From: Ian Jackson Date: Wed, 1 Oct 2014 17:19:20 +0000 (+0100) Subject: polypath: Provide privsep mode X-Git-Tag: proposed.polypath.v4~10 X-Git-Url: http://www.chiark.greenend.org.uk/ucgi/~ian/git?p=secnet.git;a=commitdiff_plain;h=bc78a2b7c5f6ac9a1b0863dfcf77fba015187acf polypath: Provide privsep mode If secnet is going to drop privileges, it won't be able to call setsockopt(,,SO_BINDTODEVICE,) to set up the new polypath sockets. Provide an arrangement where this is done by a child forked before we drop privilege. Add some comments to the existing just-broken-down interface change handling code, saying whether and how they are used in the non-privsep and privsep cases. Signed-off-by: Ian Jackson --- diff --git a/README b/README index 3e8a157..fd3d31f 100644 --- a/README +++ b/README @@ -258,8 +258,12 @@ is one. network-manager always sets up a default route. The result is that the wifi always has a default route (so is useable); ppp (being a point-to-point link) does not need one. -The use of polypath currently requires that secnet have root -privilege, to make the setsockopt(,,SO_BINDTODEVICE,) call. +The use of polypath requires that secnet be started with root +privilege, to make the setsockopt(,,SO_BINDTODEVICE,) calls. If the +configuration specifies that secnet should drop privilege (see +`userid' above), secnet will keep a special process around for this +purpose; that process will handle local network interface changes but +does not deal with any packets, key exchange, etc. polypath support is only available when secnet is built against an IPv6-capable version of adns (because it wants features in the newer diff --git a/polypath.c b/polypath.c index 8249207..48d06a1 100644 --- a/polypath.c +++ b/polypath.c @@ -31,8 +31,12 @@ struct polypath { struct buffer_if lbuf; int monitor_fd; pid_t monitor_pid; + int privsep_incoming_fd; + int privsep_ipcsock_fd; }; +static void polypath_phase_shutdown(void *sst, uint32_t newphase); + #define LG 0, st->uc.cc.cl.description, &st->uc.cc.loc static const char *const default_loopback_ifname_pats[] = { @@ -389,6 +393,7 @@ static void subproc_problem(struct polypath *st, assert(!"not reached"); } +/* Used in non-privsep case, and in privsep child */ static void afterpoll_monitor(struct polypath *st, struct pollfd *fd, polypath_ppml_callback_type *callback) { @@ -404,6 +409,7 @@ static void afterpoll_monitor(struct polypath *st, struct pollfd *fd, subproc_problem(st,alr,emsg); } +/* Used in non-privsep case only - glue for secnet main loop */ static void polypath_afterpoll_monitor(void *state, struct pollfd *fds, int nfds) { @@ -453,6 +459,7 @@ static bool_t polypath_sendmsg(void *commst, struct buffer_if *buf, return allreasonable; } +/* Non-privsep: called in (sole) child. Privsep: in grandchild. */ static void child_monitor(struct polypath *st, int childfd) { dup2(childfd,1); @@ -462,6 +469,7 @@ static void child_monitor(struct polypath *st, int childfd) exit(-1); } +/* General utility function. */ static void start_subproc(struct polypath *st, void (*make_fdpair)(int[2]), void (*child)(struct polypath *st, int childfd), const char *desc) @@ -493,14 +501,275 @@ static void start_subproc(struct polypath *st, void (*make_fdpair)(int[2]), st->uc.cc.cl.description, desc, (long)st->monitor_pid); } +/* Non-privsep only: glue for forking the monitor, from the main loop */ static void polypath_phase_startmonitor(void *sst, uint32_t newphase) { struct polypath *st=sst; - start_subproc(st,pipe_cloexec,child_monitor,"interface monitor"); + start_subproc(st,pipe_cloexec,child_monitor, + "interface monitor (no privsep)"); register_for_poll(st,polypath_beforepoll, polypath_afterpoll_monitor,"polypath"); } +/*----- Privsep-only: -----*/ + +/* + * We use two subprocesses, a child and a grandchild. These are + * forked before secnet drops privilege. + * + * The grandchild is