Merge branch 'se/rebase' into next
[gitweb.git] / daemon.c
index 0ae9af0ad4d6c58d8c7cc0b615d8f9a7bc75572c..776749e3432fca916981d73d04a1fa19ed3d88b8 100644 (file)
--- a/daemon.c
+++ b/daemon.c
-#include "cache.h"
-#include "pkt-line.h"
 #include <signal.h>
 #include <sys/wait.h>
 #include <sys/socket.h>
+#include <sys/time.h>
+#include <sys/poll.h>
+#include <netdb.h>
 #include <netinet/in.h>
 #include <arpa/inet.h>
+#include <syslog.h>
+#include "pkt-line.h"
+#include "cache.h"
+#include "exec_cmd.h"
+
+static int log_syslog;
+static int verbose;
+static int reuseaddr;
+
+static const char daemon_usage[] =
+"git-daemon [--verbose] [--syslog] [--inetd | --port=n] [--export-all]\n"
+"           [--timeout=n] [--init-timeout=n] [--strict-paths]\n"
+"           [--base-path=path] [--user-path | --user-path=path]\n"
+"           [--reuseaddr] [directory...]";
+
+/* List of acceptable pathname prefixes */
+static char **ok_paths = NULL;
+static int strict_paths = 0;
+
+/* If this is set, git-daemon-export-ok is not required */
+static int export_all_trees = 0;
+
+/* Take all paths relative to this one if non-NULL */
+static char *base_path = NULL;
+
+/* If defined, ~user notation is allowed and the string is inserted
+ * after ~user/.  E.g. a request to git://host/~alice/frotz would
+ * go to /home/alice/pub_git/frotz with --user-path=pub_git.
+ */
+static char *user_path = NULL;
 
-static const char daemon_usage[] = "git-daemon [--inetd | --port=n]";
+/* Timeout, and initial timeout */
+static unsigned int timeout = 0;
+static unsigned int init_timeout = 0;
 
