmd-list-users "lists all users") ("DD" ,#'cmd-dd "[d]isplays [d]ata: your state of affairs") ("QUIT" ,#'cmd-quit "politely says good-bye") ("DATE" ,#'cmd-date "displays the current date at this server") ("UNLOCK-ACCOUNT" ,#'cmd-unlock-account "unlocks an account") ("REPL" ,#'cmd-repl "lets you hack away"))) (defun set-up-tables! () (labels ((build-commands-assoc (ls) (if (null ls) nil (cons (apply #'make-command-pair (car ls)) (build-commands-assoc (cdr ls))))) (make-command-pair (name fn desc) (cons name (make-command :fn fn :verb name :description desc)))) (setf *commands-assoc* (sort (build-commands-assoc (table-of-commands)) #'string-lessp :key #'car)))) (defun get-command (key) (let ((cmd (assoc key *commands-assoc* :test #'string=))) (labels ((unrecognized-command () (make-command :fn #'(lambda (r) (make-response :code 400 :data "unrecognized command" :request r)) :verb 'unrecognized :description "a command for all commands typed wrongly"))) (or (cdr cmd) (unrecognized-command))))) @ %def *commands-assoc* set-up-tables! get-command \section{Representation of requests and responses} Each command is given through a text line written by the user. Let's call this text line the [[request]]. When \lp\ parses the request, it will extract (from the request) a verb and some arguments. We will take a verbatim copy of everything the user has said, possibly for debugging purposes. How do we represent a [[response]]? A [[response]] is always a reaction to a [[request]]. The NNTP protocol always specifies an integer as the code to a response, which is what we call the \verb|code| in the response. Long responses end with a period and we mark such responses with [[multi-line]]. <>= (defstruct request verb args said) (defstruct response code data request multi-line) @ <>= (defun empty-response () (make-response :code 400 :data "I beg your pardon?")) (defun prepend-response-with (message r) (make-response :code (response-code r) :data (list message (crlf) (response-data r)) :multi-line (response-multi-line r) :request (response-request r))) @ \section{Main loop} Every command consumes a [[request]] and produces a [[response]]. What does \lp\ do as its main loop? It repetitively reads a line from the user, processes that line and always replies something back. Then \lp\ is back at waiting for the user to say something else again. If the user says {\tt QUIT}, then we should identify it and terminate \lp's execution. That's why [[send-response!]] returns the [[request]] itself---so we can cascade actions based on a user's request. <
>= (defun main-loop () (handler-case (loop (let* ((bs (nntp-read-line)) (ln (bytes->string bs))) (stderr "<<< ~a~%" ln) ;; for logging (let ((r (send-response! (dispatch-line ln)))) (when (response-quit? r) (return))))) (end-of-file (c) (declare (ignore c)) (stderr "a client disconnected with EOF~%")) (error (e) (stderr "fatal: ~A" e)))) @ %def main-loop <
>= (defun request-quit? (r) (and r (string= 'quit (request-verb r)))) (defun response-quit? (r) (and r (request-quit? (response-request r)))) (defun server-start () (set-up-tables!) (send-banner!) (main-loop)) (defun main () (let ((app (cli/command))) (clingon:run app))) (defun send-banner! () (send-response! (make-response :code 200 :data (welcome-message)))) (defun welcome-message () (cl-ppcre:regex-replace-all "HASH" "<>" *git-hash*)) @ %def main welcome-message \noindent It's always useful to know which version exactly we're dealing with: % \begin{verbatim} %./loop.exe 200 Welcome! I am LOOP 9575ac2. Say ``help'' for a menu. \end{verbatim} % So we put a release tag on \lp's banner. <>= Welcome! I am <> <> hash HASH. Say ``help'' for a menu. @ \noindent We take the opportunity and describe \lp's package, information which we also use in [[loop.asd]]. <>= LOOP @ <>= An NNTP server for a circle of friends. @ <>= 0.0.1 @ \section{Parsing of requests} The commands themselves we call {\tt verbs}; everything else the user types we call {\tt args}. Observe that upper and lower case letters are equivalent in request verbs. <>= (defun parse-request (r) (let* ((collapsed-s (str:collapse-whitespaces (request-said r))) (ls (str:split " " collapsed-s :omit-nulls 'please))) ;; What are we going to do with a null request? (cond ((null ls) (make-request :said (request-said r))) (t (let ((verb (car ls)) (args (cdr ls))) (make-request :said (request-said r) :verb (str:upcase verb) :args args)))))) @ %def parse-request \section{Parsing of command-line arguments} We're using the clingon library as per Vincent Dardel suggestion in ``The Common Lisp Cookbook''. We begin with writing a description of the program and options it understands. Notice I don't know how to support a two-argument option, so I hacked a solution away. What we need to do is to implement a new option. The library is extensible. However, I could be misremembering it here, but I recall having decided to give up on the clingon library because it does too much. I think it even handles {\tt SIGINT} on our backs. Nevertheless, do notice that the library is doing a pretty good for a start, so unless we understand exactly what it was that we didn't like, we'll keep it. <>= (defun cli/options () (list (clingon:make-option :string :description "creates a new account" :long-name "create-account" :key :create-account) (clingon:make-option :string :description "changes password" :long-name "change-passwd" :key :change-passwd) (clingon:make-option :flag :description "lists accounts" :short-name #\l :long-name "list-accounts" :key :list-accounts) (clingon:make-option :flag :description "runs a REPL right now" :short-name #\r :long-name "repl" :key :repl) (clingon:make-option :flag :description "disables the NNTP REPL" :long-name "disable-nntp-repl" :key :disable-nntp-repl) (clingon:make-option :flag :description "logging (on stderr)" :long-name "logging" :key :logging))) @ We implement first the procedures that handle options that represent an entire program. For example, saying [[--list-accounts]] is like running a program [[./list-accounts]]. <>= (defun cli/list-accounts () (println (str:join (crlf-string) (list-users)))) (defun cli/create-account (username args) (let ((invited-by (car args))) (cond ((null invited-by) (println "Must specify who invites the new account.")) ((get-account username) (println "Username account ``~a'' already exists." username)) ((not (get-account invited-by)) (println "Invited-by account ``~a'' doesn't exist." invited-by)) (t (multiple-value-bind (okay? pass-or-error) (new-account! username invited-by) (if okay? (progn (println "Okay, account ``~a'' created with password ``~a''." username pass-or-error) (notify-user-created username)) (println "Sorry, ~a." pass-or-error))))))) (defun cli/change-passwd (username args) (let* ((random-passwd (random-string 6)) (given-passwd (car args)) (new-passwd (or given-passwd random-passwd))) (if (not (get-account username)) (println "No such account ``~a''." username) (multiple-value-bind (okay? problem) (change-passwd! username new-passwd) (if okay? (println "Okay, account ~a now has password ``~a''." username new-passwd) (println "Sorry, could not change password: ~a." problem)))))) @ %def cli/change-passwd cli/list-accounts cli/create-account Now let's write the main procedure in command-line parsing. Notice that because of the design of the [[clingon]] library, command-line parsing becomes the main procedure of \lp. In other words, \lp's service starts with [[server-start]]. <>= (defun cli/main-with-handlers (cmd) (handler-case (cli/main cmd) (end-of-file () (print/finish "^D~%") (uiop:quit 0)) (interactive-interrupt () (print/finish "^C~%") (uiop:quit 0)))) (defun cli/main (cmd) (read-accounts!) (let ((args (clingon:command-arguments cmd)) (run-server t) (repl (clingon:getopt cmd :repl)) (ca (clingon:getopt cmd :create-account)) (pa (clingon:getopt cmd :change-passwd)) (la (clingon:getopt cmd :list-accounts)) (logging (clingon:getopt cmd :logging)) (disable-nntp-repl (clingon:getopt cmd :disable-nntp-repl))) (setf *debug* logging) (when (or ca pa la) (setf run-server nil)) (when la (cli/list-accounts)) (when ca (cli/create-account ca args)) (when pa (cli/change-passwd pa args)) (when repl (repl (make-request :verb "repl" :args '(command-line)))) (when disable-nntp-repl (setq *enable-nntp-repl* nil)) (when run-server (server-start)))) (defun cli/command () (clingon:make-command :name "loop" :description "<>" :version "<>" :license "GPL v3" :options (cli/options) :handler #'cli/main-with-handlers)) @ %def cli/options cli/command \section{Request dispatching mechanism} Dispatching requests means consuming one and invoking the correct procedure that will process the request. The invoked procedure must produce a [[response]]. The work of dispatching is just delivering the task to an response-producing procedure and then raising the response to whoever needs to catch it. For example, [[response-quit?]] is used by [[main-loop]] to identify when the user has issued {\tt QUIT}, in which case we terminate [[main-loop]]. <>= (defun dispatch (r) (let* ((verb (request-verb r))) (if (null verb) (empty-response) (funcall (command-fn (get-command verb)) r)))) (defun dispatch-line (ln) (dispatch (parse-request (make-request :said ln)))) @ %def dispatch dispatch-line \section{Representation and parsing of articles} An article is made of two parts, the head and the body. We do need to parse the head, but we never parse the body: we don't want to intervene in anything that a user might be doing in the body of the article; in fact, we don't need read the body in memory because it might be large. When we need to send an article to a client, we then read it chunk by chunk and stream it out with constant memory use. The headers, however, are mostly under the jurisdiction of the server. This assumption is justified by the fact that servers must understand headers. For example, how do we know to which groups an article was posted? We look at the header {\tt newsgroups}. The assumption we make here is that headers are always one-byte encoded, which is not the case of ASCII, but it is the case of ISO-8859-1. The procedure [[code-char]] realizes this ideal, so we're out of trouble here. So, in the structure below, [[article-head]] is a string, while [[article-body]] is a vector of bytes. To get a list of pairs out of the set of all headers of an article, we can ask [[parse-headers]]. <>= (defstruct article head body) @ %def article-head article-body \section{How to extract articles from the database} We don't extract the entire article from the file system. We first extract just the headers. If it's ever needed, we read the article chunk by chunk and hand them to a client that asked for it. Let's focus now on reading the headers. First, we need to read the bytes off of the file system into memory. What if the article is huge? We can't read it all. What we do is to read just enough of the article until we find all of the head. That's the job of [[read-headers-raw]]. The method below is to assume every head fits in a certain fixed size. We'll be right nearly 100\% of the time. If we were to find a very long head, we'll then allocate more and more and crash, which is fine because a \lp\ process handles just a single client. Ideally, we'd stop at