the same interface monitor helper script as used + * in the non-privsep case. But its lines are read by the child + * instead of by the main secnet. The child is responsible for + * creating the actual socket (binding it, etc.). Each socket is + * passed to secnet proper via fd passing, along with a data message + * describing the interface name and address. The child does not + * retain a list of current interfaces and addresses - it trusts the + * interface monitor to get that right. secnet proper still maintains + * that data structure. + * + * The result is that much of the non-privsep code can be reused, but + * just plumbed together differently. + * + * The child does not retain the socket after passing it on. + * Interface removals are handled similarly but without any fd. + * + * The result is that the configuration's limits on which interfaces + * and ports secnet may use are enforced by the privileged child. + */ + +struct privsep_mdata { + bool_t add; + char ifname[100]; + union iaddr ia; +}; + +static void papp_bad(struct polypath *st, void *badctx, const char *m, int ev) +{ + const struct privsep_mdata *mdata=(const void*)st->lbuf.start; + const char *addr_str=badctx; + + lg_perror(LG,M_WARNING,ev, + "error processing polypath address change %s %s [%s]: %s", + mdata->add ? "+" : "-", + mdata->ifname, addr_str, m); +} + +static void polypath_afterpoll_privsep(void *state, struct pollfd *fds, + int nfds) +/* In secnet proper; receives messages from child. */ +{ + struct polypath *st=state; + + if (nfds<1) return; + + int revents=fds[0].revents; + + const char *badbit=pollbadbit(revents); + if (badbit) subproc_problem(st,async_linebuf_broken,badbit); + + if (!(revents & POLLIN)) return; + + for (;;) { + if (st->lbuf.size==sizeof(struct privsep_mdata)) { + const struct privsep_mdata *mdata=(const void*)st->lbuf.start; + if (mdata->add && st->privsep_incoming_fd<0) + fatal("polypath (privsep): got add message data but no fd"); + if (!mdata->add && st->privsep_incoming_fd>=0) + fatal("polypath (privsep): got remove message data with fd"); + if (!memchr(mdata->ifname,0,sizeof(mdata->ifname))) + fatal("polypath (privsep): got ifname with no terminating nul"); + int af=mdata->ia.sa.sa_family; + if (!(af==AF_INET6 || af==AF_INET)) + fatal("polypath (privsep): got message data but bad AF %d",af); + const char *addr_str=iaddr_to_string(&mdata->ia); + polypath_record_ifaddr(st,papp_bad,(void*)addr_str, + mdata->add,mdata->ifname,addr_str, + &mdata->ia, st->privsep_incoming_fd); + st->privsep_incoming_fd=-1; + st->lbuf.size=0; + } + struct msghdr msg; + int fd; + size_t cmsgdatalen=sizeof(fd); + char cmsg_control_buf[CMSG_SPACE(cmsgdatalen)]; + struct iovec iov; + + FILLZERO(msg); + msg.msg_iov=&iov; + msg.msg_iovlen=1; + + iov.iov_base=st->lbuf.start+st->lbuf.size; + iov.iov_len=sizeof(struct privsep_mdata)-st->lbuf.size; + + if (st->privsep_incoming_fd<0) { + msg.msg_control=cmsg_control_buf; + msg.msg_controllen=sizeof(cmsg_control_buf); + } + + ssize_t got=recvmsg(st->monitor_fd,&msg,0); + if (got<0) { + if (errno==EINTR) continue; + if (iswouldblock(errno)) break; + fatal_perror("polypath (privsep): recvmsg failed"); + } + if (got==0) + subproc_problem(st,async_linebuf_eof,0); + + st->lbuf.size+=got; + + if (msg.msg_controllen) { + size_t cmsgdatalen=sizeof(st->privsep_incoming_fd); + struct cmsghdr *h=CMSG_FIRSTHDR(&msg); + if (!