/*! \brief releases expired connections and cleans up bad ones (state<0) */
static inline void tcp_receive_timeout(void)
{
struct tcp_connection* con;
struct tcp_connection* next;
unsigned int ticks;
ticks=get_ticks();
for (con=tcp_conn_lst; con; con=next) {
next=con->c_next; /* safe for removing */
if (con->state<0){ /* kill bad connections */
/* S_CONN_BAD or S_CONN_ERROR, remove it */
/* fd will be closed in release_tcpconn */
io_watch_del(&io_w, con->fd, -1, IO_FD_CLOSING);
tcpconn_listrm(tcp_conn_lst, con, c_next, c_prev);
con->state=S_CONN_BAD;
release_tcpconn(con, CONN_ERROR, tcpmain_sock);
continue;
}
if (con->timeout<=ticks){
LM_DBG("%p expired - (%d, %d) lt=%d\n",
con, con->timeout, ticks,con->lifetime);
/* fd will be closed in release_tcpconn */
io_watch_del(&io_w, con->fd, -1, IO_FD_CLOSING);
tcpconn_listrm(tcp_conn_lst, con, c_next, c_prev);
if (con->msg_attempts)
release_tcpconn(con, CONN_ERROR, tcpmain_sock);
else
release_tcpconn(con, CONN_RELEASE, tcpmain_sock);
}
}
}
/*! \brief
* handles an io event on one of the watched tcp connections
*
* \param tcpconn - pointer to the tcp_connection for which we have an io ev.
* \param fd_i - index in the fd_array table (needed for delete)
* \return handle_* return convention, but on success it always returns 0
* (because it's one-shot, after a succesfull execution the fd is
* removed from tcp_main's watch fd list and passed to a child =>
* tcp_main is not interested in further io events that might be
* queued for this fd)
*/
inline static int handle_tcpconn_ev(struct tcp_connection* tcpconn, int fd_i)
{
int fd;
/* is refcnt!=0 really necessary?
* No, in fact it's a bug: I can have the following situation: a send only
* tcp connection used by n processes simultaneously => refcnt = n. In
* the same time I can have a read event and this situation is perfectly
* valid. -- andrei
*/
#if 0
if ((tcpconn->refcnt!=0)){
/* FIXME: might be valid for sigio_rt iff fd flags are not cleared
* (there is a short window in which it could generate a sig
* that would be catched by tcp_main) */
LM_CRIT("io event on referenced tcpconn (%p), refcnt=%d, fd=%d\n",
tcpconn, tcpconn->refcnt, tcpconn->s);
return -1;
}
#endif
/* pass it to child, so remove it from the io watch list */
LM_DBG("data available on %p %d\n", tcpconn, tcpconn->s);
if (io_watch_del(&io_h, tcpconn->s, fd_i, 0)==-1) goto error;
tcpconn->flags|=F_CONN_REMOVED;
tcpconn_ref(tcpconn); /* refcnt ++ */
if (send2child(tcpconn)<0){
LM_ERR("no children available\n");
TCPCONN_LOCK;
tcpconn->refcnt--;
if (tcpconn->refcnt==0){
fd=tcpconn->s;
_tcpconn_rm(tcpconn);
close(fd);
}else tcpconn->timeout=0; /* force expire*/
TCPCONN_UNLOCK;
}
return 0; /* we are not interested in possibly queued io events,
the fd was either passed to a child, or closed */
error:
return -1;
}
/*! \brief very inefficient for now - FIXME
* keep in sync with tcpconn_destroy, the "delete" part should be
* the same except for io_watch_del..