a certain limit and give up on the article, which is easy to do: throw an exception if the size of the array goes beyond the limit. We work with two buffers, the accumulator [[acc]] and the temporary [[buffer]]. We fill in [[buffer]]. Now we memorize the length of [[acc]] before we possibly modify its size by copying-extending it. Now we know the position of where the new chunk of the head has been allocated. It's time to search just the newly-read chunk for \verb|\r\n\r\n|. A [[search]] call should do. However, what happens if a chunk read from the stream ended right in the middle of the separator? We would be appending the tail of the separator right now, so we must go back a little bit before we search so that we can find the separator even in cases when we had to read it in two calls to [[read-sequence]]. (Such is life in programming.) That's why we subtract 3 from [[start-search]]. If we don't find the separator still, that's because the head is really long, so we repeat. When we found it, we know that the head starts at zero and goes up to the position of the separator plus the four bytes of the separator itself. Why do we include the four bytes in the head? That makes it easy later one to know where we would find the body---exactly after where the head ends. \section{How to parse articles} Let's now describe how we implement the parsing of articles. The input is a vector of bytes, bytes that represent the article written on the file system. First, search for the \verb|\r\n\r\n|, which is bytes \verb|#(13 10 13 10)|, that separates head and body. The headers themselves can be turned into a string. For the body, we're going to use a great feature of Common Lisp---a zero-copy array window into the original array [[v]], the argument of the procedure. The feature is requested by the [[:displaced-to]] keyword, which specifies the original array, where our window begins is specified by the long offset keyword. The length of our window has to be the entire article body, which is the length of the vector minus the head and the separator \verb|\r\n\r\n|, which is 4 bytes long. An article that would have no separator and no body would not be an article. But still, I'd rather be very careful here and create an empty body if no separator could ever be found. While \Lp would never create such an article, it could be tomorrow that we might parse foreign articles for whatever reason. <>= (defun parse-article (v) (let ((pos (search #(13 10 13 10) v))) (if pos (make-article :head (map 'string #'code-char (subseq v 0 pos)) :body (make-array (- (length v) (+ pos 4)) :element-type (array-element-type v) :displaced-to v :displaced-index-offset (+ pos 4))) (make-article :head (map 'string #'code-char v) :body (make-array 0 :element-type (array-element-type v)))))) (defun hs-space-collapsed (hs) (cl-ppcre:regex-replace-all (fmt "~a[ ~a]+" (crlf-string) #\tab) hs " ")) (defun hs-lines (lines) (str:split (crlf-string) lines)) (defun parse-header (header) (let* ((h (str:collapse-whitespaces header)) (pos (search ":" h))) (when (null pos) (throw 'article-syntax-error (values nil (fmt "missing colon in header |~a|" h)))) (when (<= (length h) (+ 2 pos)) (throw 'article-syntax-error (values nil (fmt "empty header ~a" h)))) (multiple-value-bind (key val) (values (subseq h 0 pos) (subseq h (+ 2 pos))) (cons (str:downcase key) val)))) (defun parse-headers (hs) (let ((ls (hs-lines (hs-space-collapsed hs)))) (mapcar #'(lambda (h) (parse-header h)) ls))) (defun string-integer? (s) (ignore-errors (parse-integer s))) @ %def parse-article parse-headers We now write some procedures that we use when we're build the {\em overview} of the command \verb|XOVER|. We write them here because they're still part of parsing an article. <>= (defun get-header (key hs) (let ((pair (assoc key hs :test #'string=))) (if pair (cdr pair) ""))) (defun fetch-headers (g i) (let* ((a-string (fetch-article g i)) (a-parsed (parse-article a-string)) (headers (parse-headers (article-head a-parsed)))) (enrich-headers headers a-string a-parsed))) (defun enrich-headers (hs a-raw a-parsed) (append hs `(("line-count" . ,(fmt "~a" (nlines (article-body a-parsed)))) ("byte-count" . ,(fmt "~a" (length a-raw)))))) (defun nlines (v) (loop for b across v count (= b 10))) @ %def get-header fetch-headers The procedures [[extract-mid]] and [[lookup]] also belong belong in this section. Do they really? It seems so. <>= (defun extract-mid (a) (lookup "message-id" (parse-headers (article-head (parse-article a))))) (defun lookup (key table) (cdr (assoc key table :test #'string=))) @ %def extract-mid lookup \section{How to stream off articles through the network} We consider the body of articles to be the poster's property, so we don't want to touch it in any way. For example, we never treat it as a string because we would have to guess an encoding. It's the article viewer's---the NNTP client's---responsibility of interpreting such bytes. For us, articles are made of a string head and a sequence-of-bytes body. Of great relevance is the fact that some articles could be huge, so we don't read them wholly into memory. When it's necessary to send an entire article to a client, we just stream the bytes to the client chunk by chunk, giving us a constant-memory guarantee. We describe now how we do this. You can see in [[parse-article]] that we use a {\em displaced array}. A displaced array is a great feature of Common Lisp in which a new array is made but references an already allocated memory. How are articles handled by LOOP? Begin at [[fetch-article]], which depends on [[read-headers-raw]], a procedure that reads only so much of an article so as to consume the entirety of headers, using [[vector-push-extend]]. So the way only to attack LOOP is to send it an article whose headers are huge, implying we must limit article's head size, though we don't currently do that. As soon as headers are consumed, [[read-headers-raw]] finishes. Of greater relevance here is [[typical-cmd-response]] because it is the one that actually streams off bytes to a client. How does it do it? It first reads the headers and then it marks off where the body begins. How does it mark the body? By using the structure called [[file-stream-payload]], our choice of how to represent an article body. The structure records the path of the file and where in the streaming it is; the {\em where} is the current byte to be written to the network, discovered by Lisp's [[file-position]], the portable version of [[POSIX fseek]]. Where does a body begin? An article is made of a head, a separator and a body. So the body begins at the separator ends. You can verify this information by reading [[read-headers-raw]]; notice how it returns the head by adding 4 bytes after the position of where the separator was found---in other words, the head includes the separator, so the body begins right where [[read-headers-raw]] ends. The product of [[typical-cmd-response]] is a response. When an entire article has been requested---which is what happens when someone issues a [[BODY]] or an [[ARTICLE]] command---, [[send-response!]] will need to read the body's bytes. If you now look at [[send-response!]] you can see that it delegates the actual work to [[send-chunk!]], which, in turn, delegates it to [[stream-file-to-socket]], which does the job. How does it do the job? It knows exactly where the body begins. It then has only the job of reading a chunk (of fixed sized) of that file at where the body begins. Then it writes the entire file to the client's network connection. How does it do it? While it's able to get bytes from [[read-sequence]], it writes them one by one to the network socket. Why one one by one? To make sure we can do {\em dot-stuffing}, otherwise we could've delegated that job to [[write-sequence]]. Can we do a better job with respect to the source code writing? Not really. Procedure [[send-response!]] needs to do different things based on the type of content. This desired type-dispatching is performed by [[send-chunk!]] and then we need to write a particular procedure for handling the file-stream-payload representation. {\bf Dot stuffing}. Should dot stuffing be done on the fly? The alternative would be when storing the article. But, no, we should store the article precisely as the user has sent it. It is to be treated as an important historical document. We don't touch user property, which is the article's body. <>= (defstruct file-stream-payload (path "" :type string) (start-pos 0 :type integer)) (defun fetch-article (g i) (in-group g (read-headers-raw (fmt "~a" i)))) (defun read-headers-raw (path) (with-open-file (in path :element-type '(unsigned-byte 8) :if-does-not-exist :error) (when in (let ((acc (make-array 8192 :element-type '(unsigned-byte 8) :adjustable t :fill-pointer 0)) (buffer (make-array 8192 :element-type '(unsigned-byte 8))) (sep #(13 10 13 10))) (loop (let ((n (read-sequence buffer in))) (when (zerop n) (return acc)) ; eof without having found the separator (let ((start-search (length acc))) (dotimes (i n) ; copy all bytes to the accumulator (vector-push-extend (aref buffer i) acc)) ;; Search for the separator. We subtract 3 from our ;; start index to catch a separator that was broken into ;; two calls of read-sequence. (let ((pos (search sep acc :start2 (max 0 (- start-search 3))))) (when pos (return (subseq acc 0 (+ pos 4)))))))))))) (defun stream-file-to-socket (payload out-stream) "Opens the file, seeks to the start of the body, and streams to the socket while applying on-the-fly NNTP dot-stuffing." (with-open-file (in (file-stream-payload-path payload) :element-type '(unsigned-byte 8)) (file-position in (file-stream-payload-start-pos payload)) (let ((buf (make-array 8192 :element-type '(unsigned-byte 8))) (at-line-start t)) ; The body implicitly starts at the beginning of a line (loop for n = (read-sequence buf in) while (plusp n) do (loop for i from 0 below n for byte = (aref buf i) do (when (and at-line-start (= byte 46)) ; 46 is '.' (write-byte 46 out-stream)) (write-byte byte out-stream) (setf at-line-start (= byte 10))))))) (defun send-chunk! (chunk s) (etypecase chunk (string (write-sequence (string->bytes chunk) s)) (vector (write-sequence chunk s)) (list (dolist (c chunk) (send-chunk! c s))) (integer (write-byte chunk s)) (file-stream-payload (stream-file-to-socket chunk s)))) (defun send-response! (r) (send-chunk! (integer->string (response-code r)) *standard-output*) (send-chunk! " " *standard-output*) (send-chunk! (response-data r) *standard-output*) (send-chunk! (crlf) *standard-output*) (when (response-multi-line r) (send-chunk! "." *standard-output*) (send-chunk! (crlf) *standard-output*)) (force-output) r) @ %def read-headers-raw fetch-article file-stream-payload \section{Commands} \subsection{{\tt HELP}} When someone asks for help, we present a table of commands. <>= (defun cmd-help (r) (labels ((menu (ls) (loop for item in ls collect (display-fn item))) (display-fn (cmd-pair) (let ((cmd (cdr cmd-pair))) (fmt "~A ~A" (command-verb cmd) (command-description cmd))))) (let ((lines (menu *commands-assoc*))) (prepend-response-with "What's on the menu today?" (make-response :code 200 :multi-line 'yes :request r :data (str:join (crlf-string) lines)))))) @ \subsection{{\tt