-static int upload(char *dir, int dirlen)
+static void logreport(int priority, const char *err, va_list params)
 {
-       if (chdir(dir) < 0)
+       /* We should do a single write so that it is atomic and output
+        * of several processes do not get intermingled. */
+       char buf[1024];
+       int buflen;
+       int maxlen, msglen;
+
+       /* sizeof(buf) should be big enough for "[pid] \n" */
+       buflen = snprintf(buf, sizeof(buf), "[%ld] ", (long) getpid());
+
+       maxlen = sizeof(buf) - buflen - 1; /* -1 for our own LF */
+       msglen = vsnprintf(buf + buflen, maxlen, err, params);
+
+       if (log_syslog) {
+               syslog(priority, "%s", buf);
+               return;
+       }
+
+       /* maxlen counted our own LF but also counts space given to
+        * vsnprintf for the terminating NUL.  We want to make sure that
+        * we have space for our own LF and NUL after the "meat" of the
+        * message, so truncate it at maxlen - 1.
+        */
+       if (msglen > maxlen - 1)
+               msglen = maxlen - 1;
+       else if (msglen < 0)
+               msglen = 0; /* Protect against weird return values. */
+       buflen += msglen;
+
+       buf[buflen++] = '\n';
+       buf[buflen] = '\0';
+
+       write(2, buf, buflen);
+}
+
+static void logerror(const char *err, ...)
+{
+       va_list params;
+       va_start(params, err);
+       logreport(LOG_ERR, err, params);
+       va_end(params);
+}
+
+static void loginfo(const char *err, ...)
+{
+       va_list params;
+       if (!verbose)
+               return;
+       va_start(params, err);
+       logreport(LOG_INFO, err, params);
+       va_end(params);
+}
+
+static int avoid_alias(char *p)
+{
+       int sl, ndot;
+
+       /* 
+        * This resurrects the belts and suspenders paranoia check by HPA
+        * done in <435560F7.4080006@zytor.com> thread, now enter_repo()
+        * does not do getcwd() based path canonicalizations.
+        *
+        * sl becomes true immediately after seeing '/' and continues to
+        * be true as long as dots continue after that without intervening
+        * non-dot character.
+        */
+       if (!p || (*p != '/' && *p != '~'))
+               return -1;
+       sl = 1; ndot = 0;
+       p++;
+
+       while (1) {
+               char ch = *p++;
+               if (sl) {
+                       if (ch == '.')
+                               ndot++;
+                       else if (ch == '/') {
+                               if (ndot < 3)
+                                       /* reject //, /./ and /../ */
+                                       return -1;
+                               ndot = 0;
+                       }
+                       else if (ch == 0) {
+                               if (0 < ndot && ndot < 3)
+                                       /* reject /.$ and /..$ */
+                                       return -1;
+                               return 0;
+                       }
+                       else
+                               sl = ndot = 0;
+               }
+               else if (ch == 0)
+                       return 0;
+               else if (ch == '/') {
+                       sl = 1;
+                       ndot = 0;
+               }
+       }
+}
+
+static char *path_ok(char *dir)
+{
+       static char rpath[PATH_MAX];
+       char *path;
+
+       if (avoid_alias(dir)) {
+               logerror("'%s': aliased", dir);
+               return NULL;
+       }
+
+       if (*dir == '~') {
+               if (!user_path) {
+                       logerror("'%s': User-path not allowed", dir);
+                       return NULL;
+               }
+               if (*user_path) {
+                       /* Got either "~alice" or "~alice/foo";
+                        * rewrite them to "~alice/%s" or
+                        * "~alice/%s/foo".
+                        */
+                       int namlen, restlen = strlen(dir);
+                       char *slash = strchr(dir, '/');
+                       if (!slash)
+                               slash = dir + restlen;
+                       namlen = slash - dir;
+                       restlen -= namlen;
+                       loginfo("userpath <%s>, request <%s>, namlen %d, restlen %d, slash <%s>", user_path, dir, namlen, restlen, slash);
+                       snprintf(rpath, PATH_MAX, "%.*s/%s%.*s",
+                                namlen, dir, user_path, restlen, slash);
+                       dir = rpath;
+               }
+       }
+       else if (base_path) {
+               if (*dir != '/') {
+                       /* Allow only absolute */
+                       logerror("'%s': Non-absolute path denied (base-path active)", dir);
+                       return NULL;
+               }
+               else {
+                       snprintf(rpath, PATH_MAX, "%s%s", base_path, dir);
+                       dir = rpath;
+               }
+       }
+
+       path = enter_repo(dir, strict_paths);
+
+       if (!path) {
+               logerror("'%s': unable to chdir or not a git archive", dir);
+               return NULL;
+       }
+
+       if ( ok_paths && *ok_paths ) {
+               char **pp;
+               int pathlen = strlen(path);
+
+               /* The validation is done on the paths after enter_repo
+                * appends optional {.git,.git/.git} and friends, but 
+                * it does not use getcwd().  So if your /pub is
+                * a symlink to /mnt/pub, you can whitelist /pub and
+                * do not have to say /mnt/pub.
+                * Do not say /pub/.
+                */
+               for ( pp = ok_paths ; *pp ; pp++ ) {
+                       int len = strlen(*pp);
+                       if (len <= pathlen &&
+                           !memcmp(*pp, path, len) &&
+                           (path[len] == '\0' ||
+                            (!strict_paths && path[len] == '/')))
+                               return path;
+               }
+       }
+       else {
+               /* be backwards compatible */
+               if (!strict_paths)
+                       return path;
+       }
+
+       logerror("'%s': not in whitelist", path);
+       return NULL;            /* Fallthrough. Deny by default */
+}
+
+static int upload(char *dir)
+{
+       /* Timeout as string */
+       char timeout_buf[64];
+       const char *path;
+
+       loginfo("Request for '%s'", dir);
+
+       if (!(path = path_ok(dir)))
                return -1;
-       chdir(".git");
 
        /*
         * Security on the cheap.
         *
-        * We want a readable HEAD, usable "objects" directory, and 
+        * We want a readable HEAD, usable "objects" directory, and
         * a "git-daemon-export-ok" flag that says that the other side
         * is ok with us doing this.
+        *
+        * path_ok() uses enter_repo() and does whitelist checking.
+        * We only need to make sure the repository is exported.
         */
-       if (access("git-daemon-export-ok", F_OK) ||
-           access("objects/00", X_OK) ||
-           access("HEAD", R_OK))
+
+       if (!export_all_trees && access("git-daemon-export-ok", F_OK)) {
+               logerror("'%s': repository not exported.", path);
+               errno = EACCES;
                return -1;
+       }
 
        /*
         * We'll ignore SIGTERM from now on, we have a
@@ -32,8 +257,10 @@ static int upload(char *dir, int dirlen)
         */
        signal(SIGTERM, SIG_IGN);
 