* \todo FIXME (very inefficient for now)
*/
static inline void tcpconn_timeout(int force)
{
struct tcp_connection *c, *next;
unsigned int ticks;
unsigned h;
int fd;
ticks=get_ticks();
TCPCONN_LOCK; /* fixme: we can lock only on delete IMO */
for(h=0; h<TCP_ID_HASH_SIZE; h++){
c=tcpconn_id_hash[h];
while(c){
next=c->id_next;
if (force ||((c->refcnt==0) && (ticks>c->timeout))) {
if (!force)
LM_DBG("timeout for hash=%d - %p"
" (%d > %d)\n", h, c, ticks, c->timeout);
fd=c->s;
#ifdef USE_TLS
if (c->type==PROTO_TLS)
tls_close(c, fd);
#endif
_tcpconn_rm(c);
if ((!force)&&(fd>0)&&(c->refcnt==0)) {
if (!(c->flags & F_CONN_REMOVED)){
io_watch_del(&io_h, fd, -1, IO_FD_CLOSING);
c->flags|=F_CONN_REMOVED;
}
close(fd);
}
tcp_connections_no--;
}
c=next;
}
}
TCPCONN_UNLOCK;
}
static ticks_t tcpconn_read_timeout(ticks_t t, struct timer_ln* tl, void* data)
{
struct tcp_connection *c;
c=(struct tcp_connection*)data;
/* or (struct tcp...*)(tl-offset(c->timer)) */
if (likely(!(c->state<0) && TICKS_LT(t, c->timeout))){
/* timeout extended, exit */
return (ticks_t)(c->timeout - t);
}
/* if conn->state is ERROR or BAD => force timeout too */
if (unlikely(io_watch_del(&io_w, c->fd, -1, IO_FD_CLOSING)<0)){
LM_ERR("io_watch_del failed for %p"
" id %d fd %d, state %d, flags %x, main fd %d\n",
c, c->id, c->fd, c->state, c->flags, c->s);
}
tcpconn_listrm(tcp_conn_lst, c, c_next, c_prev);
LM_WARN("%s:%d %s releasing\n", __FILE__, __LINE__, __PRETTY_FUNCTION__);
release_tcpconn(c, (c->state<0)?CONN_ERROR:CONN_RELEASE, tcpmain_sock);
return 0;
}
/*! \brief handles io from a "generic" ser process (get fd or new_fd from a tcp_send)
*
* \param p - pointer in the ser processes array (pt[]), to the entry for
* which an io event was detected
* \param fd_i - fd index in the fd_array (usefull for optimizing
* io_watch_deletes)
* \return handle_* return convention:
* - -1 on error reading from the fd,
* - 0 on EAGAIN or when no more io events are queued
* (receive buffer empty),
* - >0 on successfull reads from the fd (the receive buffer might
* be non-empty).
*/
inline static int handle_ser_child(struct process_table* p, int fd_i)
{
struct tcp_connection* tcpconn;
long response[2];
int cmd;
int bytes;
int ret;
int fd;
ret=-1;
if (p->unix_sock<=0){
/* (we can't have a fd==0, 0 is never closed )*/
LM_CRIT("fd %d for %d (pid %d)\n",
p->unix_sock, (int)(p-&pt[0]), p->pid);
goto error;
}
/* get all bytes and the fd (if transmitted)
* (this is a SOCK_STREAM so read is not atomic) */
bytes=receive_fd(p->unix_sock, response, sizeof(response), &fd,
MSG_DONTWAIT);
if (bytes<(int)sizeof(response)){
/* too few bytes read */
if (bytes==0){
/* EOF -> bad, child has died */
LM_DBG("dead child %d, pid %d"
" (shutting down?)\n", (int)(p-&pt[0]), p->pid);
/* don't listen on it any more */
io_watch_del(&io_h, p->unix_sock, fd_i, 0);
goto error; /* child dead => no further io events from it */
}else if (bytes<0){
/* EAGAIN is ok if we try to empty the buffer
* e.g: SIGIO_RT overflow mode or EPOLL ET */
if ((errno!=EAGAIN) && (errno!=EWOULDBLOCK)){
LM_CRIT("read from child %d (pid %d): %s [%d]\n",
(int)(p-&pt[0]), p->pid, strerror(errno), errno);
ret=-1;
}else{
ret=0;
}
/* try to ignore ? */
goto end;
}else{
/* should never happen */