(st->privsep_incoming_fd==-1 && + h && + h->cmsg_level==SOL_SOCKET && + h->cmsg_type==SCM_RIGHTS && + h->cmsg_len==CMSG_LEN(cmsgdatalen) && + !CMSG_NXTHDR(&msg,h))) + subproc_problem(st,async_linebuf_broken,"bad cmsg"); + memcpy(&st->privsep_incoming_fd,CMSG_DATA(h),cmsgdatalen); + assert(st->privsep_incoming_fd>=0); + } + + } +} + +static void privsep_handle_ifaddr(struct polypath *st, + bad_fn_type *bad, void *badctx, + bool_t add, const char *ifname, + const char *ifaddr, + const union iaddr *ia, int fd_dummy) +/* In child: handles discovered wanted interfaces, making sockets + and sending them to secnet proper. */ +{ + struct msghdr msg; + struct iovec iov; + struct udpsock us={ .fd=-1 }; + size_t cmsgdatalen=sizeof(us.fd); + char cmsg_control_buf[CMSG_SPACE(cmsgdatalen)]; + + assert(fd_dummy==-1); + + struct privsep_mdata mdata; + FILLZERO(mdata); + mdata.add=add; + + size_t l=strlen(ifname); + if (l>=sizeof(mdata.ifname)) BAD("interface name too long"); + strcpy(mdata.ifname,ifname); + + COPY_OBJ(mdata.ia,*ia); + + iov.iov_base=&mdata; + iov.iov_len =sizeof(mdata); + + FILLZERO(msg); + msg.msg_iov=&iov; + msg.msg_iovlen=1; + + if (add) { + COPY_OBJ(us.addr,*ia); + bool_t ok=polypath_make_socket(st,bad,badctx,&us,ifname); + if (!ok) goto out; + + msg.msg_control=cmsg_control_buf; + msg.msg_controllen=sizeof(cmsg_control_buf); + + struct cmsghdr *h=CMSG_FIRSTHDR(&msg); + h->cmsg_level=SOL_SOCKET; + h->cmsg_type =SCM_RIGHTS; + h->cmsg_len =CMSG_LEN(cmsgdatalen); + memcpy(CMSG_DATA(h),&us.fd,cmsgdatalen); + } + + while (iov.iov_len) { + ssize_t got=sendmsg(st->privsep_ipcsock_fd,&msg,0); + if (got<0) { + if (errno!=EINTR) fatal_perror("polypath privsep sendmsg"); + got=0; + } else { + assert(got>0); + assert((size_t)got<=iov.iov_len); + } + iov.iov_base=(char*)iov.iov_base+got; + iov.iov_len-=got; + msg.msg_control=0; + msg.msg_controllen=0; + } + + out: + if (us.fd>=0) close(us.fd); +} + +static void child_privsep(struct polypath *st, int ipcsockfd) +/* Privsep child main loop. */ +{ + struct pollfd fds[2]; + + enter_phase(PHASE_CHILDPERSIST); + + st->privsep_ipcsock_fd=ipcsockfd; + start_subproc(st,pipe_cloexec,child_monitor, + "interface monitor (grandchild)"); + for (;;) { + int nfds=1; + int r=polypath_beforepoll(st,fds,&nfds,0); + assert(nfds==1); + assert(!r); + + fds[1].fd=st->privsep_ipcsock_fd; + fds[1].events=POLLIN; + + r=poll(fds,ARRAY_SIZE(fds),-1); + + if (r<0) { + if (errno==EINTR) continue; + fatal_perror("polypath privsep poll"); + } + if (fds[1].revents) { + if (fds[1].revents & (POLLHUP|POLLIN)) { + polypath_phase_shutdown(st,PHASE_SHUTDOWN); + exit(0); + } + fatal("polypath privsep poll parent socket revents=%#x", + fds[1].revents); + } + if (fds[0].revents & POLLNVAL) + fatal("polypath privsep poll child socket POLLNVAL"); + afterpoll_monitor(st,fds,privsep_handle_ifaddr); + } +} + +static void privsep_socketpair(int *fd) +{ + int r=socketpair(AF_UNIX,SOCK_STREAM,0,fd); + if (r) fatal_perror("socketpair(AF_UNIX,SOCK_STREAM,,)"); + setcloexec(fd[0]); + setcloexec(fd[1]); +} + +static void polypath_phase_startprivsep(void *sst, uint32_t newphase) +{ + struct polypath *st=sst; + + if (!will_droppriv()) { + add_hook(PHASE_RUN, polypath_phase_startmonitor,st); + return; + } + + start_subproc(st,privsep_socketpair,child_privsep, + "socket generator (privsep interface handler)"); + + BUF_FREE(&st->lbuf); + buffer_destroy(&st->lbuf); + buffer_new(&st->lbuf,sizeof(struct privsep_mdata)); + BUF_ALLOC(&st->lbuf,"polypath mdata buf"); + st->privsep_incoming_fd=-1; + + register_for_poll(st,polypath_beforepoll, + polypath_afterpoll_privsep,"polypath"); +} + static void polypath_phase_shutdown(void *sst, uint32_t newphase) { struct polypath *st=sst; @@ -522,6 +791,8 @@ static void polypath_phase_childpersist(void *sst, uint32_t newphase) #undef BAD #undef BADE +/*----- generic closure and module setup -----*/ + static list_t *polypath_apply(closure_t *self, struct cloc loc, dict_t *context, list_t *args) { @@ -557,7 +828,8 @@ static list_t *polypath_apply(closure_t *self, struct cloc loc, st->monitor_fd=-1; st->monitor_pid=0; - add_hook(PHASE_RUN, polypath_phase_startmonitor,st); + add_hook(PHASE_GETRESOURCES, polypath_phase_startprivsep,st); + add_hook(PHASE_SHUTDOWN, polypath_phase_shutdown, st); add_hook(PHASE_CHILDPERSIST,polypath_phase_childpersist,st);