fis-gtm/sr_port/repl_comm.c

496 lines
18 KiB
C
Raw Permalink Normal View History

/****************************************************************
* *
* Copyright 2001, 2011 Fidelity Information Services, Inc *
* *
* This source code contains the intellectual property *
* of its copyright holder(s), and is made available *
* under a license. If you do not know the terms of *
* the license, please stop and do not read further. *
* *
****************************************************************/
#include "mdef.h"
#include "gtm_stdio.h"
#include "gtm_string.h"
#include "gtmio.h"
#include <sys/types.h>
#include "gtm_socket.h"
#include "gtm_inet.h"
#include <sys/time.h>
#include <errno.h>
#include "gtm_fcntl.h"
#include "gtm_unistd.h"
#include "gtm_stat.h"
#ifdef GTM_USE_POLL_FOR_SUBSECOND_SELECT
#include <sys/poll.h>
#endif
#include "repl_msg.h"
#include "repl_errno.h"
#include "repl_dbg.h"
#include "iosp.h"
#include "repl_comm.h"
#include "repl_sp.h"
#include "min_max.h"
#include "rel_quant.h"
#include "repl_log.h"
#include "iotcpdef.h"
#include "gtmmsg.h"
#include "gt_timer.h"
/* These statistics are useful and should perhaps be collected - Vinaya 2003/08/18
*
* Common:
*
* Send:
* # calls to repl_send
* # calls to repl_send with unaligned buffer
* # bytes repl_send called with, distribute into buckets
* % of input bytes repl_send actually sent, or # bytes actually sent distributed into buckets
* # calls to select, and timeout distributed into buckets
* # calls to send
* # calls to select that were interrupted (EINTR)
* # calls to select that were unsuccessful due to system resource shortage (EAGAIN)
* # calls to select that timed out
* # calls to send that were interrupted (EINTR)
* # calls to send that failed due to the message size being too big (EMSGSIZE)
* # calls to send that would have blocked (EWOULDBLOCK)
*
* Receive:
* # calls to repl_recv
* # calls to repl_recv with unaligned buffer
* # bytes repl_recv called with, distribute into buckets
* % of input length repl_recv actually received, or # bytes actuall received distributed into buckets
* # calls to select, and timeout distributed into buckets
* # calls to recv
* # calls to select that were interrupted (EINTR)
* # calls to select that were unsuccessful due to system resource shortage (EAGAIN)
* # calls to select that timed out
* # calls to recv that were interrupted (EINTR)
* # calls to recv that failed due to the connection reset (bytes received == 0)
* # calls to recv that would have blocked (EWOULDBLOCK)
*/
GBLDEF int repl_max_send_buffsize, repl_max_recv_buffsize;
#if defined(__hppa) || defined(__vms)
#define REPL_SEND_TRACE_BUFF_SIZE 65536
#define REPL_RECV_TRACE_BUFF_SIZE 65536
#else
#define REPL_SEND_TRACE_BUFF_SIZE 1048576
#define REPL_RECV_TRACE_BUFF_SIZE 1048576
#endif
#define REPL_SEND_SIZE_TRACE_SIZE 1024
#define REPL_RECV_SIZE_TRACE_SIZE 1024
STATICDEF int repl_send_trace_buff_pos = 0;
STATICDEF unsigned char * repl_send_trace_buff = 0;
STATICDEF int repl_send_size_trace_pos = 0;
STATICDEF int repl_send_size_trace[REPL_SEND_SIZE_TRACE_SIZE];
STATICDEF int repl_recv_trace_buff_pos = 0;
STATICDEF unsigned char * repl_recv_trace_buff = 0;
STATICDEF int repl_recv_size_trace_pos = 0;
STATICDEF int repl_recv_size_trace[REPL_RECV_SIZE_TRACE_SIZE];
int repl_send(int sock_fd, unsigned char *buff, int *send_len, boolean_t skip_pipe_ready_check, struct timeval *max_pipe_ready_wait)
{ /* On entry, *send_len is the number of bytes to be sent
* On exit, *send_len contains the number of bytes sent
*/
int send_size, status, eintr_cnt, eagain_cnt, ewouldblock_cnt, emsgsize_cnt;
ssize_t bytes_sent;
long wait_val;
fd_set output_fds;
struct timeval timeout;
#ifdef GTM_USE_POLL_FOR_SUBSECOND_SELECT
long poll_timeout;
unsigned long poll_nfds;
struct pollfd poll_fdlist[1];
#endif
int space_to_end;
if (!repl_send_trace_buff)
repl_send_trace_buff = malloc(REPL_SEND_TRACE_BUFF_SIZE);
/* Note: there is no corresponding free for this malloc since it is only done once per process and will not
* accumulate across multiple process invocations. It will be "freed" when the mupip process exits.
