Please note that diffs are not public domain; they are subject to the copyright notices on the relevant files. =================================================================== RCS file: /ftp/cvs/cvsroot/src/lib/librumpclient/rumpclient.c,v rcsdiff: /ftp/cvs/cvsroot/src/lib/librumpclient/rumpclient.c,v: warning: Unknown phrases like `commitid ...;' are present. retrieving revision 1.18 retrieving revision 1.51.2.1 diff -u -p -r1.18 -r1.51.2.1 --- src/lib/librumpclient/rumpclient.c 2011/01/24 17:47:51 1.18 +++ src/lib/librumpclient/rumpclient.c 2012/11/20 03:00:45 1.51.2.1 @@ -1,4 +1,4 @@ -/* $NetBSD: rumpclient.c,v 1.18 2011/01/24 17:47:51 pooka Exp $ */ +/* $NetBSD: rumpclient.c,v 1.51.2.1 2012/11/20 03:00:45 tls Exp $ */ /* * Copyright (c) 2010, 2011 Antti Kantee. All Rights Reserved. @@ -29,13 +29,36 @@ * Client side routines for rump syscall proxy. */ +#include "rumpuser_port.h" + +/* + * We use kqueue on NetBSD, poll elsewhere. Theoretically we could + * use kqueue on other BSD's too, but I haven't tested those. We + * want to use kqueue because it will give us the ability to get signal + * notifications but defer their handling to a stage where we do not + * hold the communication lock. Taking a signal while holding on to + * that lock may cause a deadlock. Therefore, block signals throughout + * the RPC when using poll. This unfortunately means that the normal + * SIGINT way of stopping a process while it is undergoing rump kernel + * RPC will not work. If anyone know which Linux system call handles + * the above scenario correctly, I'm all ears. + */ + +#ifdef __NetBSD__ +#define USE_KQUEUE +#endif + #include -__RCSID("$NetBSD"); +__RCSID("$NetBSD: rumpclient.c,v 1.51.2.1 2012/11/20 03:00:45 tls Exp $"); #include -#include #include #include +#include + +#ifdef USE_KQUEUE +#include +#endif #include #include @@ -43,6 +66,7 @@ __RCSID("$NetBSD"); #include #include +#include #include #include #include @@ -65,13 +89,17 @@ int (*host_connect)(int, const struct so int (*host_fcntl)(int, int, ...); int (*host_poll)(struct pollfd *, nfds_t, int); ssize_t (*host_read)(int, void *, size_t); -ssize_t (*host_sendto)(int, const void *, size_t, int, - const struct sockaddr *, socklen_t); +ssize_t (*host_sendmsg)(int, const struct msghdr *, int); int (*host_setsockopt)(int, int, int, const void *, socklen_t); +int (*host_dup)(int); +#ifdef USE_KQUEUE int (*host_kqueue)(void); int (*host_kevent)(int, const struct kevent *, size_t, struct kevent *, size_t, const struct timespec *); +#endif + +int (*host_execve)(const char *, char *const[], char *const[]); #include "sp_common.c" @@ -82,27 +110,95 @@ static struct spclient clispc = { static int kq = -1; static sigset_t fullset; -static int doconnect(int); -static int handshake_req(struct spclient *, uint32_t *, int, bool); +static int doconnect(void); +static int handshake_req(struct spclient *, int, void *, int, bool); + +/* + * Default: don't retry. Most clients can't handle it + * (consider e.g. fds suddenly going missing). + */ +static time_t retrytimo = 0; -int didrecon; +/* always defined to nothingness for now */ +#define ERRLOG(a) static int -send_with_recon(struct spclient *spc, const void *data, size_t dlen) +send_with_recon(struct spclient *spc, struct iovec *iov, size_t iovlen) { + struct timeval starttime, curtime; + time_t prevreconmsg; + unsigned reconretries; int rv; - do { - rv = dosend(spc, data, dlen); + for (prevreconmsg = 0, reconretries = 0;;) { + rv = dosend(spc, iov, iovlen); if (__predict_false(rv == ENOTCONN || rv == EBADF)) { - if ((rv = doconnect(1)) != 0) + /* no persistent connections */ + if (retrytimo == 0) { + rv = ENOTCONN; + break; + } + if (retrytimo == RUMPCLIENT_RETRYCONN_DIE) + _exit(1); + + if (!prevreconmsg) { + prevreconmsg = time(NULL); + gettimeofday(&starttime, NULL); + } + if (reconretries == 1) { + if (retrytimo == RUMPCLIENT_RETRYCONN_ONCE) { + rv = ENOTCONN; + break; + } + fprintf(stderr, "rump_sp: connection to " + "kernel lost, trying to reconnect ...\n"); + } else if (time(NULL) - prevreconmsg > 120) { + fprintf(stderr, "rump_sp: still trying to " + "reconnect ...\n"); + prevreconmsg = time(NULL); + } + + /* check that we aren't over the limit */ + if (retrytimo > 0) { + struct timeval tmp; + + gettimeofday(&curtime, NULL); + timersub(&curtime, &starttime, &tmp); + if (tmp.tv_sec >= retrytimo) { + fprintf(stderr, "rump_sp: reconnect " + "failed, %lld second timeout\n", + (long long)retrytimo); + return ENOTCONN; + } + } + + /* adhoc backoff timer */ + if (reconretries < 10) { + usleep(100000 * reconretries); + } else { + sleep(MIN(10, reconretries-9)); + } + reconretries++; + + if ((rv = doconnect()) != 0) continue; - if ((rv = handshake_req(&clispc, NULL, 0, true)) != 0) + if ((rv = handshake_req(&clispc, HANDSHAKE_GUEST, + NULL, 0, true)) != 0) continue; - rv = ENOTCONN; + + /* + * ok, reconnect succesful. we need to return to + * the upper layer to get the entire PDU resent. + */ + if (reconretries != 1) + fprintf(stderr, "rump_sp: reconnected!\n"); + rv = EAGAIN; + break; + } else { + _DIAGASSERT(errno != EAGAIN); break; } - } while (__predict_false(rv != 0)); + } return rv; } @@ -126,41 +222,68 @@ cliwaitresp(struct spclient *spc, struct /* are we free to receive? */ if (spc->spc_istatus == SPCSTATUS_FREE) { - struct kevent kev[8]; - int gotresp, dosig, rv, i; + int gotresp, dosig, rv; spc->spc_istatus = SPCSTATUS_BUSY; pthread_mutex_unlock(&spc->spc_mtx); dosig = 0; for (gotresp = 0; !gotresp; ) { - switch (readframe(spc)) { - case 0: - rv = host_kevent(kq, NULL, 0, - kev, __arraycount(kev), NULL); +#ifdef USE_KQUEUE + struct kevent kev[8]; + int i; + + /* + * typically we don't have a frame waiting + * when we come in here, so call kevent now + */ + rv = host_kevent(kq, NULL, 0, + kev, __arraycount(kev), NULL); + + if (__predict_false(rv == -1)) { + goto activity; + } + + /* + * XXX: don't know how this can happen + * (timeout cannot expire since there + * isn't one), but it does happen. + * treat it as an expectional condition + * and go through tryread to determine + * alive status. + */ + if (__predict_false(rv == 0)) + goto activity; + + for (i = 0; i < rv; i++) { + if (kev[i].filter == EVFILT_SIGNAL) + dosig++; + } + if (dosig) + goto cleanup; + + /* + * ok, activity. try to read a frame to + * determine what happens next. + */ + activity: +#else /* USE_KQUEUE */ + struct pollfd pfd; + + pfd.fd = clispc.spc_fd; + pfd.events = POLLIN; - /* - * XXX: don't know how this can - * happen (timeout cannot expire - * since there isn't one), but - * it does happen - */ - if (__predict_false(rv == 0)) - continue; - - for (i = 0; i < rv; i++) { - if (kev[i].filter - == EVFILT_SIGNAL) - dosig++; - } - if (dosig) - goto cleanup; + rv = host_poll(&pfd, 1, -1); +#endif /* !USE_KQUEUE */ + switch (readframe(spc)) { + case 0: continue; case -1: imalive = false; goto cleanup; default: + /* case 1 */ break; } @@ -209,12 +332,12 @@ cliwaitresp(struct spclient *spc, struct } static int -syscall_req(struct spclient *spc, int sysnum, +syscall_req(struct spclient *spc, sigset_t *omask, int sysnum, const void *data, size_t dlen, void **resp) { struct rsp_hdr rhdr; struct respwait rw; - sigset_t omask; + struct iovec iov[2]; int rv; rhdr.rsp_len = sizeof(rhdr) + dlen; @@ -222,83 +345,117 @@ syscall_req(struct spclient *spc, int sy rhdr.rsp_type = RUMPSP_SYSCALL; rhdr.rsp_sysnum = sysnum; - pthread_sigmask(SIG_SETMASK, &fullset, &omask); + IOVPUT(iov[0], rhdr); + IOVPUT_WITHSIZE(iov[1], __UNCONST(data), dlen); + do { putwait(spc, &rw, &rhdr); - if ((rv = send_with_recon(spc, &rhdr, sizeof(rhdr))) != 0) { - unputwait(spc, &rw); - continue; - } - if ((rv = send_with_recon(spc, data, dlen)) != 0) { + if ((rv = send_with_recon(spc, iov, __arraycount(iov))) != 0) { unputwait(spc, &rw); continue; } - rv = cliwaitresp(spc, &rw, &omask, false); - } while (rv == ENOTCONN || rv == EAGAIN); - pthread_sigmask(SIG_SETMASK, &omask, NULL); + rv = cliwaitresp(spc, &rw, omask, false); + if (rv == ENOTCONN) + rv = EAGAIN; + } while (rv == EAGAIN); *resp = rw.rw_data; return rv; } static int -handshake_req(struct spclient *spc, uint32_t *auth, int cancel, bool haslock) +handshake_req(struct spclient *spc, int type, void *data, + int cancel, bool haslock) { struct handshake_fork rf; + const char *myprogname = NULL; /* XXXgcc */ struct rsp_hdr rhdr; struct respwait rw; sigset_t omask; + size_t bonus; + struct iovec iov[2]; int rv; + if (type == HANDSHAKE_FORK) { + bonus = sizeof(rf); + } else { +#ifdef __NetBSD__ + /* would procfs work on NetBSD too? */ + myprogname = getprogname(); +#else + int fd = open("/proc/self/comm", O_RDONLY); + if (fd == -1) { + myprogname = "???"; + } else { + static char commname[128]; + + memset(commname, 0, sizeof(commname)); + if (read(fd, commname, sizeof(commname)) > 0) { + char *n; + + n = strrchr(commname, '\n'); + if (n) + *n = '\0'; + myprogname = commname; + } else { + myprogname = "???"; + } + close(fd); + } +#endif + bonus = strlen(myprogname)+1; + } + /* performs server handshake */ - rhdr.rsp_len = sizeof(rhdr) + (auth ? sizeof(rf) : 0); + rhdr.rsp_len = sizeof(rhdr) + bonus; rhdr.rsp_class = RUMPSP_REQ; rhdr.rsp_type = RUMPSP_HANDSHAKE; - if (auth) - rhdr.rsp_handshake = HANDSHAKE_FORK; - else - rhdr.rsp_handshake = HANDSHAKE_GUEST; + rhdr.rsp_handshake = type; + + IOVPUT(iov[0], rhdr); pthread_sigmask(SIG_SETMASK, &fullset, &omask); if (haslock) putwait_locked(spc, &rw, &rhdr); else