+       snprintf(timeout_buf, sizeof timeout_buf, "--timeout=%u", timeout);
+
        /* git-upload-pack only ever reads stuff, so this is safe */
-       execlp("git-upload-pack", "git-upload-pack", ".", NULL);
+       execl_git_cmd("upload-pack", "--strict", timeout_buf, ".", NULL);
        return -1;
 }
 
@@ -42,15 +269,17 @@ static int execute(void)
        static char line[1000];
        int len;
 
+       alarm(init_timeout ? init_timeout : timeout);
        len = packet_read_line(0, line, sizeof(line));
+       alarm(0);
 
        if (len && line[len-1] == '\n')
                line[--len] = 0;
 
-       if (!strncmp("git-upload-pack /", line, 17))
-               return upload(line + 16, len - 16);
+       if (!strncmp("git-upload-pack ", line, 16))
+               return upload(line+16);
 
-       fprintf(stderr, "got bad connection '%s'\n", line);
+       logerror("Protocol error: '%s'", line);
        return -1;
 }
 
@@ -71,23 +300,23 @@ static int max_connections = 25;
 
 /* These are updated by the signal handler */
 static volatile unsigned int children_reaped = 0;
-pid_t dead_child[MAX_CHILDREN];
+static pid_t dead_child[MAX_CHILDREN];
 
 /* These are updated by the main loop */
 static unsigned int children_spawned = 0;
 static unsigned int children_deleted = 0;
 
-struct child {
+static struct child {
        pid_t pid;
        int addrlen;
-       struct sockaddr_in address;
+       struct sockaddr_storage address;
 } live_child[MAX_CHILDREN];
 
-static void add_child(int idx, pid_t pid, struct sockaddr_in *addr, int addrlen)
+static void add_child(int idx, pid_t pid, struct sockaddr *addr, int addrlen)
 {
        live_child[idx].pid = pid;
        live_child[idx].addrlen = addrlen;
-       live_child[idx].address = *addr;
+       memcpy(&live_child[idx].address, addr, addrlen);
 }
 
 /*
@@ -177,9 +406,11 @@ static void check_max_connections(void)
        }
 }
 
-static void handle(int incoming, struct sockaddr_in *addr, int addrlen)
+static void handle(int incoming, struct sockaddr *addr, int addrlen)
 {
        pid_t pid = fork();
+       char addrbuf[256] = "";
+       int port = -1;
 
        if (pid) {
                unsigned idx;
@@ -199,60 +430,236 @@ static void handle(int incoming, struct sockaddr_in *addr, int addrlen)
        dup2(incoming, 0);
        dup2(incoming, 1);
        close(incoming);
+
+       if (addr->sa_family == AF_INET) {
+               struct sockaddr_in *sin_addr = (void *) addr;
+               inet_ntop(AF_INET, &sin_addr->sin_addr, addrbuf, sizeof(addrbuf));
+               port = sin_addr->sin_port;
+
+#ifndef NO_IPV6
+       } else if (addr->sa_family == AF_INET6) {
+               struct sockaddr_in6 *sin6_addr = (void *) addr;
+
+               char *buf = addrbuf;
+               *buf++ = '['; *buf = '\0'; /* stpcpy() is cool */
+               inet_ntop(AF_INET6, &sin6_addr->sin6_addr, buf, sizeof(addrbuf) - 1);
+               strcat(buf, "]");
+
+               port = sin6_addr->sin6_port;
+#endif
+       }
+       loginfo("Connection from %s:%d", addrbuf, port);
+
        exit(execute());
 }
 
 static void child_handler(int signo)
 {
        for (;;) {
-               pid_t pid = waitpid(-1, NULL, WNOHANG);
+               int status;
+               pid_t pid = waitpid(-1, &status, WNOHANG);
 
                if (pid > 0) {
                        unsigned reaped = children_reaped;
                        dead_child[reaped % MAX_CHILDREN] = pid;
                        children_reaped = reaped + 1;
+                       /* XXX: Custom logging, since we don't wanna getpid() */
+                       if (verbose) {
+                               char *dead = "";
+                               if (!WIFEXITED(status) || WEXITSTATUS(status) > 0)
+                                       dead = " (with error)";
+                               if (log_syslog)
+                                       syslog(LOG_INFO, "[%d] Disconnected%s", pid, dead);
+                               else
+                                       fprintf(stderr, "[%d] Disconnected%s\n", pid, dead);
+                       }
                        continue;
                }
                break;
        }
 }
 