*/
send_size = *send_len;
/* VMS returns SYSTEM-F-INVBUFLEN if send_size is larger than the hard limit VMS_MAX_TCP_SEND_SIZE (64K - 1 on some
* impelementations, 64K - 512 on some others). VMS_MAX_TCP_SEND_SIZE may be larger than repl_max_send_buffsize, and
* empirically we have noticed send() successfully sending repl_max_send_buffsize or more bytes.
*/
VMS_ONLY(send_size = MIN(send_size, VMS_MAX_TCP_SEND_SIZE);)
REPL_DPRINT3("repl_send: called with send_size %d, limiting send_size to %d\n", *send_len, send_size);
*send_len = 0;
if (!skip_pipe_ready_check)
{ /* Wait till connection pipe is ready for sending */
assert(max_pipe_ready_wait->tv_sec == 0); /* all callers pass sub-second timeout. We take advantage of this fact
* to avoid division while computing wait_val */
assert(max_pipe_ready_wait->tv_usec >= 0 && max_pipe_ready_wait->tv_usec < MICROSEC_IN_SEC);
#ifndef GTM_USE_POLL_FOR_SUBSECOND_SELECT
FD_ZERO(&output_fds);
FD_SET(sock_fd, &output_fds);
#else
poll_fdlist[0].fd = sock_fd;
poll_fdlist[0].events = POLLOUT;
poll_nfds = 1;
poll_timeout = max_pipe_ready_wait->tv_usec / 1000; /* convert to millisecs */
#endif
/* the check for EINTR below is valid and should not be converted to an EINTR wrapper macro, because EAGAIN is also
* being checked */
for (timeout = *max_pipe_ready_wait, eintr_cnt = eagain_cnt = 0;
#ifndef GTM_USE_POLL_FOR_SUBSECOND_SELECT
-1 == (status = select(sock_fd + 1, NULL, &output_fds, NULL, &timeout))
#else
-1 == (status = poll(&poll_fdlist[0], poll_nfds, poll_timeout))
#endif
&& (EINTR == errno || EAGAIN == errno); )
{
if (EINTR == errno)
{
wait_val = (0 == eintr_cnt) ? (max_pipe_ready_wait->tv_usec >> 1) : (wait_val >> 1);
if (++eintr_cnt > REPL_COMM_MAX_INTR_CNT)
{ /* assume timeout and give up. Note, this may result in a sleep different from intended.
* But, we can live with it as there is no impact on the user */
status = 0;
break;
}
timeout.tv_sec = 0;
timeout.tv_usec = (gtm_tv_usec_t)wait_val;
#ifdef GTM_USE_POLL_FOR_SUBSECOND_SELECT
poll_timeout = wait_val / 1000; /* convert to millisecs */
#endif
REPL_DPRINT5("repl_send: select interrupted, changing timeout from tv_sec %ld tv_usec %ld to "
"tv_sec %ld tv_usec %ld\n", 0, (wait_val << 1), timeout.tv_sec, timeout.tv_usec);
} else
{ /* resource starved system; relinquish the processor in the hope that we may get the required
* resources the next time around */
if (0 == ++eagain_cnt % REPL_COMM_LOG_EAGAIN_INTERVAL)
repl_log(stderr, TRUE, TRUE, "Communication subsystem warning: System appears to be "
"resource starved; EAGAIN returned from repl_send/select %d times\n", eagain_cnt);
rel_quant();
}
#ifndef GTM_USE_POLL_FOR_SUBSECOND_SELECT
FD_SET(sock_fd, &output_fds); /* Linux/gcc does not like this in the iterator */
#endif
}
} else
status = 1; /* assume connection pipe is ready for sending */
if (0 < status)
{ /* Ready for sending */
/*
* the check for EINTR below is valid and should not be converted to an EINTR wrapper macro, because other errno
* values are being checked.
