diff options
Diffstat (limited to 'erts/emulator/drivers/common/inet_drv.c')
| -rw-r--r-- | erts/emulator/drivers/common/inet_drv.c | 159 |
1 files changed, 138 insertions, 21 deletions
diff --git a/erts/emulator/drivers/common/inet_drv.c b/erts/emulator/drivers/common/inet_drv.c index 627e1c8ec2..47eb5df7dd 100644 --- a/erts/emulator/drivers/common/inet_drv.c +++ b/erts/emulator/drivers/common/inet_drv.c @@ -813,6 +813,7 @@ static size_t my_strnlen(const char *s, size_t maxlen) #define INET_OPT_PKTOPTIONS 45 /* IP(V6)_PKTOPTIONS get ancillary data */ #define INET_OPT_TTL 46 /* IP_TTL */ #define INET_OPT_RECVTTL 47 /* IP_RECVTTL ancillary data */ +#define TCP_OPT_NOPUSH 48 /* super-Nagle, aka TCP_CORK */ /* SCTP options: a separate range, from 100: */ #define SCTP_OPT_RTOINFO 100 #define SCTP_OPT_ASSOCINFO 101 @@ -957,6 +958,12 @@ static size_t my_strnlen(const char *s, size_t maxlen) typedef struct _tcp_descriptor tcp_descriptor; +#if defined(TCP_CORK) +#define INET_TCP_NOPUSH TCP_CORK +#elif defined(TCP_NOPUSH) && !defined(__DARWIN__) +#define INET_TCP_NOPUSH TCP_NOPUSH +#endif + #define BIN_REALLOC_MARGIN(x) ((x)/4) /* 25% */ /* The general purpose sockaddr */ @@ -1319,7 +1326,10 @@ static int tcp_deliver(tcp_descriptor* desc, int len); static int tcp_shutdown_error(tcp_descriptor* desc, int err); +#ifdef HAVE_SENDFILE static int tcp_inet_sendfile(tcp_descriptor* desc); +static int tcp_sendfile_aborted(tcp_descriptor* desc, int socket_error); +#endif static int tcp_inet_output(tcp_descriptor* desc, HANDLE event); static int tcp_inet_input(tcp_descriptor* desc, HANDLE event); @@ -5184,6 +5194,71 @@ static int hwaddr_libdlpi_lookup(const char *ifnm, } #endif +#ifdef HAVE_GETIFADDRS +/* Returns 0 for success and errno() for failure */ +static int call_getifaddrs(inet_descriptor* desc_p, struct ifaddrs **ifa_pp) +{ + int result, save_errno; +#ifdef HAVE_SETNS + int current_ns; + + current_ns = 0; + if (desc_p->netns != NULL) { + int new_ns; + /* Temporarily change network namespace for this thread + * over the getifaddrs() call + */ + current_ns = open("/proc/self/ns/net", O_RDONLY); + if (current_ns == INVALID_SOCKET) + return sock_errno(); + new_ns = open(desc_p->netns, O_RDONLY); + if (new_ns == INVALID_SOCKET) { + save_errno = sock_errno(); + while (close(current_ns) == INVALID_SOCKET && + sock_errno() == EINTR); + return save_errno; + } + if (setns(new_ns, CLONE_NEWNET) != 0) { + save_errno = sock_errno(); + while (close(new_ns) == INVALID_SOCKET && + sock_errno() == EINTR); + while (close(current_ns) == INVALID_SOCKET && + sock_errno() == EINTR); + return save_errno; + } + else { + while (close(new_ns) == INVALID_SOCKET && + sock_errno() == EINTR); + } + } +#endif + save_errno = 0; + result = getifaddrs(ifa_pp); + if (result < 0) + save_errno = sock_errno(); +#ifdef HAVE_SETNS + if (desc_p->netns != NULL) { + /* Restore network namespace */ + if (setns(current_ns, CLONE_NEWNET) != 0) { + /* XXX Failed