putwait(spc, &rw, &rhdr); - rv = dosend(spc, &rhdr, sizeof(rhdr)); - if (auth) { - memcpy(rf.rf_auth, auth, AUTHLEN*sizeof(*auth)); + if (type == HANDSHAKE_FORK) { + memcpy(rf.rf_auth, data, sizeof(rf.rf_auth)); /* uh, why? */ rf.rf_cancel = cancel; - rv = send_with_recon(spc, &rf, sizeof(rf)); + IOVPUT(iov[1], rf); + } else { + IOVPUT_WITHSIZE(iov[1], __UNCONST(myprogname), bonus); } + rv = send_with_recon(spc, iov, __arraycount(iov)); if (rv || cancel) { if (haslock) unputwait_locked(spc, &rw); else unputwait(spc, &rw); if (cancel) { - pthread_sigmask(SIG_SETMASK, &omask, NULL); - return rv; + goto out; } } else { rv = cliwaitresp(spc, &rw, &omask, haslock); } - pthread_sigmask(SIG_SETMASK, &omask, NULL); if (rv) - return rv; + goto out; rv = *(int *)rw.rw_data; free(rw.rw_data); + out: + pthread_sigmask(SIG_SETMASK, &omask, NULL); return rv; } static int -prefork_req(struct spclient *spc, void **resp) +prefork_req(struct spclient *spc, sigset_t *omask, void **resp) { struct rsp_hdr rhdr; struct respwait rw; - sigset_t omask; + struct iovec iov[1]; int rv; rhdr.rsp_len = sizeof(rhdr); @@ -306,18 +463,20 @@ prefork_req(struct spclient *spc, void * rhdr.rsp_type = RUMPSP_PREFORK; rhdr.rsp_error = 0; - pthread_sigmask(SIG_SETMASK, &fullset, &omask); + IOVPUT(iov[0], rhdr); + do { putwait(spc, &rw, &rhdr); - rv = send_with_recon(spc, &rhdr, sizeof(rhdr)); + rv = send_with_recon(spc, iov, __arraycount(iov)); if (rv != 0) { unputwait(spc, &rw); continue; } - rv = cliwaitresp(spc, &rw, &omask, false); - } while (rv == ENOTCONN || rv == EAGAIN); - pthread_sigmask(SIG_SETMASK, &omask, NULL); + rv = cliwaitresp(spc, &rw, omask, false); + if (rv == ENOTCONN) + rv = EAGAIN; + } while (rv == EAGAIN); *resp = rw.rw_data; return rv; @@ -352,6 +511,7 @@ send_copyin_resp(struct spclient *spc, u int wantstr) { struct rsp_hdr rhdr; + struct iovec iov[2]; if (wantstr) dlen = MIN(dlen, strlen(data)+1); @@ -362,10 +522,12 @@ send_copyin_resp(struct spclient *spc, u rhdr.rsp_type = RUMPSP_COPYIN; rhdr.rsp_sysnum = 0; + IOVPUT(iov[0], rhdr); + IOVPUT_WITHSIZE(iov[1], data, dlen); + if (resp_sendlock(spc) != 0) return; - (void)dosend(spc, &rhdr, sizeof(rhdr)); - (void)dosend(spc, data, dlen); + (void)SENDIOV(spc, iov); sendunlock(spc); } @@ -373,6 +535,7 @@ static void send_anonmmap_resp(struct spclient *spc, uint64_t reqno, void *addr) { struct rsp_hdr rhdr; + struct iovec iov[2]; rhdr.rsp_len = sizeof(rhdr) + sizeof(addr); rhdr.rsp_reqno = reqno; @@ -380,10 +543,12 @@ send_anonmmap_resp(struct spclient *spc, rhdr.rsp_type = RUMPSP_ANONMMAP; rhdr.rsp_sysnum = 0; + IOVPUT(iov[0], rhdr); + IOVPUT(iov[1], addr); + if (resp_sendlock(spc) != 0) return; - (void)dosend(spc, &rhdr, sizeof(rhdr)); - (void)dosend(spc, &addr, sizeof(addr)); + (void)SENDIOV(spc, iov); sendunlock(spc); } @@ -392,15 +557,18 @@ rumpclient_syscall(int sysnum, const voi register_t *retval) { struct rsp_sysresp *resp; + sigset_t omask; void *rdata; int rv; + pthread_sigmask(SIG_SETMASK, &fullset, &omask); + DPRINTF(("rumpsp syscall_req: syscall %d with %p/%zu\n", sysnum, data, dlen)); - rv = syscall_req(&clispc, sysnum, data, dlen, &rdata); + rv = syscall_req(&clispc, &omask, sysnum, data, dlen, &rdata); if (rv) - return rv; + goto out; resp = rdata; DPRINTF(("rumpsp syscall_resp: syscall %d error %d, rv: %d/%d\n", @@ -410,6 +578,8 @@ rumpclient_syscall(int sysnum, const voi rv = resp->rsys_error; free(rdata); + out: + pthread_sigmask(SIG_SETMASK, &omask, NULL); return rv; } @@ -473,27 +643,52 @@ handlereq(struct spclient *spc) static unsigned ptab_idx; static struct sockaddr *serv_sa; +/* dup until we get a "good" fd which does not collide with stdio */ static int -doconnect(int retry) +dupgood(int myfd, int mustchange) +{ + int ofds[4]; + int sverrno; + unsigned int i; + + for (i = 0; (myfd <= 2 || mustchange) && myfd != -1; i++) { + assert(i < __arraycount(ofds)); + ofds[i] = myfd; + myfd = host_dup(myfd); + if (mustchange) { + i--; /* prevent closing old fd */ + mustchange = 0; + } + } + + sverrno = 0; + if (myfd == -1 && i > 0) + sverrno = errno; + + while (i-- > 0) { + host_close(ofds[i]); + } + + if (sverrno) + errno = sverrno; + + return myfd; +} + +static int +doconnect(void) { - time_t prevreconmsg; - unsigned reconretries; struct respwait rw; struct rsp_hdr rhdr; - struct kevent kev[NSIG+1]; char banner[MAXBANNER]; - struct pollfd pfd; - int s, error, flags, i; + int s, error, flags; ssize_t n; if (kq != -1) host_close(kq); kq = -1; + s = -1; - prevreconmsg = 0; - reconretries = 0; - - again: if (clispc.spc_fd != -1) host_close(clispc.spc_fd); clispc.spc_fd = -1; @@ -525,85 +720,52 @@ doconnect(int retry) free(clispc.spc_buf); clispc.spc_off = 0; - s = host_socket(parsetab[ptab_idx].domain, SOCK_STREAM, 0); + s = dupgood(host_socket(parsetab[ptab_idx].domain, SOCK_STREAM, 0), 0); if (s == -1) return -1; - pfd.fd = s; - pfd.events = POLLIN; - while (host_connect(s, serv_sa, (socklen_t)serv_sa->sa_len) == -1) { + while (host_connect(s, serv_sa, parsetab[ptab_idx].slen) == -1) { if (errno == EINTR) continue; - if (!retry) { - error = errno; - fprintf(stderr, "rump_sp: client connect failed: %s\n", - strerror(errno)); - errno = error; - return -1; - } - - if (prevreconmsg == 0) { - fprintf(stderr, "rump_sp: connection to kernel lost, " - "trying to reconnect ...\n"); - prevreconmsg = time(NULL); - } - if (time(NULL) - prevreconmsg > 120) { - fprintf(stderr, "rump_sp: still trying to " - "reconnect ...\n"); - prevreconmsg = time(NULL); - } - - /* adhoc backoff timer */ - if (reconretries++ < 10) { - usleep(100000 * reconretries); - } else { - sleep(MIN(10, reconretries-9)); - } - goto again; + ERRLOG(("rump_sp: client connect failed: %s\n", + strerror(errno))); + return -1; } if ((error = parsetab[ptab_idx].connhook(s)) != 0) { - error = errno; - fprintf(stderr, "rump_sp: connect hook failed\n"); - errno = error; + ERRLOG(("rump_sp: connect hook failed\n")); return -1; } - if ((n = host_read(s, banner, sizeof(banner)-1)) < 0) { - error = errno; - fprintf(stderr, "rump_sp: failed to read banner\n"); - errno = error; + if ((n = host_read(s, banner, sizeof(banner)-1)) <= 0) { + ERRLOG(("rump_sp: failed to read banner\n")); return -1; } if (banner[n-1] != '\n') { - fprintf(stderr, "rump_sp: invalid banner\n"); - errno = EINVAL; + ERRLOG(("rump_sp: invalid banner\n")); return -1; } banner[n] = '\0'; - /* parse the banner some day */ + /* XXX parse the banner some day */ flags = host_fcntl(s, F_GETFL, 0); if (host_fcntl(s, F_SETFL, flags | O_NONBLOCK) == -1) { - fprintf(stderr, "rump_sp: socket fd NONBLOCK: %s\n", - strerror(errno)); - errno = EINVAL; + ERRLOG(("rump_sp: socket fd NONBLOCK: %s\n", strerror(errno))); return -1; } clispc.spc_fd = s; clispc.spc_state = SPCSTATE_RUNNING; clispc.spc_reconnecting = 0; - if (prevreconmsg) { - fprintf(stderr, "rump_sp: reconnected!\n"); - } +#ifdef USE_KQUEUE +{ + struct kevent kev[NSIG+1]; + int i; /* setup kqueue, we want all signals and the fd */ - if ((kq = host_kqueue()) == -1) { - error = errno; - fprintf(stderr, "rump_sp: cannot setup kqueue"); - errno = error; + if ((kq = dupgood(host_kqueue(), 0)) == -1) { + ERRLOG(("rump_sp: cannot setup kqueue")); return -1; } @@ -613,11 +775,11 @@ doconnect(int retry) EV_SET(&kev[NSIG], clispc.spc_fd, EVFILT_READ, EV_ADD|EV_ENABLE, 0, 0, 0); if (host_kevent(kq, kev, NSIG+1, NULL, 0, NULL) == -1) { - error = errno; - fprintf(stderr, "rump_sp: kevent() failed"); - errno = error; + ERRLOG(("rump_sp: kevent() failed")); return -1; } +} +#endif /* USE_KQUEUE */ return 0; } @@ -633,93 +795,163 @@ doinit(void) return 0; } -void *(*rumpclient_dlsym)(void *, const char *); +void *rumpclient__dlsym(void *, const char *); +void * +rumpclient__dlsym(void *handle, const char *symbol) +{ + + return dlsym(handle, symbol); +} +void *rumphijack_dlsym(void *, const char *) + __attribute__((__weak__, alias("rumpclient__dlsym"))); + +static pid_t init_done = 0; int -rumpclient_init() +rumpclient_init(void) { char *p; int error; + int rv = -1; + int hstype; + pid_t mypid; - /* dlsym overrided by rumphijack? */ - if (!rumpclient_dlsym) - rumpclient_dlsym = dlsym; + /* + * Make sure we're not riding the context of a previous + * host fork. Note: it's *possible* that after n>1 forks + * we have the same pid as one of our exited parents, but + * I'm pretty sure there are 0 practical implications, since + * it means generations would have to skip rumpclient init. + */ + if (init_done == (mypid = getpid())) + return 0; + + /* kq does not traverse fork() */ + if (init_done != 0) + kq = -1; + init_done = mypid; + + sigfillset(&fullset); /* - * sag mir, wo die symbol sind. zogen fort, der krieg beginnt. + * sag mir, wo die symbols sind. zogen fort, der krieg beginnt. * wann wird man je verstehen? wann wird man je verstehen? */ #define FINDSYM2(_name_,_syscall_) \ - if ((host_##_name_ = rumpclient_dlsym(RTLD_NEXT, \ - #_syscall_)) == NULL) \ - /* host_##_name_ = _syscall_ */; + if ((host_##_name_ = rumphijack_dlsym(RTLD_NEXT, \ + #_syscall_)) == NULL) { \ + if (rumphijack_dlsym == rumpclient__dlsym) \ + host_##_name_ = _name_; /* static fallback */ \ + if (host_##_name_ == NULL) \ + errx(1, "cannot find %s: %s", #_syscall_, \ + dlerror()); \ + } #define FINDSYM(_name_) FINDSYM2(_name_,_name_) - FINDSYM2(socket,__socket30); - FINDSYM(close); - FINDSYM(connect); - FINDSYM(fcntl); - FINDSYM(poll); - FINDSYM(read); - FINDSYM(sendto); - FINDSYM(setsockopt); - FINDSYM(kqueue); - FINDSYM(kevent); +#ifdef __NetBSD__ + FINDSYM2(socket,__socket30) +#else + FINDSYM(socket) +#endif + + FINDSYM(close) + FINDSYM(connect) + FINDSYM(fcntl) + FINDSYM(poll) + FINDSYM(read) + FINDSYM(sendmsg) + FINDSYM(setsockopt) + FINDSYM(dup) + FINDSYM(execve) + +#ifdef USE_KQUEUE + FINDSYM(kqueue) +#if !