*/
/* Trace last REPL_SEND_SIZE_TRACE_SIZE sizes of what was sent */
repl_send_size_trace[repl_send_size_trace_pos++] = send_size;
repl_send_size_trace_pos %= ARRAYSIZE(repl_send_size_trace);
/* Trace last REPL_SEND_TRACE_BUFF_SIZE bytes sent. */
if (send_size > REPL_SEND_TRACE_BUFF_SIZE)
{
/* if the message size > our buffer just copy the last our buffer-size-worth starting from the
* beginning of our buffer and reset pos to the beginning of our buffer.
*/
memcpy(repl_send_trace_buff, buff + send_size - REPL_SEND_TRACE_BUFF_SIZE, REPL_SEND_TRACE_BUFF_SIZE);
repl_send_trace_buff_pos = 0;
}
else
{
space_to_end = REPL_SEND_TRACE_BUFF_SIZE - repl_send_trace_buff_pos;
if (send_size > space_to_end)
{
memcpy(repl_send_trace_buff + repl_send_trace_buff_pos, buff, space_to_end);
memcpy(repl_send_trace_buff, buff + space_to_end, send_size - space_to_end);
}
else
{
memcpy(repl_send_trace_buff + repl_send_trace_buff_pos, buff, send_size);
}
repl_send_trace_buff_pos = (repl_send_trace_buff_pos + send_size) % REPL_SEND_TRACE_BUFF_SIZE;
}
if (0 < send_size)
{
for (ewouldblock_cnt = emsgsize_cnt = 0;
(((bytes_sent = send(sock_fd, (char *)buff, send_size, 0)) < 0)
&& (EINTR == errno || EMSGSIZE == errno || EWOULDBLOCK == errno)); )
{
if (EINTR == errno)
continue;
assert(EMSGSIZE != errno); /* since we use blocking sockets, we don't expect to see EMSGSIZE */
if (EMSGSIZE == errno)
{ /* Reduce the send size if possible */
if (send_size > REPL_COMM_MIN_SEND_SIZE)
{
if ((send_size >> 1) <= REPL_COMM_MIN_SEND_SIZE)
send_size = REPL_COMM_MIN_SEND_SIZE;
else
send_size >>= 1;
}
if (0 == ++emsgsize_cnt % REPL_COMM_LOG_EMSGSIZE_INTERVAL)
repl_log(stderr, TRUE, TRUE, "Communication subsystem warning: System "
"appears to be clogged; EMSGSIZE returned from send "
"%d times\n", emsgsize_cnt);
} else
{
if (0 == ++ewouldblock_cnt % REPL_COMM_LOG_EWDBLCK_INTERVAL)
repl_log(stderr, TRUE, TRUE, "Communication subsystem warning: System "
"appears to be running slow; EWOULDBLOCK returned from "
"send %d times\n", ewouldblock_cnt);
rel_quant(); /* we hope the cause for blocking would have cleared by the time
* we get scheduled next time around */
}
}
} else
bytes_sent = 0;
if (0 <= bytes_sent)
{
*send_len = (int)bytes_sent;
REPL_DPRINT2("repl_send: returning with send_len %ld\n", bytes_sent);
return (SS_NORMAL);
}
repl_errno = EREPL_SEND;
return (ERRNO);
} else if (0 == status)
return (SS_NORMAL);
repl_errno = EREPL_SELECT;
return (ERRNO);
}
int repl_recv(int sock_fd, unsigned char *buff, int *recv_len, boolean_t skip_data_avail_check, struct timeval *max_data_avail_wait)
{ /* On entry *recv_len is the maximum length to be received.