LM_CRIT("too few bytes received (%d)\n", bytes );
ret=0; /* something was read so there is no error; otoh if
receive_fd returned less then requested => the receive
buffer is empty => no more io queued on this fd */
goto end;
}
}
ret=1; /* something was received, there might be more queued */
LM_DBG("read response= %lx, %ld, fd %d from %d (%d)\n",
response[0], response[1], fd, (int)(p-&pt[0]), p->pid);
cmd=response[1];
tcpconn=(struct tcp_connection*)response[0];
if (tcpconn==0){
LM_CRIT("null tcpconn pointer received from child %d (pid %d)"
"%lx, %lx\n", (int)(p-&pt[0]), p->pid, response[0], response[1]) ;
goto end;
}
switch(cmd){
case CONN_ERROR:
if (!(tcpconn->flags & F_CONN_REMOVED) && (tcpconn->s!=-1)){
io_watch_del(&io_h, tcpconn->s, -1, IO_FD_CLOSING);
tcpconn->flags|=F_CONN_REMOVED;
}
tcpconn_destroy(tcpconn); /* will close also the fd */
break;
case CONN_GET_FD:
/* send the requested FD */
/* WARNING: take care of setting refcnt properly to
* avoid race condition */
if (send_fd(p->unix_sock, &tcpconn, sizeof(tcpconn),
tcpconn->s)<=0){
LM_ERR("send_fd failed\n");
}
break;
case CONN_NEW:
/* update the fd in the requested tcpconn*/
/* WARNING: take care of setting refcnt properly to
* avoid race condition */
if (fd==-1){
LM_CRIT(" cmd CONN_NEW: no fd received\n");
break;
}
tcpconn->s=fd;
/* add tcpconn to the list*/
tcpconn_add(tcpconn);
/* update the timeout*/
tcpconn->timeout=get_ticks()+tcp_con_lifetime;
io_watch_add(&io_h, tcpconn->s, F_TCPCONN, tcpconn);
tcpconn->flags&=~F_CONN_REMOVED;
break;
default:
LM_CRIT("unknown cmd %d\n", cmd);
}
end:
return ret;
error:
return -1;
}
/*! \brief handles io from a tcp child process
* \param tcp_c - pointer in the tcp_children array, to the entry for
* which an io event was detected
* \param fd_i - fd index in the fd_array (usefull for optimizing
* io_watch_deletes)
* \return handle_* return convention: -1 on error, 0 on EAGAIN (no more
* io events queued), >0 on success. success/error refer only to
* the reads from the fd.
*/
inline static int handle_tcp_child(struct tcp_child* tcp_c, int fd_i)
{
struct tcp_connection* tcpconn;
long response[2];
int cmd;
int bytes;
if (tcp_c->unix_sock<=0){
/* (we can't have a fd==0, 0 is never closed )*/
LM_CRIT("fd %d for %d (pid %d, ser no %d)\n", tcp_c->unix_sock,
(int)(tcp_c-&tcp_children[0]), tcp_c->pid, tcp_c->proc_no);
goto error;
}
/* read until sizeof(response)
* (this is a SOCK_STREAM so read is not atomic) */
bytes=recv_all(tcp_c->unix_sock, response, sizeof(response), MSG_DONTWAIT);
if (bytes<(int)sizeof(response)){
if (bytes==0){
/* EOF -> bad, child has died */
LM_DBG("dead tcp child %d (pid %d, no %d)"
" (shutting down?)\n", (int)(tcp_c-&tcp_children[0]),
tcp_c->pid, tcp_c->proc_no );
/* don't listen on it any more */
io_watch_del(&io_h, tcp_c->unix_sock, fd_i, 0);
goto error; /* eof. so no more io here, it's ok to return error */
}else if (bytes<0){
/* EAGAIN is ok if we try to empty the buffer
* e.g.: SIGIO_RT overflow mode or EPOLL ET */
if ((errno!=EAGAIN) && (errno!=EWOULDBLOCK)){
LM_CRIT("read from tcp child %ld (pid %d, no %d) %s [%d]\n",
(long)(tcp_c-&tcp_children[0]), tcp_c->pid,
tcp_c->proc_no, strerror(errno), errno );
}else{
bytes=0;
}
/* try to ignore ? */
goto end;
}else{
/* should never happen */