AUTHINFO}}\label{sec:authinfo} The implementation of {\tt AUTHINFO}. When we connect to \lp\ directly from a keyboard, it's a bit painful to authenticate with two commands---{\tt AUTHINFO user} and {\tt AUTHINFO pass}. So we also implemented the command {\tt LOGIN}---see Section~\ref{sec:login}. To check the user's password, we use the procedure [[pass?]] that's defined in the implementation of {\tt PASSWD}. <>= Syntax error. Say ``authinfo USER /your-name/'' then ``authinfo PASS /your-pass/''. @ <>= (defun cmd-authinfo (r) (let* ((args (mapcar #'str:upcase (request-args r)))) (cond ((not (= (length args) 2)) (bad-input r "No, no: I take exactly two arguments.")) (t (multiple-value-bind (cmd arg) (apply #'values args) (cond ((string= cmd "USER") (setf (client-username *client*) arg) (make-response :code 381 :request r :data (fmt "Hey, ~a, please tell us your password." arg))) ((string= cmd "PASS") (if (authinfo-check (client-username *client*) arg) (progn (log-user-in!) (make-response :code 281 :request r :data (fmt "Welcome, ~a." (client-username *client*)))) (make-response :code 400 :request r :data "Sorry. Wrong password."))) (t (make-response :code 400 :request r :data "<>")))))))) (defun authinfo-check (username passwd) (pass? username passwd)) (defun auth? () (eq 'yes (client-auth? *client*))) (defun log-user-in! () (setf (client-auth? *client*) 'yes) (let ((u (get-account (client-username *client*)))) (setf (account-seen u) (get-universal-time))) (write-accounts!)) @ %def auth? log-user-in! \subsection{{\tt CREATE-ACCOUNT}} We allow authenticated members to invite their friends, which creates a tree of people. ({\em An idea}. We could envision that each tree trunk manages the rest of the tree underneath it. So I invite you, I could change your password, say, or handle any problems you might have. This decentralizes system administration, easing the support burden.) The name of each user must conform to the expression <
>= ^[^\\s]+$ @ Same as in @<>. We'll let users create whatever complicated user names they want. If they can type it up, it's their problem. <>= (defun user-name-conforms? (u) (conforms-to? u "<>")) (defun cmd-create-account (r) (with-auth (with-n-args 1 r (let* ((args (mapcar #'str:upcase (request-args r))) (username (car args))) (multiple-value-bind (username pass-or-error) (new-account! username (client-username *client*)) (if (not username) (make-response :code 400 :request r :data (fmt "~a. Choose a new name." pass-or-error)) (progn (notify-user-created username) (make-response :code 200 :request r :data (fmt "Okay, account ~a created with password ``~a''." username pass-or-error))))))))) @ <>= (defstruct account username seen last-post friends pass pass-locked pass-locked-why creation) (defparameter *accounts* nil) @ %def *accounts* <>= (defun read-accounts! () (let ((*package* (find-package '#:loop))) (with-open-file (s "accounts.lisp" :direction :input) (setq *accounts* (read s)))) *accounts*) (defun string->array (s) (make-array (length s) :element-type '(unsigned-byte 8) :initial-contents (map 'vector #'char-code s))) (defun string->sha256 (s) (let ((d (ironclad:make-digest :sha256))) (ironclad:produce-digest (ironclad:update-digest d (string->array s))))) (defun new-account! (username invited-by) (let* ((u (str:upcase username)) (p (random-string 6)) (a (make-account :username u :pass (string->sha256 (str:upcase p)) :creation (get-universal-time)))) (multiple-value-bind (okay? reason) (user-name-conforms? u) (declare (ignore reason)) (cond ((not okay?) (values nil (fmt "username must conform to <>"))) ((get-account u) (values nil (fmt "account ~a already exists" u))) (t (push u (account-friends (get-account invited-by))) (push a *accounts*) (write-accounts!) (values (str:upcase username) p)))))) @ %def CREATE-ACCOUNT new-account! To write [[write-accounts!]] we will need locking, no way to avoid this given the parallel daily-nature of \Lp. Two processes in parallel may ask for the writing of [[accounts.lisp]]. The process that loses the race would have its modifications lost. What do we need to do? We need locking no matter what. <>= (defmacro with-file-lock ((lock-path &key (sleep-time 0.1) (retries 50)) &body body) (let ((lock-file (gensym "LOCK-FILE-")) (stream (gensym "STREAM-")) (attempt (gensym "ATTEMPT-"))) `(let ((,lock-file ,lock-path)) (unwind-protect (progn (loop for ,attempt from 1 to ,retries do (handler-case (with-open-file (,stream ,lock-file :direction :output :if-exists :error :if-does-not-exist :create) (format ,stream "~a~%" (get-universal-time)) (return)) (file-error () (if (= ,attempt ,retries) (error "could not acquire lock on ~a after ~a attempts." ,lock-file ,retries) (sleep ,sleep-time))))) ,@body) (ignore-errors (delete-file ,lock-file)))))) (defun write-accounts! () (with-file-lock ("accounts.lck") (let ((name (loop (let* ((tmp (random-string 10)) (name (fmt "~a.tmp" tmp))) (when (ignore-errors (with-open-file (s name :direction :output :if-exists :error :if-does-not-exist :create) (write *accounts* :stream s) t)) (return name)))))) (if (ignore-errors (rename-file name "accounts.lisp")) (values t *accounts*) (values nil (fmt "could not rename ~a to accounts.lisp" name)))))) (defun get-account (username) (loop for u in *accounts* do (when (string= (str:upcase username) (account-username u)) (return u)))) @ %def read-accounts! write-accounts! get-account \subsection{{\tt UNLOCK-ACCOUNT}} Inactive accounts are removed or locked---see Section \ref{sec:inactive-users}. When an account is locked, any member can unlock it. <>= (defun cmd-unlock-account (r) (with-auth (with-n-args 1 r (let* ((args (mapcar #'str:upcase (request-args r))) (username (car args))) (cond ((not (get-account username)) (make-response :code 400 :request r :data (fmt "No such account ~a." username))) ((not (locked? username)) (make-response :code 400 :request r :data (fmt "Can't unlock ~a because it's not locked." username))) (t (unlock-account! username) (notify-user-unlocked username) (make-response :code 200 :request r :data (fmt "Okay, account ~a unlocked." username)))))))) (defun unlock-account! (username) (let ((u (get-account username))) (cond ((not u) (values nil "no such account")) ((not (locked? username)) (values nil "account isn't locked")) (t (setf (account-pass u) (account-pass-locked u)) (setf (account-pass-locked u) nil) (setf (account-pass-locked-why u) nil))))) @ %def unlock-account! \subsection{{\tt LOGIN}}\label{sec:login} Besides {\tt AUTHINFO} in Section~\ref{sec:authinfo}, we also implement a more convenient command for authentication when we interact with \lp\ through a command-line interface. Instead of having to say two commands, we can just say {\tt login user password}. <>= (defun cmd-login (r) (let* ((args (mapcar #'str:upcase (request-args r)))) (cond ((not (= (length args) 2)) (bad-input r "Usage: login your-username your-password")) (t (multiple-value-bind (name pass) (apply #'values args) (cond ((pass? name pass) (log-user-in-as! name) (make-response :code 200 :request r :data (fmt "Welcome, ~a." name))) (t (make-response :code 400 :request r :data (fmt "Wrong password."))))))))) (defun log-user-in-as! (name) (setf (client-username *client*) name) (log-user-in!)) @ \subsection{{\tt PASSWD}}\label{sec:passwd} A change of password is made with {\tt PASSWD current new}. Observe that we are duplicating code from other command procedures. I think there is a macro emerging here called [[with-upcase-args]]. %% TODO <>= (defun cmd-passwd (r) (with-auth (let* ((args (mapcar #'str:upcase (request-args r)))) (cond ((not (= (length args) 2)) (bad-input r "Usage: passwd current-password new-password")) (t (multiple-value-bind (cur new) (apply #'values args) (cond ((pass? (client-username *client*) cur) (multiple-value-bind (okay? problem) (change-passwd! (client-username *client*) new) (if okay? (make-response :code 200 :request r :data "You got it. Password changed.") (make-response :code 500 :request r :data (fmt "Sorry: ~a" problem))))) (t (make-response :code 400 :request r :data (fmt "Sorry. Wrong password.")))))))))) (defun pass? (username pass) (let ((u (get-account username))) (and u (cond ((null (account-pass u)) nil) ((integerp (account-pass u)) (eq (sxhash pass) (account-pass u))) ((arrayp (account-pass u)) (equalp (string->sha256 pass) (account-pass u))) (t (error "pass? expected to find INTEGERP or ARRAYP but found ~a" (type-of (account-pass u)))))))) (defun change-passwd! (username newpass) (let ((u (get-account username))) (when (not u) (error "I could not find account ~a." username)) (setf (account-pass u) (string->sha256 (str:upcase newpass))) (write-accounts!))) @ %def PASSWD pass? change-passwd! \subsection{{\tt USERS}}\label{sec:users} The tree of users and their friends is public. Anyone can know who invited who. <>= (defun cmd-list-users (r) (with-auth (prepend-response-with "List of current users:" (make-response :code 200 :request r :multi-line 'yes :data (str:join (crlf-string) (list-users)))))) (defun size-of-longest-username () (loop for u in *accounts* maximizing (length (account-username u)))) (defun list-users () (mapcar #'(lambda (row) (cadr row)) (sort (loop for u in *accounts* collect (list (account-username u) (fmt "~v@a~a, ~a, invited ~a" (size-of-longest-username) (account-username u) (if (locked? (account-username u)) (fmt " (account locked: ~a)" (account-pass-locked-why u)) "") (if (last-time-seen (account-username u)) (fmt "last seen on ~a" (last-time-seen (account-username u))) "never logged in") (or (account-friends u) "nobody")))) #'string<= :key #'(lambda (row) (car row))))) (defun universal-to-human (s) (format-timestring nil (universal-to-timestamp s) :format +asctime-format+)) (defun last-time-seen (username) (let ((u (get-account username))) (if u (let ((s (account-seen u))) (if s (universal-to-human s)))))) @ %def list-users last-time-seen size-of-longest-username \subsection{{\tt LIST}}\label{sec:list} The database of groups and articles is a UNIX directory. We just need to discover which directories exist and produce a listing. The heavy work here is finding the index interval of articles in the group. (I think we should already be optimizing this by merely caching the information in a file that is read at start-up. I think we should even cache the overview of the group.) %% TODO <>= (defstruct group name high low) (defun cmd-list (r) (prepend-response-with "Get in the loop! Lots to choose from." (make-response :code 215 :multi-line 'yes :data (str:join (crlf-string) (build-groups-lines (build-groups-structs))) :request r))) (defun build-groups-lines (ls) (reverse (mapcar #'(lambda (g) (fmt "~a ~a ~a y" (group-name g) (group-high g) (group-low g))) ls))) (defun build-groups-structs () (let ((ret-ls nil)) (dolist (g (list-groups) ret-ls) (multiple-value-bind (low high len) (group-high-low g) (declare (ignore len)) (setf ret-ls (cons (make-group :name g :high high :low low) ret-ls)))))) (defun between? (x from to) (<= from x to)) (declaim (inline between?)) (defun filesize (path) (sb-posix:stat-size (sb-posix:stat path))) (defun zero-file? (path) (= (filesize path) 0)) (defun temporary-article? (path) (or (zero-file? path) (cl-ppcre:scan "\.tmp$" (namestring path)))) (defun article-ready? (path) (not (temporary-article? path))) <> (defun get-articles (g &optional from to) (in-groups ;; We might want to optimize this some day. That's a ;; problem to be studied. (let ((as (articles->integers (remove-if #'temporary-article? (loop-list-files (truename g)))))) (sort (remove-if-not #'(lambda (x) (between? x (or from x) (or to x))) as) #'<)))) (defun group-high-low (g) (let* ((articles (get-articles g)) (sorted-ints (sort articles #'<))) (values (or (car sorted-ints) 0) (or (car (last sorted-ints)) 0) (length sorted-ints)))) (defun articles->integers (ls) (remove-if #'null (mapcar #'(lambda (g) (ignore-errors (parse-integer (basename (uiop:unix-namestring g))))) ls))) (defun list-groups () (let ((groups (in-groups (loop-list-directories (truename "."))))) (sort (mapcar #'(lambda (g) (basename (uiop:unix-namestring g))) groups) #'string-lessp))) (defun last-char (s) (char s (1- (length s)))) @ In [[basename]], what we want to do is---if the path ends with a slash, we invoke [[pathname]]. Otherwise, we invoke {\tt file-namestring}. <>= (defun basename (path) (let ((s (str:collapse-whitespaces path))) (if (char= #\/ (last-char s)) (car (last (pathname-directory s))) (file-namestring s)))) @ %def get-articles group-high-low To list directories and files, I have been using \href{https://github.com/Shinmera/filesystem-utils/tree/master}{{\tt filesystem-utils}} by Yukari Hafner. I found an issue with both [[list-directories]] and [[list-files]] in a fresh install of FreeBSD 14.2 and in a Debian 8.11 codename jessie. The issue is that the [[#+cffi]] chunk of the source code incorrectly produced [[NIL]]. (Dramatically, the same was not true in a FreeBSD 14.1.) The source code had an alternative chunk of code for [[#-cffi]] and I discovered that this alternative worked on these systems I tested. So, as a workaround, I incorporate these procedures below using the chunk [[#-cffi]] to get \Lp\ working on these systems. <>= (defun loop-directory* (directory &rest args &key &allow-other-keys) #+allegro (apply #'directory directory :directories-are-files NIL :follow-symbolic-links NIL args) #+(or clozure digitool) (apply #'directory directory :follow-links NIL args) #+clisp (apply #'directory directory :circle T :if-does-not-exist :ignore args) #+(or cmucl scl) (apply #'directory directory :follow-links NIL :truenamep NIL args) #+lispworks (apply #'directory directory :link-transparency NIL args) #+sbcl (apply #'directory directory :resolve-symlinks NIL args) #-(or allegro clozure digitool clisp cmucl scl lispworks sbcl) (apply #'directory directory args)) (defun loop-list-files (directory) (let ((directory (pathname-utils:to-directory directory))) (let* ((directory (pathname-utils:pathname* directory)) (entries (ignore-errors (loop-directory* (merge-pathnames pathname-utils:*wild-file* directory))))) (remove-if #'directory-p entries)))) (defun loop-list-directories (directory) (let ((directory (pathname-utils:to-directory directory))) (let* (#-(or abcl cormanlisp genera xcl) (wild (merge-pathnames #-(or abcl allegro cmucl lispworks sbcl scl xcl) pathname-utils:*wild-directory* #+(or abcl allegro cmucl lispworks sbcl scl xcl) "*.*" directory)) (dirs #+(or abcl xcl) (system:list-directory directory) #+cormanlisp (cl::directory-subdirs directory) #+genera (handler-case (loop for (p . k) in (fs:directory-list directory) when (eql :directory k) collect p) (fs:directory-not-found () nil)) #+clozure (ignore-errors (directory* wild :directories T :files NIL)) #+mcl (ignore-errors (directory* wild :directories T)) #-(or abcl xcl cormanlisp genera clozure mcl) (loop-directory* wild))) (loop for path in dirs when (directory-p path) collect (pathname-utils:force-directory path))))) @ %def loop-list-directories loop-list-files \subsection{{\tt GROUP}}\label{sec:group} We just need to verify if the group exists and modify [[*client*]]. <>= (defun cmd-group (r) (with-auth (with-n-args 1 r (let ((g (car (request-args r)))) (with-group g r (set-group! g) (multiple-value-bind (low high len) (group-high-low g) (let ((ln (fmt "~a ~a ~a ~a" len low high g))) (setf (client-article *client*) low) (make-response :code 211 :request r :data ln)))))))) (defun group? (g) (in-groups (ignore-errors (directory-p g)))) (defun set-group! (g) (setf (client-group *client*) g)) @ %def group? \subsection{{\tt BODY}, {\tt HEAD}, {\tt ARTICLE} e {\tt NEXT}} \label{sec:typical-cmds} Here we have an illustration of the expressive power of first-class procedures. We want to implement the commands {\tt BODY}, {\tt HEAD}, {\tt ARTICLE} and {\tt NEXT}. The ways to invoke them are (1) with no argument, (2) with a single integer argument and (3) with a text argument. In case (1), we use the number of the current article---which is kept in [[*client*]]. In case (2), the NNTP client specifies the article number. In case (3), the NNTP client specifies the exact {\tt message-id}, which forces us to query the index---see Section~\ref{sec:index} for the implementation of the index. <>= (defun typical-cmd-head-body-article (r fn-name) (with-auth (with-group-set (let ((args (request-args r))) (cond ((null args) (funcall fn-name r (client-group *client*) (client-article *client*))) ((= 1 (length args)) (let* ((n-or-mid (car args))) (cond ((string-integer? n-or-mid) (funcall fn-name r (client-group *client*) n-or-mid)) (t (multiple-value-bind (group n-str) (lookup-index n-or-mid) (if (and group n-str) (funcall fn-name r group n-str) (bad-input r (fmt "Unknown article ~a." n-or-mid)))))))) (t (bad-input r "No, no: it takes at most two arguments."))))))) (defun cmd-head (r) (typical-cmd-head-body-article r #'head-response)) (defun cmd-body (r) (typical-cmd-head-body-article r #'body-response)) (defun cmd-article (r) (typical-cmd-head-body-article r #'article-response)) (defun article-response (r g i) (typical-cmd-response 220 r g i :article)) (defun head-response (r g i) (typical-cmd-response 221 r g i :head)) (defun body-response (r g i) (typical-cmd-response 222 r g i :body)) @ In processing such commands, we need to verify the existence of files {\em et cetera}. The procedure [[fetch-article]] returns [[nil]] when the article requested doesn't exist and it throws [[sb-posix:syscall-error]] due to [[sb-posix:stat-size]], which we need to find the size of the article. We need to know the file size not only to allocate an array at the right size, but also because we must provide the size when producing the \verb|OVERVIEW|. If a problem such as [[sb-posix:syscall-error]] appears, we just inform the client and terminate the request---nothing else to do. Our error checking strategy below is the following. If the user requests an article we don't have, we figure it out with [[fetch-article]]. If a system error come up such as a failed system call, we report it to the user and log it---there's probably nothing else we can do. The same is the case for all other errors. What I want to do is to log it. What's important is that we seem to have enough code to treat any future case we might want to add and we're not ignoring any errors at all. <>= (defun typical-cmd-response (code r g i requested-part) (handler-case (let* ((path (fmt "groups/~a/~a" g i)) (headers-raw (read-headers-raw path))) (if (null headers-raw) (make-response :code 400 :request r :data (fmt "article ~a/~a does not exist" g i)) (let* ((a (parse-article headers-raw)) ;; FIX: Pass the raw bytes to extract-mid, not the parsed struct (mid (extract-mid headers-raw)) (header-length (length headers-raw)) (payload (ecase requested-part (:head (article-head a)) (:body (make-file-stream-payload :path path :start-pos header-length)) (:article (list (article-head a) (crlf) (crlf) (make-file-stream-payload :path path :start-pos header-length)))))) (prepend-response-with (fmt "~a ~a" i mid) (make-response :multi-line 'yes :code code :request r :data payload))))) (sb-posix:syscall-error (c) (let ((msg (fmt "article ~a/~a: ~a" g i c))) (stderr msg) (make-response :code 400 :request r :data msg))) (error (c) (let ((msg (fmt "unexpected condition when reading article ~a/~a: ~a" g i c))) (stderr msg) (make-response :code 400 :request r :data msg))))) @ %def typical-cmd-response The command \verb|NEXT| has a slight different semantics. <>= (defun cmd-next (r) (with-auth (let ((g (client-group *client*)) (n-cur (client-article *client*))) (cond ((not g) (bad-input :code 412 r "must say GROUP first")) (t (multiple-value-bind (low high len) (group-high-low g) (declare (ignore low len)) (cond ((= n-cur high) (bad-input r "you are at the last article already")) (t (article-next! r g))))))))) (defun article-next! (r g) (setf (client-article *client*) (1+ (client-article *client*))) (let ((cur (client-article *client*))) (make-response :code 223 :request r :data (fmt "~a ~a" cur (mid-by-name g cur))))) (defun mid-by-name (g name) (extract-mid (fetch-article g name))) @ %def cmd-next article-next! mid-by-name \subsection{{\tt XOVER}}\label{sec:xover} The procedure [[cmd-xover]] is used to figure out what the user said. Once we have that figured out, we invoke [[xover]], which finishes the work. Notice that when the argument [[to]] from [[xover]] is [[NIL]], then the user is asking for articles indexed from the integer [[fr]] to the last one. <>= (defun cmd-xover (r) (with-auth (with-group-set (let ((args (request-args r))) (cond ((null args) (xover r (client-article *client*) (client-article *client*))) ((= 1 (length args)) (multiple-value-bind (s v) (cl-ppcre:scan-to-strings "([0-9]+)([-]?)([0-9]*)" (car args)) (cond ((not s) (make-response :code 502 :request r :data "bad syntax")) (t (let ((fr (parse-integer (aref v 0))) (hifen (aref v 1)) (to (ignore-errors (parse-integer (aref v 2))))) (when (not (string= hifen "-")) (setq to fr)) (xover r fr to)))))) (t (make-response :code 502 :request r :data "bad syntax"))))))) (defun xover (r from to) (assert (client-group *client*)) (let* ((g (client-group *client*)) (ls (get-articles g from to))) (cond ((= 0 (length ls)) (make-response :code 420 :request r :data "no articles in the range")) (t (prepend-response-with "Okay, your overview follows..." (make-response :code 224 :request r :multi-line 'yes :data (str:join (crlf-string) (loop for i in ls collect (xover-format-line i (remove-if-not #'(lambda (h) (member (car h) (xover-headers) :test #'string=)) (fetch-headers g i))))))))))) (defun xover-format-line (i hs) (str:concat (fmt "~a~a" i #\tab) (str:join #\tab (mapcar #'(lambda (h) (get-header h hs)) (xover-headers))))) (defun xover-headers () '("subject" "from" "date" "message-id" "references" "line-count" "byte-count")) @ \subsection{{\tt MODE READER}}\label{sec:mode-reader} So, we're always in reader mode, so we just ignore this command. <>= (defun cmd-mode (r) ;; Whatever. (make-response :code 200 :request r :data "Sure thing.")) @ %def \subsection{{\tt DATE}}\label{sec:date} It's always useful to know the time and date at a computer. We should surely format it a bit better than what {\tt now} does. <>= (defun cmd-date (r) (make-response :code 201 :request r :data (format-timestring nil (now)))) @ %def \subsection{{\tt QUIT}}\label{sec:quit} The use of {\tt QUIT} has a conection to [[main-loop]]: when the user says {\tt QUIT}, [[main-loop]] must terminate. <>= (defun cmd-quit (r) (make-response :code 205 :data "Good-bye." :request r)) @ %def \subsection{{\tt DD}}\label{sec:dd} The command {\tt DD} means ``[d]isplay client [d]ata structures''. It shows to the client the internal state of how the server sees it. I've used only for debugging and it's not really useful any longer. I'm going to remove this very soon. <>= (defun cmd-dd (r) (make-response :code 200 :data (fmt "state: ~a" *client*) :request r)) @ %def \subsection{{\tt POST}}\label{sec:post} If the client says {\tt POST}, then we continue to read line after line until we find \verb|.