* On exit, *recv_len will contain the number of bytes received.
*/
int status, max_recv_len, eintr_cnt, eagain_cnt;
ssize_t bytes_recvd;
long wait_val;
fd_set input_fds;
struct timeval timeout;
#ifdef GTM_USE_POLL_FOR_SUBSECOND_SELECT
long poll_timeout;
unsigned long poll_nfds;
struct pollfd poll_fdlist[1];
#endif
int space_to_end;
if (!repl_recv_trace_buff)
repl_recv_trace_buff = malloc(REPL_RECV_TRACE_BUFF_SIZE);
/* Note: there is no corresponding free for this malloc since it is only done once per process and will not
* accumulate across multiple process invocations. It will be "freed" when the mupip process exits.
*/
assert(FD_INVALID != sock_fd);
max_recv_len = *recv_len;
/* VMS returns SYSTEM-F-INVBUFLEN if max_recv_len is larger than the hard limit VMS_MAX_TCP_SEND_SIZE (64K - 1 on some
* impelementations, 64K - 512 on some others). Although VMS_MAX_TCP_RECV_SIZE may be larger than repl_max_recv_buffsize,
* we have empirically noticed recv() returning with repl_max_recv_buffsize or fewer bytes.
*/
VMS_ONLY(max_recv_len = MIN(max_recv_len, VMS_MAX_TCP_RECV_SIZE);)
REPL_DPRINT3("repl_recv: called with max_recv_len %d, limiting max_recv_len to %d\n", *recv_len, max_recv_len);
*recv_len = 0;
if (!skip_data_avail_check)
{
assert(max_data_avail_wait->tv_sec == 0); /* all callers pass sub-second timeout. We take advantage of this fact
* to avoid division while computing wait_val */
assert(max_data_avail_wait->tv_usec >= 0 && max_data_avail_wait->tv_usec < MICROSEC_IN_SEC);
#ifndef GTM_USE_POLL_FOR_SUBSECOND_SELECT
FD_ZERO(&input_fds);
FD_SET(sock_fd, &input_fds);
#else
poll_fdlist[0].fd = sock_fd;
poll_fdlist[0].events = POLLIN;
poll_nfds = 1;
poll_timeout = max_data_avail_wait->tv_usec / 1000; /* convert to millisecs */
#endif
/* the check for EINTR below is valid and should not be converted to an EINTR wrapper macro, because EAGAIN is also
* being checked */
for (timeout = *max_data_avail_wait, eintr_cnt = eagain_cnt = 0;
#ifndef GTM_USE_POLL_FOR_SUBSECOND_SELECT
-1 == (status = select(sock_fd + 1, &input_fds, NULL, NULL, &timeout))
#else
-1 == (status = poll(&poll_fdlist[0], poll_nfds, poll_timeout))
#endif
&& (EINTR == errno || EAGAIN == errno); )
{
if (EINTR == errno)
{
wait_val = (0 == eintr_cnt) ? (max_data_avail_wait->tv_usec >> 1) : (wait_val >> 1);
if (++eintr_cnt > REPL_COMM_MAX_INTR_CNT)
{ /* assume timeout and give up. Note, this may result in a sleep different from intended.