-static int serve(int port)
+static int set_reuse_addr(int sockfd)
+{
+       int on = 1;
+
+       if (!reuseaddr)
+               return 0;
+       return setsockopt(sockfd, SOL_SOCKET, SO_REUSEADDR,
+                         &on, sizeof(on));
+}
+
+#ifndef NO_IPV6
+
+static int socksetup(int port, int **socklist_p)
 {
+       int socknum = 0, *socklist = NULL;
+       int maxfd = -1;
+       char pbuf[NI_MAXSERV];
+
+       struct addrinfo hints, *ai0, *ai;
+       int gai;
+
+       sprintf(pbuf, "%d", port);
+       memset(&hints, 0, sizeof(hints));
+       hints.ai_family = AF_UNSPEC;
+       hints.ai_socktype = SOCK_STREAM;
+       hints.ai_protocol = IPPROTO_TCP;
+       hints.ai_flags = AI_PASSIVE;
+
+       gai = getaddrinfo(NULL, pbuf, &hints, &ai0);
+       if (gai)
+               die("getaddrinfo() failed: %s\n", gai_strerror(gai));
+
+       for (ai = ai0; ai; ai = ai->ai_next) {
+               int sockfd;
+               int *newlist;
+
+               sockfd = socket(ai->ai_family, ai->ai_socktype, ai->ai_protocol);
+               if (sockfd < 0)
+                       continue;
+               if (sockfd >= FD_SETSIZE) {
+                       error("too large socket descriptor.");
+                       close(sockfd);
+                       continue;
+               }
+
+#ifdef IPV6_V6ONLY
+               if (ai->ai_family == AF_INET6) {
+                       int on = 1;
+                       setsockopt(sockfd, IPPROTO_IPV6, IPV6_V6ONLY,
+                                  &on, sizeof(on));
+                       /* Note: error is not fatal */
+               }
+#endif
+
+               if (set_reuse_addr(sockfd)) {
+                       close(sockfd);
+                       continue;
+               }
+
+               if (bind(sockfd, ai->ai_addr, ai->ai_addrlen) < 0) {
+                       close(sockfd);
+                       continue;       /* not fatal */
+               }
+               if (listen(sockfd, 5) < 0) {
+                       close(sockfd);
+                       continue;       /* not fatal */
+               }
+
+               newlist = realloc(socklist, sizeof(int) * (socknum + 1));
+               if (!newlist)
+                       die("memory allocation failed: %s", strerror(errno));
+
+               socklist = newlist;
+               socklist[socknum++] = sockfd;
+
+               if (maxfd < sockfd)
+                       maxfd = sockfd;
+       }
+
+       freeaddrinfo(ai0);
+
+       *socklist_p = socklist;
+       return socknum;
+}
+
+#else /* NO_IPV6 */
+
+static int socksetup(int port, int **socklist_p)
+{
+       struct sockaddr_in sin;
        int sockfd;
-       struct sockaddr_in addr;
 
-       signal(SIGCHLD, child_handler);
-       sockfd = socket(PF_INET, SOCK_STREAM, IPPROTO_IP);
+       sockfd = socket(AF_INET, SOCK_STREAM, 0);
        if (sockfd < 0)
-               die("unable to open socket (%s)", strerror(errno));
-       memset(&addr, 0, sizeof(addr));
-       addr.sin_port = htons(port);
-       addr.sin_family = AF_INET;
-       if (bind(sockfd, (void *)&addr, sizeof(addr)) < 0)
-               die("unable to bind to port %d (%s)", port, strerror(errno));
-       if (listen(sockfd, 5) < 0)
-               die("unable to listen to port %d (%s)", port, strerror(errno));
+               return 0;
+
+       memset(&sin, 0, sizeof sin);
+       sin.sin_family = AF_INET;
+       sin.sin_addr.s_addr = htonl(INADDR_ANY);
+       sin.sin_port = htons(port);
+
+       if (set_reuse_addr(sockfd)) {
+               close(sockfd);
+               return 0;
+       }
+
+       if ( bind(sockfd, (struct sockaddr *)&sin, sizeof sin) < 0 ) {
+               close(sockfd);
+               return 0;
+       }
+
+       if (listen(sockfd, 5) < 0) {
+               close(sockfd);
+               return 0;
+       }
+
+       *socklist_p = xmalloc(sizeof(int));
+       **socklist_p = sockfd;
+       return 1;
+}
+
+#endif
+
+static int service_loop(int socknum, int *socklist)
+{
+       struct pollfd *pfd;
+       int i;
+
+       pfd = xcalloc(socknum, sizeof(struct pollfd));
+
+       for (i = 0; i < socknum; i++) {
+               pfd[i].fd = socklist[i];
+               pfd[i].events = POLLIN;
+       }
+
+       signal(SIGCHLD, child_handler);
 