LM_CRIT("too few bytes received (%d)\n", bytes );
bytes=0; /* something was read so there is no error; otoh if
receive_fd returned less then requested => the receive
buffer is empty => no more io queued on this fd */
goto end;
}
}
LM_DBG("reader response= %lx, %ld from %d \n",
response[0], response[1], (int)(tcp_c-&tcp_children[0]));
cmd=response[1];
tcpconn=(struct tcp_connection*)response[0];
if (tcpconn==0){
/* should never happen */
LM_CRIT("null tcpconn pointer received from tcp child %d (pid %d):"
"%lx, %lx\n", (int)(tcp_c-&tcp_children[0]), tcp_c->pid,
response[0], response[1]) ;
goto end;
}
switch(cmd){
case CONN_RELEASE:
tcp_c->busy--;
if (tcpconn->state==S_CONN_BAD){
tcpconn_destroy(tcpconn);
break;
}
/* update the timeout (lifetime) */
set_tcp_timeout( tcpconn );
tcpconn_put(tcpconn);
/* must be after the de-ref*/
io_watch_add(&io_h, tcpconn->s, F_TCPCONN, tcpconn);
tcpconn->flags&=~F_CONN_REMOVED;
LM_DBG("cmd CONN_RELEASE %p refcnt= %d\n",
tcpconn, tcpconn->refcnt);
break;
case CONN_ERROR:
case CONN_DESTROY:
case CONN_EOF:
/* WARNING: this will auto-dec. refcnt! */
tcp_c->busy--;
/* main doesn't listen on it => we don't have to delete it
if (tcpconn->s!=-1)
io_watch_del(&io_h, tcpconn->s, -1, IO_FD_CLOSING);
*/
tcpconn_destroy(tcpconn); /* closes also the fd */
break;
default:
LM_CRIT("unknown cmd %d from tcp reader %d\n",
cmd, (int)(tcp_c-&tcp_children[0]));
}
end:
return bytes;
error:
return -1;
}
/*! \brief
* handle io routine, based on the fd_map type
* (it will be called from io_wait_loop* )
* params: fm - pointer to a fd hash entry
* idx - index in the fd_array (or -1 if not known)
* return: -1 on error, or when we are not interested any more on reads
* from this fd (e.g.: we are closing it )
* 0 on EAGAIN or when by some other way it is known that no more
* io events are queued on the fd (the receive buffer is empty).
* Usefull to detect when there are no more io events queued for
* sigio_rt, epoll_et, kqueue.
* >0 on successfull read from the fd (when there might be more io
* queued -- the receive buffer might still be non-empty)
*/
inline static int handle_io(struct fd_map* fm, int idx)
{
int ret;
int n;
struct tcp_connection* con;
int s;
long resp;
switch(fm->type){
case F_TCPMAIN:
again:
ret=n=receive_fd(fm->fd, &con, sizeof(con), &s, 0);
LM_DBG("received n=%d con=%p, fd=%d\n", n, con, s);
if (n<0){
if (errno == EWOULDBLOCK || errno == EAGAIN){
ret=0;
break;
}else if (errno == EINTR) goto again;
else{
LM_CRIT("read_fd: %s \n", strerror(errno));
abort(); /* big error*/
}
}
if (n==0){
LM_WARN("0 bytes read\n");
break;
}
if (con==0){
LM_CRIT("null pointer\n");
break;
}
con->fd=s;
if (s==-1) {
LM_ERR("read_fd:no fd read\n");
goto con_error;
}
if (con==tcp_conn_lst){
LM_CRIT("duplicate"
" connection received: %p, id %d, fd %d, refcnt %d"
" state %d (n=%d)\n", con, con->id, con->fd,
con->refcnt, con->state, n);
release_tcpconn(con, CONN_ERROR, tcpmain_sock);
break; /* try to recover */
}
/* reset the per process TCP req struct */
init_tcp_req(¤t_req);
/* 0 attempts so far for this SIP MSG */
con->msg_attempts = 0;
/* must be before io_watch_add, io_watch_add might catch some
* already existing events => might call handle_io and
* handle_io might decide to del. the new connection =>