\r\n|. Having done that, we must check whether we have a conformant article at hand. The definition of conformant is given by [[conforms?]]: the article must have the headers required by [[headers-required-from-clients]]. <>= <> (defun suggest-message-id (&optional (n 20)) (fmt "<~a@loop>" (random-string n))) (defun random-string (size) (let* ((universe "abcdefghijklmnopqrstuvwxyz") (len (length universe)) (state (make-random-state t)) mid) (dotimes (c size) (setq mid (cons (char universe (random len state)) mid))) (coerce mid 'string))) @ Now it's time to look at the header \verb|newsgroups|. The name of each group must conform to the expression <>= ^[^\\s/]+$ @ In other words, let group names go wild. They cannot contain a slash or a space of any kind anywhere on the name---more literally: they must begin with any character that's not a space and must have at least one character. The problem wish slashes is that each group will be a directory on a UNIX file system, so we cannot let slashes appear. People should have total freedom in naming groups. If users create groups that mess up the local organization, then people should discuss the matter and find a solution. Let's let people mess it up instead of trying to stop them. <>= (defun cmd-post (r) (with-auth (let ((suggested (suggest-message-id))) (send-response! (make-response :code 340 :data (fmt "Okay, go ahead. Suggested message-id ~a." suggested))) (let ((tmp (fmt "post-~a.tmp" (random-string 12)))) (unwind-protect (progn (in-tmp (nntp-read-article-to-disk tmp)) (stderr "finished writing article to disk in tmp/~a~%" tmp) (let* ((headers-raw (in-tmp (read-headers-raw tmp))) ;; headers-raw includes the terminating CRLF CRLF; strip it ;; so parse-headers does not see an empty final line. (headers-str (map 'string #'code-char (subseq headers-raw 0 (max 0 (- (length headers-raw) 4)))))) (stderr "finished reading the headers of the tmp file~%") (stderr "headers string: ~a~%" headers-str) (multiple-value-bind (okay? error) (conforms-headers? headers-str) (cond ((not okay?) (make-response :code 400 :request r :data (fmt "Sorry. Your article doesn't conform: ~a." error))) (t (multiple-value-bind (code reply) (post tmp headers-raw) (stderr "I'd post this article now.~%") (update-last-post-date! (client-username *client*)) (make-response :code code :request r :data reply))))))) (ignore-errors (delete-file tmp))))))) @ It's time to write the action of posting. The article has already been streamed to a temporary file by [[nntp-read-article-to-disk]]; we hold only the headers in memory. [[post]] receives the path and the raw header bytes, ensures {\tt message-id} and {\tt date} are present (rewriting the temp file only when a header is missing), then places a copy into each destination group. Cross-posting still duplicates the article on the file system for now. (We will replace the copies with symbolic links once the index work is finished.) <>= (defun post (tmp-path headers-raw) (let* ((headers-str (map 'string #'code-char (subseq headers-raw 0 (max 0 (- (length headers-raw) 4))))) (headers (parse-headers headers-str)) (ngs (newsgroups-header->list (get-header "newsgroups" headers))) (body-start (length headers-raw))) ;; This has gotten complicated. We need to move TMP-PATH into the ;; first newsgroup and then symlink the article to all other ;; groups. This means we give a different treatment to ;; groups. This isn't so good. Looks like we should create a ;; file-system index of articles and then symlink each article to ;; each group---a more general treatment. Do this crazy thing now ;; and then generalize only later once you've cleaned this code up ;; so much. So here's your NEXT STEP: forget symlinks for now; ;; copy the temporary article to each directory before you post ;; it; IMPLEMENT COPY-TEMPORARY-FILE-TO and test the system. (dolist (ng ngs) (save-article-insist ng (get-next-article-id ng) tmp-path)) (values 240 "Thank you! Your article has been saved."))) (defun newsgroups-header->list (s) (mapcar #'(lambda (n) (str:trim (string-downcase n))) (str:split "," s))) @ %def post <>= (defun update-last-post-date! (username) (let ((u (get-account username))) (setf (account-last-post u) (get-universal-time)))) @ %def update-last-post-date! The job of [[save-article-try]] is to atomically allocate an article name in a group. The atomicity is achieved with [[:if-exists]] in the [[with-open-file]] macro. When [[if-exists]] is {\tt nil}, the {\tt open} system call uses the flag \verb|O_EXCL|, giving us an atomic operation. So the first [[with-open-file]] allocates the name. If successful, we copy the already-streamed article from [[src-path]] into a temporary file next to it and atomically rename it into place. <>= (defun rename-no-extension (old new) (rename-file (merge-pathnames old (uiop:getcwd)) (merge-pathnames (make-pathname :name new :type :unspecific) (uiop:getcwd)))) (defun save-article-try (name-try src-path) (stderr "going to try to save ~a with name ~a~%" src-path name-try) (let ((name (fmt "~a" name-try)) (tmp (fmt "~a.tmp" name-try))) (with-open-file (s name :direction :output :if-exists nil ; atomic O_EXCL allocation :if-does-not-exist :create) (when (null s) (stderr "warning: save-article-try: ~a exists~%" name) (return-from save-article-try 'name-exists))) (stderr "going to rename in save-article-try~%") ;; Because of my mess with directory paths, I'm having to do the ;; ../../ here. This is bad design. We need to find a solution ;; for this. But don't come up with the use ;; absolute-paths-solution. Relative paths is always better. (stderr "current-directory: ~a~%" (uiop:getcwd)) (stderr "default-pathname: ~a~%" *default-pathname-defaults*) (stderr "rename-no-extension ~a ~a~%" (fmt "tmp/~a" src-path) (fmt "groups/local.test/~a" name)) (rename-no-extension (fmt "../../tmp/~a" src-path) name))) @ %def save-article-try The procedure [[save-article-insist]] insists on calling [[save-article-try]] until it finds an article name that has not been allocated. The argument [[name]] is an integer, incremented at each iteration (and it's the loop for x from x that makes the increments). <>= (defun save-article-insist (g name src-path) (loop for name from name do (in-group g (when (not (eql 'name-exists (save-article-try name src-path))) (return (values name t)))))) (defun get-next-article-id (g) (multiple-value-bind (low high len) (group-high-low g) (declare (ignore low len)) (1+ high))) @ %def save-article-insist get-next-article-id {\bf How to read lines in the NNTP protocol?} We need a new presentation here and a clean up operation. <>= (defun nntp-read-article-to-disk (tmp-path) "Streams an article to disk." (let ((success nil)) (with-open-file (out tmp-path :direction :output :element-type '(unsigned-byte 8) :if-exists :error) (loop ;; 1. Fetch a clean line from your highly optimized buffer (let ((line (nntp-read-line))) ;; 2. Detect terminator: A line containing exactly the byte 46 ('.') (when (end-of-article? line) (setf success t) (return tmp-path)) ;; 3. Dot-unstuffing: If the line starts with '.', skip the first byte (let ((start-idx (if (and (> (length line) 0) (= (aref line 0) 46)) 1 0))) (write-sequence line out :start start-idx))) ;; 4. Restore the CRLF that `nntp-read-line` stripped away (write-byte 13 out) (write-byte 10 out))))) (defun end-of-article? (ls) ;; A line is the NNTP terminator only if it's exactly the single byte ;; 46 (#\.). LENGTH/ELT instead of EQUAL-against-a-list, since ;; NNTP-READ-LINE hands back a vector. (and ls (not (eq ls :eof)) (= (length ls) 1) (= (elt ls 0) 46))) @ %def nntp-read-line-buffered nntp-read-article-to-disk Now let's move to the complicated side of life. This is the very heart of \Lp. Our data structure of interest here is a buffer with two pointers into it, {\tt head} and {\tt tail}. The head just points to a place in the buffer where data yet to be consumed begins and {\tt tail} points to where it ends. All data coming through the network (or coming locally from someone's keyboard on standard input) is stored in this buffer. <>= (defstruct nntp-stream-buffer (data (make-array 8192 :element-type '(unsigned-byte 8))) (head 0 :type fixnum) (tail 0 :type fixnum)) (defparameter *network-buffer* (make-nntp-stream-buffer)) (defun nntp-read-line (&optional (s *standard-input*)) (nntp-read-line-buffered s *network-buffer*)) @ %def *network-buffer* Why do we need a fast procedure here? Because I/O is the slowest thing in computers. For instance, if we allow attachments in posts (which we don't), the performance penalty is clearly noticeable. There is only one line-reading path. Bytes always land in [[*network-buffer*]]; [[find-lf]] and [[extract-line]] do the framing work; the sole branch is inside [[fetch-more-bytes]], which prefers a raw POSIX {\tt read(2)} and falls back to a [[read-byte]] loop only when the stream has no underlying file descriptor (Gray streams under SLIME, string streams in tests, and the like). Terminal sessions and {\tt tcpserver} sockets both take the fast path. The NNTP protocol establishes that line termination is done with \verb|\r\n|, but we also accept a bare \verb|\n| because we talk to the server from the command line with tools that do not always send CRLF. [[extract-line]] strips a preceding CR when it is present, so both terminators produce the same clean line of bytes. How to read the procedure [[nntp-read-line-buffered]]? The procedure is an entire small program with a library of helpers. The main program sits at the bottom. We begin by looking for a linefeed in the buffer. If there is one, an entire line is ready to be extracted. If there isn't, the buffer is either empty or holds only a partial line; in either case we refill. What does refilling do? First it checks that we still have room---a pathological line that never ends must not grow the buffer forever; we signal an error and let [[main-loop]] decide what to do. Next, any bytes that already form a partial line are slid to the front of the buffer with [[replace]] (a memmove). Then [[fetch-more-bytes]] obtains more data. Finally we search again for a linefeed and either extract the completed line or recurse to refill once more. A partial line left in the buffer after an EOF is returned to the caller; the caller should treat it as a broken client and close the connection. We deliver everything we have, even when the peer has already gone away. The only asymmetry that remains is how bytes arrive. [[%unix-read]] peels the stream with [[fd-stream-of]] until it finds an {\tt sb-sys:fd-stream}, then issues a single POSIX {\tt read(2)}. That path is taken for ordinary terminals and for sockets handed to us by {\tt tcpserver}. When the stream is a Gray stream (no file descriptor anywhere in its wrapper chain), [[fd-stream-of]] signals a {\tt type-error}; [[fetch-more-bytes]] catches it and falls back to [[%portable-read]], a [[read-byte]] loop that blocks for the first byte and then slurps whatever is immediately available. The contract is identical: return as soon as at least one byte is ready. The fallback costs nothing on the path we care about---human typing or SLIME---and keeps the framing logic completely shared. <>= (defun nntp-read-line-buffered (stream buffer) (let* ((data (nntp-stream-buffer-data buffer)) (capacity (length data))) (labels ((find-lf (start end) (position 10 data :start start :end end)) (extract-line (lf-pos) (let* ((start (nntp-stream-buffer-head buffer)) ;; If the previous character is CR (13), strip it (end-line (if (and (> lf-pos start) (= (aref data (1- lf-pos)) 13)) (1- lf-pos) lf-pos)) (line (subseq data start end-line))) (setf (nntp-stream-buffer-head buffer) (1+ lf-pos)) line)) (refill-and-extract-line () (let* ((head (nntp-stream-buffer-head buffer)) (tail (nntp-stream-buffer-tail buffer)) (remaining (- tail head))) ;; 1. Prevent buffer saturation/overflow (when (= remaining capacity) (stderr "line too long in the network buffer: [~A]" (subseq data head tail)) (error "no space left on the network buffer for the current line")) ;; 2. Shift remaining bytes to the start (fast memmove) (when (> remaining 0) (replace data data :start1 0 :start2 head :end2 tail)) (setf (nntp-stream-buffer-head buffer) 0) (setf (nntp-stream-buffer-tail buffer) remaining) ;; 3. Fetch data via the unified dispatch mechanism (let* ((n (fetch-more-bytes stream data remaining (- capacity remaining))) (new-tail (+ remaining n))) (setf (nntp-stream-buffer-tail buffer) new-tail) (let ((lf-pos (find-lf remaining new-tail))) (if lf-pos (extract-line lf-pos) (refill-and-extract-line))))))) ;; --- Main flow --- (let ((lf-pos (find-lf (nntp-stream-buffer-head buffer) (nntp-stream-buffer-tail buffer)))) (if lf-pos (extract-line lf-pos) (refill-and-extract-line)))))) (defun fetch-more-bytes (stream buf start max-count) "Try the fast POSIX path; fall back to a portable read-byte loop when the stream has no underlying fd (Gray streams under SLIME, string streams in tests, etc.)." (handler-case (%unix-read stream buf start max-count) (type-error () (%portable-read stream buf start max-count)))) (defun %portable-read (stream buf start max-count) ;; Blocks for the VERY FIRST byte to respect the wait-for-data contract, ;; preventing busy-waiting loops. (let ((first-byte (read-byte stream nil :eof))) (when (eq first-byte :eof) (error 'end-of-file :stream stream)) (setf (aref buf start) first-byte) ;; Slurp any IMMEDIATELY available bytes to minimize looping overhead, ;; up to max-count. 'listen' returns T if a byte is ready on the wire. (let ((count 1)) (loop while (and (< count max-count) (listen stream)) do (let ((b (read-byte stream nil :eof))) (when (eq b :eof) (loop-finish)) (setf (aref buf (+ start count)) b) (incf count))) count))) (defun %unix-read (stream buf start max-count) ;;; One POSIX read(2) on STREAM's underlying fd. ;;; ;;; Blocks until at least one byte arrives, then returns immediately with ;;; however many bytes the kernel has buffered (up to MAX-COUNT). ;;; Retries transparently on EINTR (signals can interrupt system calls). ;;; Signals END-OF-FILE when the peer closes the connection. (let ((fd (sb-sys:fd-stream-fd (fd-stream-of stream)))) (sb-sys:with-pinned-objects (buf) (loop (multiple-value-bind (n err) (sb-unix:unix-read fd (sb-sys:sap+ (sb-sys:vector-sap buf) start) max-count) (cond ((and (null n) (= err sb-unix:eintr))) ((or (null n) (zerop n)) (error 'end-of-file :stream stream)) (t (return n)))))))) (defun fd-stream-of (stream) "Peel away SYNONYM-STREAM and TWO-WAY-STREAM wrappers until we reach the underlying FD-STREAM. Signals a clear error if no FD-STREAM is found." (etypecase stream (sb-sys:fd-stream stream) (synonym-stream (fd-stream-of (symbol-value (synonym-stream-symbol stream)))) (two-way-stream (fd-stream-of (two-way-stream-input-stream stream))))) @ %def nntp-read-line-buffered <>= (defun list->bytes (ls) (mapcar #'data->bytes ls)) (defun vector->bytes (v) (mapcar #'data->bytes (coerce v 'list))) (defun data->bytes (d) (etypecase d (null nil) (integer (list d)) (string (string->bytes d)) (cons (list->bytes d)) (vector (vector->bytes d)))) (defun add-crlf-between (ls-of-ls) ;; Add \r\n to each ``line''. Returns List-of Byte. (mapcar #'(lambda (ls) (append ls (list 13 10))) ls-of-ls)) (defun string->bytes (s) (map 'list #'char-code s)) (defun bytes->string (ls) (map 'string #'code-char ls)) @ {\bf Does the article conform?} RFC 5536 obliges every article to have exactly \verb|date|, \verb|from|, \verb|message-id|, \verb|newsgroups|, \verb|path| and \verb|subject| headers. The headers \verb|path| is \Lp's responsibility, but it's used only in a network of servers---so we're leaving that out for now. The header \verb|message-id| is added by \Lp\ if the client doesn't write it itself. We cannot let the user post to any group if one of the groups is incorrectly named---for example, when the group doesn't exist. Why don't we post to the ones that are correct and warn the user of the ones that are incorrect? Because that is not prudent. The user could be trying to publish news to be received at the same time by various groups; we would make such plans all go down the drain. So we collect a list of newsgroups that don't exist (or whose names do not conform for any reason); if we find any such group, then we refuse posting and return a 400 code with a message describing which group names failed. Otherwise the article is susceptible to be saved by [[post]]. So the message here is that if an article is vetted by [[conforms?]], then it can be saved. Any problem with the article must be reported by [[conforms?]]. Something to think about is that we have a [[catch]] here and a [[throw]] in [[parse-headers]]. We also have a [[return]] here. It's getting hard to read this procedure because it's not easy to know that a procedure has to return a certain value to match the expectation of another procedure. I don't remember what [[catch]] does. I need to review this and then add the explanation for myself. If I don't remember how this works, other beginners won't know it either. It's unclear right now what the wise measure is, so we'll keep on moving. <>= (defun conforms? (bs) (catch 'article-syntax-error (conforms-headers? (article-head (parse-article bs))))) (defun conforms-headers? (headers-str) "HEADERS-STR is the raw header block (without the body). Returns (values T NIL) on success, or (values NIL error-string)." (catch 'article-syntax-error (let ((headers (parse-headers headers-str))) (let ((result (dolist (h (headers-required-from-clients)) (when (not (lookup h headers)) (return (fmt "missing the /~a/ header" h)))))) (cond ((stringp result) (values nil result)) (t (let ((ngs (newsgroups-header->list (get-header "newsgroups" headers))) (ngs-dont-exist nil)) (dolist (g ngs) (when (or (not (group-name-conforms? g)) (not (group? g))) (push g ngs-dont-exist))) (if (not (zerop (length ngs-dont-exist))) (values nil (fmt "the ~a ~a just ~a exist" (word-plural (length ngs-dont-exist) "newsgroup") (str:join ", " (sort ngs-dont-exist #'string<)) (word-plural (length ngs-dont-exist) "doesn't"))) (values t nil))))))))) (defun text/plain? (header-s) ;; I say T when S begins with "text/plain" or when S is "". (let* ((s (str:collapse-whitespaces header-s)) (needle "text/plain") (len (min (length needle) (length s)))) (or (zerop len) (and (<= (length needle) (length s)) (string= needle s :end1 len :end2 len))))) (defun headers-required-from-clients () '("from" "newsgroups" "subject" "message-id" "date")) @ %def conforms? headers-required-from-clients Notice that to simplify matters we're duplicating articles cross-posted. What we should do is write the article to the first group in the list of \verb|newsgroups| and then make a symbolic link to all others. Why are we saying this here? This should be said in the [[post]] procedure. You know what? Let's commit what we have right now and let's tackle the index. We gotta push on through. (But while I'm gone I meant it---everything is gonna be alright.) \subsection{{\tt CREATE-GROUP}} We allow every user to create their own groups. (Are we crazy?) When someone creates a group, we post an article to {\tt local.control.news} notifying everyone that a new group has been created. People then have a chance to subscribe to the new group. We assume that when someone creates a group, it's either no problem at all or it has been discussed with the community beforehand. <>= (defun cmd-create-group (r) (with-n-args 1 r (let ((g (string-downcase (car (request-args r))))) (multiple-value-bind (okay? reason) (group-name-conforms? g) (if (not okay?) (make-response :code 580 :request r :data (fmt "group name does not conform: ~a" reason)) (progn (multiple-value-bind (path created?) (in-groups (ensure-directories-exist (concatenate 'string g "/"))) (declare (ignore created?)) (if (not path) (make-response :code 581 :request r :data (fmt "could not create group ~a" (if (group? g) "because it already exists" "but we don't know why---sorry!"))) (progn (notify-group-created g) (make-response :code 280 :request r :data (fmt "group ~a created" g))))))))))) (defun group-name-conforms? (g) (conforms-to? g "<>")) @ %def CREATE-GROUP group-name-conforms? \subsection{{\tt REPL}} \lp\ is {\em totally} hackable. Users can say {\tt repl} to have complete control over their \lxxp\ process. <>= (defun cmd-repl (r) (if *enable-nntp-repl* (with-auth (repl r)) (make-response :code 400 :data "The REPL has been *explicitly* disabled by the sysadmin. :(" :request r))) @ If your users are not the hacker-type, you can disable the NNTP REPL with the command-line option [[--disable-nntp-repl]]. We decide not to hide the command in the list of commands given by saying {\tt HELP} to \lp\ because this way users are advertised about the commands that exist---they could be having fun, but their sysadmin doesn't think they're skilled enough. <>= (defparameter *enable-nntp-repl* t) @ \section{Publication of news} If you're interested in being notified about what's going on in the \lp, then subscribe to {\tt local.control.news}. Group creation, invitations {\em et cetera} are published there. <>= (defun notify-group-created (g) (post-notification :subject (fmt "new group ~a by ~a" g (client-username *client*)) :body (fmt "Blame ~a for the group ~a just created." (client-username *client*) g))) (defun notify-user-created (u) (post-notification :subject (fmt "new account ~a by ~a" u (client-username *client*)) :body (fmt "Blame ~a for inviting ~a." (client-username *client*) u))) (defun notify-user-unlocked (u) (let ((guilty (client-username *client*))) (post-notification :subject (fmt "account ~a unlocked by ~a" u guilty) :body (fmt "Blame ~a for unlocking ~a." guilty u)))) (defun post-notification (&key subject body) (in-groups (ensure-directories-exist "local.control.news/")) (when (group? "local.control.news") (let ((tmp (fmt "notify-~a.tmp" (random-string 10))) (mid (suggest-message-id))) (unwind-protect (progn (with-open-file (out tmp :direction :output :element-type '(unsigned-byte 8) :if-exists :supersede :if-does-not-exist :create) (flet ((wline (s) (write-sequence (map 'vector #'char-code s) out) (write-byte 13 out) (write-byte 10 out))) (wline (fmt "From: Loop")) (wline (fmt "Subject: ~a" subject)) (wline (fmt "Newsgroups: local.control.news")) (wline (fmt "Message-ID: ~a" mid)) (wline (fmt "Date: ~a" (get-date))) (write-byte 13 out) (write-byte 10 out) (write-sequence (map 'vector #'char-code body) out) (write-byte 13 out) (write-byte 10 out))) (let ((headers-raw (read-headers-raw tmp))) (post tmp headers-raw))) (ignore-errors (delete-file tmp)))))) @ %def notify-group-created notify-user-created \section{Algorithm of {\tt split-vector}} How should we describe the algorithm of [[split-vector]]? The objective is to consume lists of bytes such as [[(1 2 3 13 10 4 5 6 13 10 7 8 9)]] and produce lists of bytes such as [[((1 2 3) (4 5 6) (7 8 9))]]. We use [[search]] to find [[13 10]]. When we find this line termination, we have its position in the consumed list. With the position, we collect the line and we iterate searching for the next line. <>= (defun split-vector (delim v acc &key limit (so-far 1)) (let ((len (length v))) (split-vector-helper delim v len acc limit so-far 0))) (defun split-vector-helper (delim v len acc limit so-far start) (if (zerop len) acc (let ((pos (search delim v :start2 start :end2 len))) (cond ((or (not pos) (and limit (= so-far limit))) (nreverse (cons (subseq v start len) acc))) (t (split-vector-helper delim v len (cons (subseq v start (or pos len)) acc) limit (1+ so-far) (+ pos (length delim)))))))) @ %def split-vector \section{Article index}\label{sec:index} Every NNTP server needs to have an index of articles. Each article is indexed by its message-id. For example, the article % \begin{verbatim} From: Eva Lu Newsgroups: comp.unix Subject: xv6: an introduction Date: Wed, 24 Jul 2024 22:42:50 -0300 Message-ID: <87plr25js5.fsf@tor.soy> Cancel-Lock: sha1:82E8hqL1D1WQ9xc+CfnEYbFAaJo= MIME-Version: 1.0 Content-Type: text/plain; charset=utf-8 Content-Transfer-Encoding: 8bit \end{verbatim} % is indexed by the header {\tt message-id}. If you ask for article {\tt <87plr25js5.fsf@tor.soy>}, \Lp\ will tell you that you can find the article in group {\tt comp.unix} and its numeric ID is 37. Note that the number 37 is not inside the article, but only in the name of the file stored in the file system. The index, therefore, knows in which file is each {\tt message-id}. This fact implies that you cannot rename files in the file system---of course not: you'd changing identifiers inside a database. At this very moment we're designing the index. We're removing all references to the previous index. We're going to use a two-level UNIX directory. Every message id will be hashed to a {\tt hashid} and the hard link on the file system will be written in directory {\tt index/ha/sh/id}. I believe a two-level will be enough for \Lp---it's a system for a loop of friends. Though the code is dead, I'm keeping the next section here because we'll likely keep the procedure names, so seeing here will remind me what needs to be rewritten. \section{Essential operations relative to the index} Here's how to query the index or how to insert a new article into it. If [[insert-index]] returns [[nil]], then it's because it found no errors at all. The return value, therefore, indicates which error ocurred. We don't consider an error that an article has already been added to the index. For instance, we write multiple messages to the storage when someone cross-posts, but we'll add just a single record to the index, of course. So, an article already indexed is normal situation. Sure---in the future, we will not duplicate articles in storage; we will make symbolic links. We don't do that right now because Windows doesn't really support symbolic links. <>= (defun insert-index (m g i) (handler-case (clsql:insert-records :into "indices" :attributes '(id grp article) :values (list (str:trim m) (str:trim g) (str:trim i))) (clsql-sys:sql-database-data-error (c) (cond ((= (slot-value c 'clsql-sys::error-id) 19) 'already-indexed) (t ; We should log this error. ;(format t "other error: ~a" (slot-value c 'clsql-sys::database-message)) 'sql-error))) (:no-error () nil))) (defun lookup-index (mid) (let* ((found (clsql:select [grp] [article] :from [indices] :where [= [id] (str:trim mid)])) (article (first found)) (grp (first article)) (art (second article))) (when found (values grp art)))) @ %def insert-index lookup-index \section{Procedure to import the index from the file system} When we build the index from scratch, it's important to have a procedure capable of reading all articles in the file system and index them one by one. This is what we implement right here. For each news group, we get the name of the file relative to the article---that's the numeric ID of the article---and its {\tt message-id}. That's all we need to index it. <>= (defun index-from-fs! () (loop for path in (in-groups (directory "**/*")) do (let* ((g (str:trim (first (last (pathname-directory path))))) (i (str:trim (pathname-name path))) (m (str:trim (extract-mid (fetch-article g i))))) (when (> (length m) 0) (format t "article ~a/~a indexed by ~a~%" g i m) (insert-index m g i))))) (defun remake-index-from-fs () (drop-create-index!) (index-from-fs!)) @ Here's a program to build the index from a UNIX shell. <>= <> (ql:quickload :loop :silent t) (in-package #:loop) (connect-index! "message-id.db") (drop-create-index!) (index-from-fs!) (format t "Index built.~%") @ Be careful when using this program: it will build the database [[message-id.db]], which is an operation that needs to be done only once. Here's how to use it: % \begin{verbatim} %pwd /home/me/loop %sbcl --script build-index-from-fs.lisp Index built. %ls -l message-id.db -rw-r--r-- 1 me wheel 65536 Aug 26 13:32 message-id.db \end{verbatim} \section{Deletion and locking of inactive accounts}\label{sec:inactive-users} We now implement some of the \hyperref[principles]{principles} exposed earlier on page~\pageref{principles}. The program @<> would be run by {\tt cron} every day (at midnight, say). It checks all accounts that are inactive and locks them. If you want to keep accounts forever, just don't run the program. I've decided to never delete accounts that are inactive, so I only lock them. <>= <> (ql:quickload :loop :silent t) (in-package #:loop) (read-accounts!) (connect-index! "message-id.db") (remove-inactive-users!) (write-accounts!) @ %def cron-remove-inactive-users.lisp The entire program is really [[remove-inactive-users!]]. <>= (defun remove-inactive-users! () (loop for u in *accounts* do (let ((username (account-username u))) (cond ((and (not (locked? username)) (inactive-from-never-logged-in? username)) (post-notification :subject (fmt "account ~a removed by Loop" username) :body (fmt "~a didn't log in a first time (for ~a month~a) since account creation." username *months-never-logged-in* (plural *months-never-logged-in* "s"))) (remove-account! username) (format t "Removed ~a due to never logging in.~%" username)) ((and (not (locked? username)) (inactive-from-last-seen? username)) (post-notification :subject (fmt "account ~a locked by Loop" username) :body (fmt "~a disappeared for over ~a month~a." username *months-inactive-allowed* (plural *months-inactive-allowed* "s"))) (lock-account! username (fmt "disappeared for over ~a months" *months-inactive-allowed*)) (format t "Locked ~a due to long-time-no-see.~%" username)))))) @ %def remove-inactive-users! To remove an account, we need to first remove the username (to be removed) from anyone's list of friends. So, this involves scanning the entire list of accounts. Also, notice that delete ``may modify {\em sequence}''. More important is to understand that we really must {\tt setf} the return, otherwise we might find the deletion did not take effect---for example, when deleting the first element of a list. <>= (defun remove-account! (username) (loop for u in *accounts* do (setf (account-friends u) (delete username (account-friends u) :test #'equal))) (setf *accounts* (delete-if #'(lambda (a) (equal (account-username a) username)) *accounts*))) (defun lock-account! (username why) (let ((u (get-account username))) (setf (account-pass-locked u) (account-pass u)) (setf (account-pass u) "locked") (setf (account-pass-locked-why u) why))) @ %def remove-account! Accounts that do not have a creation date up until today---Tue Sep 17 21:37:18 ESAST 2024---will have its creation dates migrated to the \Lp\ epoch, which is January 1st 2024, the exact month in which \Lp\ was written. But notice that this migration is done only once. New system administrators of \Lp\ will never need to run this. (We do not remove this set of source code chunks because they serve as an example of how to do a migration of this sort.) <>= (defun loop-epoch () (encode-timestamp 0 0 0 0 1 1 2024)) (defun migrate-add-creation-and-post-date! () (read-accounts!) (loop for u in *accounts* do (if (not (account-creation u)) (setf (account-creation u) (timestamp-to-universal (loop-epoch))) (setf (account-last-post u) (account-seen u)))) (write-accounts!)) @ %def migrate-add-creation-and-post-date! Here's a program to run the migration in a UNIX shell. <>= <> (ql:quickload :loop :silent t) (in-package #:loop) (migrate-add-creation-and-post-date!) (format t "Accounts rewritten.~%") @ Now we write the procedures that discover what accounts are inactive. <>= (defparameter *months-inactive-allowed* 3) (defparameter *months-never-logged-in* 1) @ <>= (defun user-inactive? (username) (or (inactive-from-never-logged-in? username) (inactive-from-last-seen? username))) (defun inactive-from-never-logged-in? (username) (let ((u (get-account username))) (if (ever-logged-in? username) NIL (inactive-from? username *months-never-logged-in* #'(lambda () (account-creation u)))))) (defun locked? (username) (equal "locked" (account-pass (get-account username)))) (defun inactive-from-last-post? (username) (let ((last-post (account-last-post (get-account username))) (creation (account-creation (get-account username)))) (inactive-from? username *months-inactive-allowed* (if last-post #'(lambda () last-post) #'(lambda () creation))))) (defun inactive-from-last-seen? (username) (let* ((u (get-account username)) (last-seen (account-seen u)) (creation (account-creation u))) (inactive-from? username *months-inactive-allowed* (if last-seen #'(lambda () last-seen) #'(lambda () creation))))) (defun inactive-from? (username months timestamp-source) (declare (ignore username)) (timestamp< (timestamp+ (universal-to-timestamp (funcall timestamp-source)) months :month) (now))) (defun ever-logged-in? (username) (account-seen (get-account username))) (defun never-logged-in? (username) (not (ever-logged-in? username))) (defun list-inactive-users () (loop for u in *accounts* do (format t "Username ~a is inactive? ~a~%" (account-username u) (user-inactive? (account-username u))))) @ \section{Macros to ease writing}\label{sec:dsl} These macros make up a tiny language to ease the writing of \lp. For example, when we need to access the group database, we use [[in-groups]]. When a certain command demands authentication, we use [[with-auth]]. <>= (defmacro