__NetBSD_Prereq__(5,99,7) + FINDSYM(kevent) +#else + FINDSYM2(kevent,_sys___kevent50) +#endif +#endif /* USE_KQUEUE */ + #undef FINDSYM #undef FINDSY2 - if ((p = getenv("RUMP_SERVER")) == NULL) { - errno = ENOENT; - return -1; + if ((p = getenv("RUMP__PARSEDSERVER")) == NULL) { + if ((p = getenv("RUMP_SERVER")) == NULL) { + fprintf(stderr, "error: RUMP_SERVER not set\n"); + errno = ENOENT; + goto out; + } } if ((error = parseurl(p, &serv_sa, &ptab_idx, 0)) != 0) { errno = error; - return -1; + goto out; } if (doinit() == -1) - return -1; - if (doconnect(0) == -1) - return -1; + goto out; - error = handshake_req(&clispc, NULL, 0, false); + if ((p = getenv("RUMPCLIENT__EXECFD")) != NULL) { + sscanf(p, "%d,%d", &clispc.spc_fd, &kq); + unsetenv("RUMPCLIENT__EXECFD"); + hstype = HANDSHAKE_EXEC; + } else { + if (doconnect() == -1) + goto out; + hstype = HANDSHAKE_GUEST; + } + + error = handshake_req(&clispc, hstype, NULL, 0, false); if (error) { pthread_mutex_destroy(&clispc.spc_mtx); pthread_cond_destroy(&clispc.spc_cv); if (clispc.spc_fd != -1) host_close(clispc.spc_fd); errno = error; - return -1; + goto out; } + rv = 0; - sigfillset(&fullset); - return 0; + out: + if (rv == -1) + init_done = 0; + return rv; } struct rumpclient_fork { uint32_t fork_auth[AUTHLEN]; + struct spclient fork_spc; + int fork_kq; }; struct rumpclient_fork * rumpclient_prefork(void) { struct rumpclient_fork *rpf; + sigset_t omask; void *resp; int rv; + pthread_sigmask(SIG_SETMASK, &fullset, &omask); rpf = malloc(sizeof(*rpf)); if (rpf == NULL) - return NULL; + goto out; - if ((rv = prefork_req(&clispc, &resp)) != 0) { + if ((rv = prefork_req(&clispc, &omask, &resp)) != 0) { free(rpf); errno = rv; - return NULL; + rpf = NULL; + goto out; } memcpy(rpf->fork_auth, resp, sizeof(rpf->fork_auth)); free(resp); + rpf->fork_spc = clispc; + rpf->fork_kq = kq; + + out: + pthread_sigmask(SIG_SETMASK, &omask, NULL); return rpf; } @@ -727,17 +959,21 @@ int rumpclient_fork_init(struct rumpclient_fork *rpf) { int error; + int osock; + osock = clispc.spc_fd; memset(&clispc, 0, sizeof(clispc)); - clispc.spc_fd = -1; - kq = -1; + clispc.spc_fd = osock; + + kq = -1; /* kqueue descriptor is not copied over fork() */ if (doinit() == -1) return -1; - if (doconnect(1) == -1) + if (doconnect() == -1) return -1; - error = handshake_req(&clispc, rpf->fork_auth, 0, false); + error = handshake_req(&clispc, HANDSHAKE_FORK, rpf->fork_auth, + 0, false); if (error) { pthread_mutex_destroy(&clispc.spc_mtx); pthread_cond_destroy(&clispc.spc_cv); @@ -747,3 +983,185 @@ rumpclient_fork_init(struct rumpclient_f return 0; } + +/*ARGSUSED*/ +void +rumpclient_fork_cancel(struct rumpclient_fork *rpf) +{ + + /* EUNIMPL */ +} + +void +rumpclient_fork_vparent(struct rumpclient_fork *rpf) +{ + + clispc = rpf->fork_spc; + kq = rpf->fork_kq; +} + +void +rumpclient_setconnretry(time_t timeout) +{ + + if (timeout < RUMPCLIENT_RETRYCONN_DIE) + return; /* gigo */ + + retrytimo = timeout; +} + +int +rumpclient__closenotify(int *fdp, enum rumpclient_closevariant variant) +{ + int fd = *fdp; + int untilfd, rv; + int newfd; + + switch (variant) { + case RUMPCLIENT_CLOSE_FCLOSEM: + untilfd = MAX(clispc.spc_fd, kq); + for (; fd <= untilfd; fd++) { + if (fd == clispc.spc_fd || fd == kq) + continue; + rv = host_close(fd); + if (rv == -1) + return -1; + } + *fdp = fd; + break; + + case RUMPCLIENT_CLOSE_CLOSE: + case RUMPCLIENT_CLOSE_DUP2: + if (fd == clispc.spc_fd) { + newfd = dupgood(clispc.spc_fd, 1); + if (newfd == -1) + return -1; + +#ifdef USE_KQUEUE + { + struct kevent kev[2]; + + /* + * now, we have a new socket number, so change + * the file descriptor that kqueue is + * monitoring. remove old and add new. + */ + EV_SET(&kev[0], clispc.spc_fd, + EVFILT_READ, EV_DELETE, 0, 0, 0); + EV_SET(&kev[1], newfd, + EVFILT_READ, EV_ADD|EV_ENABLE, 0, 0, 0); + if (host_kevent(kq, kev, 2, NULL, 0, NULL) == -1) { + int sverrno = errno; + host_close(newfd); + errno = sverrno; + return -1; + } + clispc.spc_fd = newfd; + } + } + if (fd == kq) { + newfd = dupgood(kq, 1); + if (newfd == -1) + return -1; + kq = newfd; +#else /* USE_KQUEUE */ + clispc.spc_fd = newfd; +#endif /* !USE_KQUEUE */ + } + break; + } + + return 0; +} + +pid_t +rumpclient_fork(void) +{ + + return rumpclient__dofork(fork); +} + +/* + * Process is about to exec. Save info about our existing connection + * in the env. rumpclient will check for this info in init(). + * This is mostly for the benefit of rumphijack, but regular applications + * may use it as well. + */ +int +rumpclient_exec(const char *path, char *const argv[], char *const envp[]) +{ + char buf[4096]; + char **newenv; + char *envstr, *envstr2; + size_t nelem; + int rv, sverrno; + + snprintf(buf, sizeof(buf), "RUMPCLIENT__EXECFD=%d,%d", + clispc.spc_fd, kq); + envstr = malloc(strlen(buf)+1); + if (envstr == NULL) { + return ENOMEM; + } + strcpy(envstr, buf); + + /* do we have a fully parsed url we want to forward in the env? */ + if (*parsedurl != '\0') { + snprintf(buf, sizeof(buf), + "RUMP__PARSEDSERVER=%s", parsedurl); + envstr2 = malloc(strlen(buf)+1); + if (envstr2 == NULL) { + free(envstr); + return ENOMEM; + } + strcpy(envstr2, buf); + } else { + envstr2 = NULL; + } + + for (nelem = 0; envp && envp[nelem]; nelem++) + continue; + + newenv = malloc(sizeof(*newenv) * (nelem+3)); + if (newenv == NULL) { + free(envstr2); + free(envstr); + return ENOMEM; + } + memcpy(&newenv[0], envp, nelem*sizeof(*envp)); + + newenv[nelem] = envstr; + newenv[nelem+1] = envstr2; + newenv[nelem+2] = NULL; + + rv = host_execve(path, argv, newenv); + + _DIAGASSERT(rv != 0); + sverrno = errno; + free(envstr2); + free(envstr); + free(newenv); + errno = sverrno; + return rv; +} + +int +rumpclient_daemon(int nochdir, int noclose) +{ + struct rumpclient_fork *rf; + int sverrno; + + if ((rf = rumpclient_prefork()) == NULL) + return -1; + + if (daemon(nochdir, noclose) == -1) { + sverrno = errno; + rumpclient_fork_cancel(rf); + errno = sverrno; + return -1; + } + + if (rumpclient_fork_init(rf) == -1) + return -1; + + return 0; +}