* must be in the list */
tcpconn_listadd(tcp_conn_lst, con, c_next, c_prev);
con->timeout=get_ticks()+TCP_CHILD_MAX_MSG_TIME;
if (io_watch_add(&io_w, s, F_TCPCONN, con)<0){
LM_CRIT("failed to add new socket to the fd list\n");
tcpconn_listrm(tcp_conn_lst, con, c_next, c_prev);
goto con_error;
}
break;
case F_TCPCONN:
con=(struct tcp_connection*)fm->data;
resp=tcp_read_req(con, &ret);
if (resp<0) {
ret=-1; /* some error occured */
io_watch_del(&io_w, con->fd, idx, IO_FD_CLOSING);
tcpconn_listrm(tcp_conn_lst, con, c_next, c_prev);
con->state=S_CONN_BAD;
release_tcpconn(con, resp, tcpmain_sock);
}
break;
case F_NONE:
LM_CRIT("empty fd map %p (%d): "
"{%d, %d, %p}\n", fm, (int)(fm-io_w.fd_hash),
fm->fd, fm->type, fm->data);
goto error;
default:
LM_CRIT("uknown fd type %d\n", fm->type);
goto error;
}
return ret;
con_error:
con->state=S_CONN_BAD;
release_tcpconn(con, CONN_ERROR, fm->fd);
return ret;
error:
return -1;
}
/* Responsible for reading the request
* * if returns >= 0 : it keeps the connection for further usage
* or releases it manually
* * if returns < 0 : the connection should be released by the
* upper layer
*/
int tcp_read_req(struct tcp_connection* con, int* bytes_read)
{
int bytes;
int total_bytes;
int resp;
long size;
struct tcp_req* req;
char c;
struct receive_info local_rcv;
char *msg_buf;
int msg_len;
bytes=-1;
total_bytes=0;
resp=CONN_RELEASE;
if (con->con_req) {
req=con->con_req;
LM_DBG("Using the per connection buff \n");
} else {
LM_DBG("Using the global ( per process ) buff \n");
req=¤t_req;
}
#ifdef USE_TLS
if (con->type==PROTO_TLS){
if (tls_fix_read_conn(con)!=0){
resp=CONN_ERROR;
goto end_req;
}
if(con->state!=S_CONN_OK) goto end_req; /* not enough data */
}
#endif
again:
if(req->error==TCP_REQ_OK){
bytes=tcp_read_headers(con,req);
//#ifdef EXTRA_DEBUG
/* if timeout state=0; goto end__req; */
LM_DBG("read= %d bytes, parsed=%d, state=%d, error=%d\n",
bytes, (int)(req->parsed-req->start), req->state,
req->error );
LM_DBG("last char=0x%02X, parsed msg=\n%.*s\n",
*(req->parsed-1), (int)(req->parsed-req->start),
req->start);
//#endif
if (bytes==-1){
LM_ERR("failed to read \n");
resp=CONN_ERROR;
goto end_req;
}
total_bytes+=bytes;
/* eof check:
* is EOF if eof on fd and req. not complete yet,
* if req. is complete we might have a second unparsed
* request after it, so postpone release_with_eof
*/
if ((con->state==S_CONN_EOF) && (req->complete==0)) {
LM_DBG("EOF\n");
resp=CONN_EOF;
goto end_req;
}
}
if (req->error!=TCP_REQ_OK){
LM_ERR("bad request, state=%d, error=%d "
"buf:\n%.*s\nparsed:\n%.*s\n", req->state, req->error,
(int)(req->pos-req->buf), req->buf,
(int)(req->parsed-req->start), req->start);
LM_DBG("- received from: port %d\n", con->rcv.src_port);
print_ip("- received from: ip ",&con->rcv.src_ip, "\n");
resp=CONN_ERROR;
goto end_req;
}
if (req->complete){
#ifdef EXTRA_DEBUG
LM_DBG("end of header part\n");
LM_DBG("- received from: port %d\n", con->rcv.src_port);
print_ip("- received from: ip ", &con->rcv.src_ip, "\n");
LM_DBG("headers:\n%.*s.\n",(int)(req->body-req->start), req->start);
#endif
if (req->has_content_len){
LM_DBG("content-length= %d\n", req->content_len);
#ifdef EXTRA_DEBUG
LM_DBG("body:\n%.*s\n", req->content_len,req->body);
#endif
}else{
req->error=TCP_REQ_BAD_LEN;