to restore network namespace. + * What to do? Tidy up and return an error... + * Note that the thread now might still be in the set namespace. + * Can this even happen? Should the emulator be aborted? + */ + if (result >= 0) { + /* We got a result but have to waste it */ + save_errno = sock_errno(); + freeifaddrs(*ifa_pp); + } + } + while (close(current_ns) == INVALID_SOCKET && + sock_errno() == EINTR); + } +#endif + return save_errno; +} +#endif /* #ifdef HAVE_GETIFADDRS */ + /* FIXME: temporary hack */ #ifndef IFHWADDRLEN #define IFHWADDRLEN 6 @@ -5261,8 +5336,8 @@ static ErlDrvSSizeT inet_ctl_ifget(inet_descriptor* desc, struct sockaddr_dl *sdlp; int found = 0; - if (getifaddrs(&ifa) == -1) - goto error; + if (call_getifaddrs(desc, &ifa) != 0) + goto error; for (ifp = ifa; ifp; ifp = ifp->ifa_next) { if ((ifp->ifa_addr->sa_family == AF_LINK) && @@ -5980,6 +6055,7 @@ static ErlDrvSSizeT inet_ctl_getifaddrs(inet_descriptor* desc_p, ErlDrvSizeT buf_size; char *buf_p; char *buf_alloc_p; + int save_errno; buf_size = GETIFADDRS_BUFSZ; buf_alloc_p = ALLOC(GETIFADDRS_BUFSZ); @@ -6014,9 +6090,9 @@ static ErlDrvSSizeT inet_ctl_getifaddrs(inet_descriptor* desc_p, } \ } while (0) - if (getifaddrs(&ifa_p) < 0) { - return ctl_error(sock_errno(), rbuf_pp, rsize); - } + if ((save_errno = call_getifaddrs(desc_p, &ifa_p)) != 0) + return ctl_error(save_errno, rbuf_pp, rsize); + ifa_free_p = ifa_p; *buf_p++ = INET_REP_OK; for (; ifa_p; ifa_p = ifa_p->ifa_next) { @@ -6538,6 +6614,19 @@ static int inet_set_opts(inet_descriptor* desc, char* ptr, int len) (long)desc->port, desc->s, ival)); break; + case TCP_OPT_NOPUSH: +#if defined(INET_TCP_NOPUSH) + proto = IPPROTO_TCP; + type = INET_TCP_NOPUSH; + DEBUGF(("inet_set_opts(%ld): s=%d, t=%d TCP_NOPUSH=%d\r\n", + (long)desc->port, desc->s, type, ival)); + break; +#else + /* inet_fill_opts always returns a value for this option, + * so we need to ignore it if not implemented, just in case */ + continue; +#endif + #if defined(HAVE_MULTICAST_SUPPORT) && defined(IPPROTO_IP) case UDP_OPT_MULTICAST_TTL: @@ -7699,6 +7788,16 @@ static ErlDrvSSizeT inet_fill_opts(inet_descriptor* desc, proto = IPPROTO_TCP; type = TCP_NODELAY; break; + case TCP_OPT_NOPUSH: +#if defined(INET_TCP_NOPUSH) + proto = IPPROTO_TCP; + type = INET_TCP_NOPUSH; + break; +#else + *ptr++ = opt; + put_int32(0, ptr); + continue; +#endif #if defined(HAVE_MULTICAST_SUPPORT) && defined(IPPROTO_IP) case UDP_OPT_MULTICAST_TTL: @@ -10220,12 +10319,11 @@ static ErlDrvSSizeT tcp_inet_ctl(ErlDrvData e, unsigned int cmd, desc->tcp_add_flags |= TCP_ADDF_SENDFILE; /* See if we can finish sending without selecting & rescheduling. */ - tcp_inet_sendfile(desc); - - if(desc->sendfile.length > 0) { - sock_select(INETP(desc), FD_WRITE, 1); + if (tcp_inet_sendfile(desc) == 0) { + if(desc->sendfile.length > 0) { + sock_select(INETP(desc), FD_WRITE, 1); + } } - return ctl_reply(INET_REP_OK, NULL, 0, rbuf, rsize); #else return ctl_error(ENOTSUP, rbuf, rsize); @@ -10436,6 +10534,7 @@ static int tcp_recv_closed(tcp_descriptor* desc) #ifdef DEBUG long port = (long) desc->inet.port; /* Used after driver_exit() */ #endif + int blocking_send = 0; DEBUGF(("tcp_recv_closed(%ld): s=%d, in %s, line %d\r\n", port, desc->inet.s, __FILE__, __LINE__)); if (IS_BUSY(INETP(desc))) { @@ -10451,7 +10550,15 @@ static int tcp_recv_closed(tcp_descriptor* desc) set_busy_port(desc->inet.port, 0); inet_reply_error_am(INETP(desc), am_closed); DEBUGF(("tcp_recv_closed(%ld): busy reply 'closed'\r\n", port)); - } else { + blocking_send = 1; + } +#ifdef HAVE_SENDFILE + if (desc->tcp_add_flags & TCP_ADDF_SENDFILE) { + tcp_sendfile_aborted(desc, ENOTCONN); + blocking_send = 1; + } +#endif + if (!blocking_send) { /* No blocking send op to reply to right now. * If next op is a send, make sure it returns {error,closed} * rather than {error,enotconn}. @@ -10503,6 +10610,11 @@ static int tcp_recv_error(tcp_descriptor* desc, int err) set_busy_port(desc->inet.port, 0); inet_reply_error_am(INETP(desc), am_closed); } +#ifdef HAVE_SENDFILE + if (desc->tcp_add_flags & TCP_ADDF_SENDFILE) { + tcp_sendfile_aborted(desc, err); + } +#endif if (!desc->inet.active) { /* We must cancel any timer here ! */ clean_multi_timers(desc, INETP(desc)->port); @@ -11165,27 +11277,31 @@ static int tcp_send_or_shutdown_error(tcp_descriptor* desc, int err) DEBUGF(("driver_failure_eof(%ld) in %s, line %d\r\n", (long)desc->inet.port, __FILE__, __LINE__)); if (desc->inet.active) { + ErlDrvTermData err_atom; if (show_econnreset) { tcp_error_message(desc, err); - tcp_closed_message(desc); - inet_reply_error(INETP(desc), err); + err_atom = error_atom(err); } else { - tcp_closed_message(desc); - inet_reply_error_am(INETP(desc), am_closed); + err_atom = am_closed; } + tcp_closed_message(desc); + if (!(desc->tcp_add_flags & TCP_ADDF_SENDFILE)) + inet_reply_error_am(INETP(desc), err_atom); + if (desc->inet.exitf) driver_exit(desc->inet.port, 0); else tcp_desc_close(desc); } else { tcp_close_check(desc); - tcp_desc_close(desc); if (desc->inet.caller) { - if (show_econnreset) - inet_reply_error(INETP(desc), err); - else - inet_reply_error_am(INETP(desc), am_closed); + if (!(desc->tcp_add_flags & TCP_ADDF_SENDFILE)) { + if (show_econnreset) + inet_reply_error(INETP(desc), err); + else + inet_reply_error_am(INETP(desc), am_closed); + } } else { /* No blocking send op to reply to right now. @@ -11194,6 +11310,7 @@ static int tcp_send_or_shutdown_error(tcp_descriptor* desc, int err) */ desc->tcp_add_flags |= TCP_ADDF_DELAYED_CLOSE_SEND; } + tcp_desc_close(desc); /* * Make sure that the next receive operation gets an {error,closed} @@ -11720,8 +11837,8 @@ socket_error: { DEBUGF(("tcp_inet_sendfile(%ld): send errno = %d (errno %d)\r\n", (long)desc->inet.port, socket_errno, errno)); - result = tcp_send_error(desc, socket_errno); tcp_sendfile_aborted(desc, socket_errno); + result = tcp_send_error(desc, socket_errno); goto done; } |