in-dir (dir &rest body) `(let ((*default-pathname-defaults* (truename ,dir))) (uiop:with-current-directory (,dir) ,@body))) @ Notice that we set [[*default-pathname-defaults*]] and we set the process' current working directory. That's not necessary because Lisp always uses [[*default-pathname-defaults*]] and does not care about the current working directory. We did this out of the fact that we used to invoke [[renameat2]] through the [[cffi]], but we don't use it any more. <>= (defmacro in-tmp (&rest body) `(in-dir "tmp/" ,@body)) (defmacro in-groups (&rest body) `(in-dir "groups/" ,@body)) (defun in-group-lambda (g fn) (in-dir g (funcall fn))) (defmacro in-group (g &rest body) `(in-group-lambda (fmt "groups/~a/" ,g) #'(lambda () ,@body))) (defmacro with-group (g r &rest body) (let ((g-var (gensym)) (r-var (gensym))) `(let ((,g-var ,g) (,r-var ,r)) (if (not (group? ,g-var)) (make-response :code 411 :request ,r-var :data (fmt "no such group ``~a''" ,g-var)) (progn ,@body))))) (defmacro with-n-args (n r &rest body) (let ((args-var (gensym)) (message-var (gensym)) (n-var n)) `(let ((,args-var (request-args r)) (,message-var ,(fmt "bad arguments: needs exactly ~a" n-var))) (if (not (= ,n-var (length ,args-var))) (make-response :code 400 :request ,r :data ,message-var) (progn ,@body))))) (defmacro with-group-set (&rest body) (let ((g-var (gensym))) `(let ((,g-var (client-group *client*))) (if (not ,g-var) (bad-input r "must say GROUP first") ,@body)))) (defmacro with-auth (&rest body) `(if (not (auth?)) (make-response :code 400 :data "You must authenticate first.") (progn ,@body))) @ %def in-groups with-group with-n-args with-group-set with-auth \section{Other procedures} Of small importance, they have nothing. <>= (defun conforms-to? (s re &optional error-msg) "Does string S conform to regular expression RE?" (let ((okay? (cl-ppcre:scan-to-strings re s))) (if okay? (values t nil) (values nil (or error-msg (fmt "must match ~a" re)))))) (defun print/finish (&rest args) (apply #'format (cons t args)) (finish-output)) (defun word-plural (n word) (let ((table '(("doesn't" . "don't") ("newsgroup" . "newsgroups")))) (let ((w (assoc word table :test #'string=))) (when (not w) (error "word not found")) (if (< n 2) (car w) (cdr w))))) (defun plural (v suffix) (if (> v 1) suffix "")) (eval-when (:compile-toplevel :load-toplevel :execute) (defun fmt (cstr &rest args) (apply #'format nil (list* cstr args)))) (defun out (stream &rest args) (apply #'format (cons stream args))) (defun stderr (&rest args) (when *debug* (apply #'out (cons *error-output* args)) (finish-output *error-output*))) (defun stdout (&rest args) (apply #'out (list* *standard-output* args))) (defun println (&rest args) (apply #'out (list* *standard-output* (str:concat (car args) "~%") (cdr args)))) (defun bad-input (r msg &key code) (make-response :code (or code 400) :data msg :request r)) (defun integer->string (n) (fmt "~a" n)) (defun crlf () (vector 13 10)) (defun crlf-string () (fmt "~c~c" #\return #\linefeed)) (defmacro mac (&rest body) `(macroexpand-1 ,@body)) @ %def bad-input crlf crlf-string \section{Tests} I studied the minimum to be able to add these tests as we comprehend better the direction in which we're going. A test system is essential for us to trust we can move forward without breaking important past decisions in the code. We should not include these tests in production code as we are doing right now. Divert them to a tests package or something that makes more sense. To run the tests, you need to invoke [[lisp-unit:run-tests]]. I believe that invoking it in a script will make all tests run. Oh, that should be included the binary, too, so that we can always test an archived project. That's a good idea. Make it a clingon option on the command line, though remember we will get rid of clingon. Of course, by including tests in the executable, we should isolate all the tests here. I know how---just define a test package and isolate it all in it. Packages are for namespace isolation. Lisp's ``systems'' are the type of packaging meant for loading. Lisp's ``packages'' are merely namespace isolation. Our test package should use \lp's package, so that we can use any of the procedures under testing. When we define a variable, it will be defined in the test package, not in \lp's, but when we are in \lp's package, we see no names of the test package. It's simple and a good solution. When testing, it's important for us not to clutter a production system---we will want to run tests on production systems. So what we need to do is to wrap any file system modification to a certain other directory in which \lp's tests will find the groups directory in place. Making that happen is as simple as changing [[*default-pathname-defaults*]]. <>= (setq lisp-unit:*print-failures* t) (define-test dispatching (assert-true (equalp (empty-response) (dispatch (make-request))))) (defun unix->nntp (s) "I substitute \n for \r\n" (str:replace-all (fmt "~a" #\linefeed) (crlf-string) s)) (defvar a-post (unix->nntp "From: root Message-id: Subject: test Newsgroups: local.test Quickest test of the West. ")) (defvar a-bad-post (unix->nntp "From: root Message-id: Subject: a bad test Newsgroups: local.test, bad.newsgroup A bad test from the biggest mouth of the south. ")) @ %def I removed the tests. The most central piece of the system has been rewritten. \section{How to produce the binary executable} Just say {\tt make loop} to your shell. <>= <> (ql:quickload :loop) (sb-ext:save-lisp-and-die #P"loop" :toplevel #'loop:main :executable t :save-runtime-options t) @ <>= #-quicklisp (let ((quicklisp-init (merge-pathnames "quicklisp/setup.lisp" (user-homedir-pathname)))) (when (probe-file quicklisp-init) (load quicklisp-init))) @ \section{How to get a log of \lp's communication} If you invoke \lxxp\ with option [[--logging]], you get logging on [[stderr]]: % \begin{verbatim} $ ./loop -s --logging > /dev/null @>>> 200 Welcome! Say ``help'' for a menu. quit @<<< quit @>>> 205 Good-bye. \end{verbatim} But remember that \lp\ runs various processes in parallel, so the question now is how to unify all processes' logs into a single one. We can't use \href{https://cr.yp.to/daemontools/multilog.html}{multilog}, for instance. The easiest solution now is to send all messages to {\tt syslog}. FreeBSD and GNU systems come with a program called {\tt logger}, which is able to write messages to {\tt syslog}. If you're using \href{https://cr.yp.to/daemontools/svscan.html}{{\tt svscan}}, then you can use a program such as % \begin{verbatim} $ cat log/run #!/bin/sh exec /usr/bin/logger -i -t loop \end{verbatim} % as your {\tt log/run} script. See {\tt logger(1)} for more information. Using {\tt logger(1)} means you need to set up {\tt syslog}, too. By the default, {\tt logger(1)} will use the {\em facility} {\tt user} and the {\em level} {\tt notice}. So you can specify in {\tt syslog.conf} the selector {\tt user.notice} and specify a log file such as {\tt /var/log/loop.log}. See {\tt syslog.conf(5)} and {\tt newsyslog(1)} for more information. \section{Package {\tt loop.lisp} as the compiler needs it} We now put together all source code chunks in the order the compiler needs to read it. One thing to keep in mind here is---I wonder if people that might read this source code would read the literate programming \LaTeX\ output or would they read [[loop.lisp]] directly. For literate programmers, it doesn't matter how [[loop.lisp]] turns out because only the compiler reads [[loop.lisp]]. But if we care about anyone who might read [[loop.lisp]], then we should perhaps tell our literate programming tools to generate a nice-looking file. For instance, I declare global variables in the chunks where it's used. But for someone reading [[loop.lisp]] directly, it is perhaps better if they would see all global variables at the top of the file. That's something to think about. <>= ;;; -*- mode: LISP; syntax: COMMON-LISP; package: LOOP; -*- (eval-when (:compile-toplevel :load-toplevel :execute) (ql:quickload '(<>) :silent t)) (clsql:file-enable-sql-reader-syntax) (defpackage #:loop (:use :common-lisp :local-time) (:import-from :lisp-unit define-test assert-true) (:import-from :org.shirakumo.filesystem-utils directory-p list-directories list-files) (:import-from :sb-sys interactive-interrupt) (:export :main)) (in-package #:loop) <> <
> <> <> <> <> <> <> <> <> <> <> <> <> <> <> <> <> <> <> <> <> <> <> <> <> <> <> <> <> <> <> <> <> <> <
> <> @ %def <>= (defparameter *debug* t) <> <> <> <> <> <> <> <> <> @ On which packages do we depend? <>= :lisp-unit :str :uiop :cl-ppcre :local-time :clsql-sqlite3 :clingon :filesystem-utils :ironclad/digest/sha256 @ The \lp\ system definition: <>= ;;; -*- Mode: LISP; Syntax: COMMON-LISP; -*- (asdf:defsystem :<> :version "<>" :description "<>" :depends-on (<>) :components ((:file "loop") (:file "version"))) @ \section{Other source code} The shell script {\tt format-def} is invoked whenever we build any Lisp source code. That's to format the source code a bit better for readers that will be reading it directly. It is not what we do. We read the documentation in PDF format and we work on the NOWEB file {\tt loop.nw}. But we know that potential readers will not do the same and will hack {\tt loop.lisp} directly. Paying respect to these readers, we try to format Lisp source code as best as possible. So we do two things: first, we produce the final source code in an order that should produce no warnings during compilation; second, we make sure there's one and only one blank line between procedure or macro definition. We don't add a blank line between global variables. The following shell script does the job. The first {\tt sed} program finds our definitions of interest and inserts a new blank line before the definition. Such action makes function definitions separated by two blank lines in some cases. We then remove the excess with the second program. Notice we need the {\tt -E} option because we're using the {\tt \textbar} metacharacter. The second program find a blank line as its first step. Then we say {\tt N} to expand the pattern space to include the next line. Then we delete the {\em first} blank line and not the second---that's what the {\tt D} command does. This strategy is explained by Dale Dougherty and Arnold Robbins in ``sed \& awk'' second edition, pages 112--114. <>= #!/bin/sh usage() { printf 'usage: %s [file.lisp]\n' $0 exit 1 } sed -E '/^\(defun |\(defmacro /{ i\ }' "$@" | \ sed '/^[ \t]*$/{ N /^[ \t]*\n$/D }' @ When we make a new release of \lp, we like to name its version as the tip of the source code repository. We get the information usually with a command line such as % \begin{verbatim} $ git log --oneline | head -1 | awk '{print $1}' 52663d1 \end{verbatim} % We achieve this with [[version.lisp]], which we generate before any release. This a bit of meta source code we use included in a meta Makefile we maintain. It's perfectly fine if you assume that [[version.lisp]] is manually written before we release a new version. When \Lp\ displays the welcome banner, the procedure [[welcome-message]] replaces the string [[HASH]] with the hash spelled out in [[version.lisp]]. \section*{Index of chunks} \nowebchunks \section*{Index of names} \nowebindex \end{document}