LM_ERR("content length not present or unparsable\n");
resp=CONN_ERROR;
goto end_req;
}
/* update the timeout - we succesfully read the request */
con->timeout=get_ticks()+TCP_CHILD_MAX_MSG_TIME;
/* if we are here everything is nice and ok*/
update_stat( pt[process_no].load, +1 );
resp=CONN_RELEASE;
#ifdef EXTRA_DEBUG
LM_DBG("calling receive_msg(%p, %d, )\n",
req->start, (int)(req->parsed-req->start));
#endif
/* rcv.bind_address should always be !=0 */
bind_address=con->rcv.bind_address;
/* just for debugging use sendipv4 as receiving socket FIXME*/
/*
if (con->rcv.dst_ip.af==AF_INET6){
bind_address=sendipv6_tcp;
}else{
bind_address=sendipv4_tcp;
}
*/
con->rcv.proto_reserved1=con->id; /* copy the id */
c=*req->parsed; /* ugly hack: zero term the msg & save the
previous char, req->parsed should be ok
because we always alloc BUF_SIZE+1 */
*req->parsed=0;
/* prepare for next request */
size=req->pos-req->parsed;
if (req->state==H_PING_CRLFCRLF) {
/* we send the reply */
if (tcp_send( con->rcv.bind_address, con->rcv.proto,CRLF,
CRLF_LEN, &(con->rcv.src_su), con->rcv.proto_reserved1) < 0) {
LM_ERR("CRLF pong - tcp_send() failed\n");
}
if (!size) {
/* we can release the connection */
io_watch_del(&io_w, con->fd, -1, IO_FD_CLOSING);
tcpconn_listrm(tcp_conn_lst, con, c_next, c_prev);
if (con->state==S_CONN_EOF)
release_tcpconn(con, CONN_EOF, tcpmain_sock);
else
release_tcpconn(con, CONN_RELEASE, tcpmain_sock);
}
} else {
msg_buf = req->start;
msg_len = req->parsed-req->start;
local_rcv = con->rcv;
if (!size) {
/* did not read any more things - we can release the connection */
LM_DBG("We're releasing the connection in state %d \n",con->state);
if (req != ¤t_req) {
/* we have the buffer in the connection tied buff -
* detach it , release the conn and free it afterwards */
con->con_req = NULL;
}
io_watch_del(&io_w, con->fd, -1, IO_FD_CLOSING);
tcpconn_listrm(tcp_conn_lst, con, c_next, c_prev);
/* if we have EOF, signal that to MAIN as well
* otherwise - just pass it back */
if (con->state==S_CONN_EOF)
release_tcpconn(con, CONN_EOF, tcpmain_sock);
else
release_tcpconn(con, CONN_RELEASE, tcpmain_sock);
} else {
LM_DBG("We still have things on the pipe - keeping connection \n");
}
if (receive_msg(msg_buf, msg_len,
&local_rcv) <0)
LM_ERR("receive_msg failed \n");
if (req != ¤t_req)
pkg_free(req);
}
*req->parsed=c;
update_stat( pt[process_no].load, -1 );
if (size) memmove(req->buf, req->parsed, size);
#ifdef EXTRA_DEBUG
LM_DBG("preparing for new request, kept %ld bytes\n", size);
#endif
req->pos=req->buf+size;
req->parsed=req->buf;
req->start=req->buf;
req->body=0;
req->error=TCP_REQ_OK;
req->state=H_SKIP_EMPTY;
req->complete=req->content_len=req->has_content_len=0;
req->bytes_to_go=0;
con->msg_attempts = 0;
/* if we still have some unparsed bytes, try to parse them too*/
if (size) goto again;
} else {
/* request not complete - check the if the thresholds are exceeded */
con->msg_attempts ++;
if (con->msg_attempts == TCP_CHILD_MAX_MSG_CHUNK) {
LM_ERR("Made %u read attempts but message is not complete yet - "
"closing connection \n",con->msg_attempts);
resp = CONN_ERROR;
goto end_req;
}
if (req == ¤t_req) {
/* let's duplicate this - most likely another conn will come in */