        for (;;) {
-               struct sockaddr_in in;
-               socklen_t addrlen = sizeof(in);
-               int incoming = accept(sockfd, (void *)&in, &addrlen);
-
-               if (incoming < 0) {
-                       switch (errno) {
-                       case EAGAIN:
-                       case EINTR:
-                       case ECONNABORTED:
-                               continue;
-                       default:
-                               die("accept returned %s", strerror(errno));
+               int i;
+
+               if (poll(pfd, socknum, -1) < 0) {
+                       if (errno != EINTR) {
+                               error("poll failed, resuming: %s",
+                                     strerror(errno));
+                               sleep(1);
+                       }
+                       continue;
+               }
+
+               for (i = 0; i < socknum; i++) {
+                       if (pfd[i].revents & POLLIN) {
+                               struct sockaddr_storage ss;
+                               unsigned int sslen = sizeof(ss);
+                               int incoming = accept(pfd[i].fd, (struct sockaddr *)&ss, &sslen);
+                               if (incoming < 0) {
+                                       switch (errno) {
+                                       case EAGAIN:
+                                       case EINTR:
+                                       case ECONNABORTED:
+                                               continue;
+                                       default:
+                                               die("accept returned %s", strerror(errno));
+                                       }
+                               }
+                               handle(incoming, (struct sockaddr *)&ss, sslen);
                        }
                }
-               handle(incoming, &in, addrlen);
        }
 }
 
+static int serve(int port)
+{
+       int socknum, *socklist;
+
+       socknum = socksetup(port, &socklist);
+       if (socknum == 0)
+               die("unable to allocate any listen sockets on port %u", port);
+
+       return service_loop(socknum, socklist);
+}
+
 int main(int argc, char **argv)
 {
        int port = DEFAULT_GIT_PORT;
@@ -271,17 +678,77 @@ int main(int argc, char **argv)
                                continue;
                        }
                }
-
                if (!strcmp(arg, "--inetd")) {
                        inetd_mode = 1;
+                       log_syslog = 1;
+                       continue;
+               }
+               if (!strcmp(arg, "--verbose")) {
+                       verbose = 1;
+                       continue;
+               }
+               if (!strcmp(arg, "--syslog")) {
+                       log_syslog = 1;
+                       continue;
+               }
+               if (!strcmp(arg, "--export-all")) {
+                       export_all_trees = 1;
+                       continue;
+               }
+               if (!strncmp(arg, "--timeout=", 10)) {
+                       timeout = atoi(arg+10);
                        continue;
                }
+               if (!strncmp(arg, "--init-timeout=", 15)) {
+                       init_timeout = atoi(arg+15);
+                       continue;
+               }
+               if (!strcmp(arg, "--strict-paths")) {
+                       strict_paths = 1;
+                       continue;
+               }
+               if (!strncmp(arg, "--base-path=", 12)) {
+                       base_path = arg+12;
+                       continue;
+               }
+               if (!strcmp(arg, "--reuseaddr")) {
+                       reuseaddr = 1;
+                       continue;
+               }
+               if (!strcmp(arg, "--user-path")) {
+                       user_path = "";
+                       continue;
+               }
+               if (!strncmp(arg, "--user-path=", 12)) {
+                       user_path = arg + 12;
+                       continue;
+               }
+               if (!strcmp(arg, "--")) {
+                       ok_paths = &argv[i+1];
+                       break;
+               } else if (arg[0] != '-') {
+                       ok_paths = &argv[i];
+                       break;
+               }
 
                usage(daemon_usage);
        }
 
-       if (inetd_mode)
+       if (log_syslog)
+               openlog("git-daemon", 0, LOG_DAEMON);
+
+       if (strict_paths && (!ok_paths || !*ok_paths)) {
+               if (!inetd_mode)
+                       die("git-daemon: option --strict-paths requires a whitelist");
+
+               logerror("option --strict-paths requires a whitelist");
+               exit (1);
+       }
+
+       if (inetd_mode) {
+               fclose(stderr); //FIXME: workaround
                return execute();
+       }
 
        return serve(port);
 }