LM_DBG("We didn't manage to read a full request. Back to child poll\n");
/* FIXME - PKG or SHM ? */
con->con_req = pkg_malloc(sizeof(struct tcp_req));
if (con->con_req == NULL) {
LM_ERR("No more mem for dynamic con request buffer\n");
resp = CONN_ERROR;
goto end_req;
}
con->con_req->content_len = req->content_len;
con->con_req->bytes_to_go = req->bytes_to_go;
con->con_req->error = req->error;
con->con_req->state = req->state;
if (req->pos != req->buf) {
/* we have read some bytes */
memcpy(con->con_req->buf,req->buf,req->pos-req->buf);
con->con_req->pos = con->con_req->buf + (req->pos-req->buf);
} else {
con->con_req->pos = con->con_req->buf;
}
if (req->start != req->buf)
con->con_req->start = con->con_req->buf + (req->start-req->buf);
else
con->con_req->start = con->con_req->buf;
if (req->parsed != req->buf)
con->con_req->parsed = con->con_req->buf + (req->parsed-req->buf);
else
con->con_req->parsed = con->con_req->buf;
/* zero out the per process req for the future SIP msg */
init_tcp_req(¤t_req);
}
}
LM_DBG("tcp_read_req end\n");
end_req:
if (bytes_read) *bytes_read=total_bytes;
return resp;
}
/* handle io routine, based on the fd_map type
* (it will be called from io_wait_loop* )
* params: fm - pointer to a fd hash entry
* idx - index in the fd_array (or -1 if not known)
* return: -1 on error, or when we are not interested any more on reads
* from this fd (e.g.: we are closing it )
* 0 on EAGAIN or when by some other way it is known that no more
* io events are queued on the fd (the receive buffer is empty).
* Usefull to detect when there are no more io events queued for
* sigio_rt, epoll_et, kqueue.
* >0 on successfull read from the fd (when there might be more io
* queued -- the receive buffer might still be non-empty)
*/
inline static int handle_io(struct fd_map* fm, short events, int idx)
{
int ret;
int n;
int read_flags;
struct tcp_connection* con;
int s;
long resp;
ticks_t t;
/* update the local config */
cfg_update();
switch(fm->type){
case F_TCPMAIN:
again:
ret=n=receive_fd(fm->fd, &con, sizeof(con), &s, 0);
LM_DBG("received n=%d con=%p, fd=%d\n", n, con, s);
if (unlikely(n<0)){
if (errno == EWOULDBLOCK || errno == EAGAIN){
ret=0;
break;
}else if (errno == EINTR) goto again;
else{
LM_CRIT("read_fd: %s \n", strerror(errno));
abort(); /* big error*/
}
}
if (unlikely(n==0)){
LM_ERR("0 bytes read\n");
goto error;
}
if (unlikely(con==0)){
LM_CRIT("null pointer\n");
goto error;
}
con->fd=s;
if (unlikely(s==-1)) {
LM_ERR("read_fd: no fd read\n");
goto con_error;
}
con->reader_pid=my_pid();
if (unlikely(con==tcp_conn_lst)){
LM_CRIT("duplicate connection received: %p, id %d, fd %d, refcnt %d"
" state %d (n=%d)\n", con, con->id, con->fd,
atomic_get(&con->refcnt), con->state, n);
goto con_error;
break; /* try to recover */
}
if (unlikely(con->state==S_CONN_BAD)){
LM_WARN("received an already bad connection: %p id %d refcnt %d\n",
con, con->id, atomic_get(&con->refcnt));
goto con_error;
}
/* if we received the fd there is most likely data waiting to
* be read => process it first to avoid extra sys calls */
read_flags=((con->flags & (F_CONN_EOF_SEEN|F_CONN_FORCE_EOF)) &&
!(con->flags & F_CONN_OOB_DATA))? RD_CONN_FORCE_EOF
:0;
#ifdef USE_TLS
repeat_1st_read:
#endif /* USE_TLS */
resp=tcp_read_req(con, &n, &read_flags);
if (unlikely(resp<0)){
/* some error occured, but on the new fd, not on the tcp
* main fd, so keep the ret value */
if (unlikely(resp!=CONN_EOF))
con->state=S_CONN_BAD;
LM_WARN("%s:%d %s releasing\n", __FILE__, __LINE__, __PRETTY_FUNCTION__);
release_tcpconn(con, resp, tcpmain_sock);
break;
}
#ifdef USE_TLS
/* repeat read if requested (for now only tls might do this) */
if (unlikely(read_flags & RD_CONN_REPEAT_READ))
goto repeat_1st_read;
#endif /* USE_TLS */
/* must be before io_watch_add, io_watch_add might catch some
* already existing events => might call handle_io and
* handle_io might decide to del. the new connection =>
* must be in the list */
tcpconn_listadd(tcp_conn_lst, con, c_next, c_prev);
t=get_ticks_raw();
con->timeout=t+S_TO_TICKS(TCP_CHILD_TIMEOUT);
/* re-activate the timer */
con->timer.f=tcpconn_read_timeout;
local_timer_reinit(&con->timer);
local_timer_add(&tcp_reader_ltimer, &con->timer,
S_TO_TICKS(TCP_CHILD_TIMEOUT), t);
if (unlikely(io_watch_add(&io_w, s, POLLIN, F_TCPCONN, con)<0)){
LM_CRIT("io_watch_add failed for %p id %d fd %d, state %d, flags %x,"
" main fd %d, refcnt %d\n",
con, con->id, con->fd, con->state, con->flags,
con->s, atomic_get(&con->refcnt));
tcpconn_listrm(tcp_conn_lst, con, c_next, c_prev);
local_timer_del(&tcp_reader_ltimer, &con->timer);
goto con_error;
}
break;
case F_TCPCONN:
con=(struct tcp_connection*)fm->data;
if (unlikely(con->state==S_CONN_BAD)){
resp=CONN_ERROR;
if (!(con->send_flags.f & SND_F_CON_CLOSE))
LM_WARN("F_TCPCONN connection marked as bad: %p id %d refcnt %d\n",
con, con->id, atomic_get(&con->refcnt));
goto read_error;
}
read_flags=((
#ifdef POLLRDHUP
(events & POLLRDHUP) |
#endif /* POLLRDHUP */
(events & (POLLHUP|POLLERR)) |
(con->flags & (F_CONN_EOF_SEEN|F_CONN_FORCE_EOF)))
&& !(events & POLLPRI))? RD_CONN_FORCE_EOF: 0;
#ifdef USE_TLS
repeat_read:
#endif /* USE_TLS */
resp=tcp_read_req(con, &ret, &read_flags);
if (unlikely(resp<0)){
read_error:
ret=-1; /* some error occured */
if (unlikely(io_watch_del(&io_w, con->fd, idx,
IO_FD_CLOSING) < 0)){
LM_CRIT("io_watch_del failed for %p id %d fd %d,"
" state %d, flags %x, main fd %d, refcnt %d\n",
con, con->id, con->fd, con->state,
con->flags, con->s, atomic_get(&con->refcnt));
}
tcpconn_listrm(tcp_conn_lst, con, c_next, c_prev);
local_timer_del(&tcp_reader_ltimer, &con->timer);
if (unlikely(resp!=CONN_EOF))
con->state=S_CONN_BAD;
LM_WARN("%s:%d %s releasing\n", __FILE__, __LINE__, __PRETTY_FUNCTION__);
release_tcpconn(con, resp, tcpmain_sock);
}else{
#ifdef USE_TLS
if (unlikely(read_flags & RD_CONN_REPEAT_READ))
goto repeat_read;
#endif /* USE_TLS */
/* update timeout */
con->timeout=get_ticks_raw()+S_TO_TICKS(TCP_CHILD_TIMEOUT);
/* ret= 0 (read the whole socket buffer) if short read &
* !POLLPRI, bytes read otherwise */
ret&=(((read_flags & RD_CONN_SHORT_READ) &&
!(events & POLLPRI)) - 1);
}
break;
case F_NONE:
LM_CRIT("empty fd map %p (%d): {%d, %d, %p}\n",
fm, (int)(fm-io_w.fd_hash),
fm->fd, fm->type, fm->data);
goto error;
default:
LM_CRIT("uknown fd type %d\n", fm->type);
goto error;
}
return ret;
con_error:
con->state=S_CONN_BAD;
LM_WARN("%s:%d %s releasing\n", __FILE__, __LINE__, __PRETTY_FUNCTION__);
release_tcpconn(con, CONN_ERROR, tcpmain_sock);
return ret;
error:
return -1;
}