* But, we can live with it as there is no impact on the user */
status = 0;
break;
}
timeout.tv_sec = 0;
timeout.tv_usec = (gtm_tv_usec_t)wait_val;
#ifdef GTM_USE_POLL_FOR_SUBSECOND_SELECT
poll_timeout = wait_val / 1000; /* convert to millisecs */
#endif
REPL_DPRINT5("repl_recv: select interrupted, changing timeout from tv_sec %ld tv_usec %ld to "
"tv_sec %ld tv_usec %ld\n", 0, (wait_val << 1), timeout.tv_sec, timeout.tv_usec);
} else
{ /* resource starved system; relinquish the processor in the hope that we may get the required
* resources the next time around */
if (0 == ++eagain_cnt % REPL_COMM_LOG_EAGAIN_INTERVAL)
repl_log(stderr, TRUE, TRUE, "Communication subsystem warning: System appears to be "
"resource starved; EAGAIN returned from select %d times\n", eagain_cnt);
rel_quant();
}
#ifndef GTM_USE_POLL_FOR_SUBSECOND_SELECT
FD_SET(sock_fd, &input_fds); /* Linux/gcc does not like this in the iterator */
#endif
}
} else
status = 1; /* assume data is available, consequence is recv() may block if data is not available */
if (0 < status)
{ /* Data available on the pipe */
while (0 > (bytes_recvd = recv(sock_fd, (char *)buff, max_recv_len, 0)) && EINTR == errno)
;
if (0 < bytes_recvd)
{
*recv_len = (int)bytes_recvd;
REPL_DPRINT2("repl_recv: returning with recv_len %ld\n", bytes_recvd);
/* Trace last REPL_RECV_SIZE_TRACE_SIZE sizes of what was received */
repl_recv_size_trace[repl_recv_size_trace_pos++] = bytes_recvd;
repl_recv_size_trace_pos %= ARRAYSIZE(repl_recv_size_trace);
/* Trace last REPL_RECV_TRACE_BUFF_SIZE bytes received. */
if (bytes_recvd > REPL_RECV_TRACE_BUFF_SIZE)
{
/* if the message size > our buffer just copy the last our buffer-size-worth starting from the
* beginning of our buffer and reset pos to the beginning of our buffer.
*/
memcpy(repl_recv_trace_buff, buff + bytes_recvd - REPL_RECV_TRACE_BUFF_SIZE,
REPL_RECV_TRACE_BUFF_SIZE);
repl_recv_trace_buff_pos = 0;
}
else
{
space_to_end = REPL_RECV_TRACE_BUFF_SIZE - repl_recv_trace_buff_pos;
if (bytes_recvd > space_to_end)
{
memcpy(repl_recv_trace_buff + repl_recv_trace_buff_pos, buff, space_to_end);
memcpy(repl_recv_trace_buff, buff + space_to_end, bytes_recvd - space_to_end);
}
else
{
memcpy(repl_recv_trace_buff + repl_recv_trace_buff_pos, buff, bytes_recvd);
}
repl_recv_trace_buff_pos = (repl_recv_trace_buff_pos + bytes_recvd) %
REPL_RECV_TRACE_BUFF_SIZE;
}
return (SS_NORMAL);
}
if (0 == bytes_recvd)
{ /* Connection reset */
errno = ECONNRESET;
} else if (EWOULDBLOCK == errno)
{ /* NOTE: Although we use non blocking sockets, it is possible to get EWOULDBLOCK error status if receive
* timeout has been set and the timeout expired before data was received (from man recv on RH 8 Linux). Some
* systems return ETIMEDOUT for the timeout condition. */
repl_log(stderr, TRUE, TRUE, "Communication subsystem warning: network I/O failed to complete; "
"socket recv() returned EWOULDBLOCK\n");
errno = ETIMEDOUT; /* will be treated as a bad connection and the connection closed */
}
repl_errno = EREPL_RECV;
return (ERRNO);
} else if (0 == status)
return (SS_NORMAL);
repl_errno = EREPL_SELECT;
return (ERRNO);
}
int repl_close(int *sock_fd)
{
int status = 0;
if (FD_INVALID != *sock_fd)
CLOSEFILE_RESET(*sock_fd, status); /* resets "*sock_fd" to FD_INVALID */
return (0 == status ? 0 : ERRNO);
}
static int get_sock_buff_size(int sockfd, int *buffsize, int which_buf)
{
int status;
GTM_SOCKLEN_TYPE optlen;
optlen = SIZEOF(*buffsize);
status = getsockopt(sockfd, SOL_SOCKET, which_buf, (void *)buffsize, (GTM_SOCKLEN_TYPE *)&optlen);
return (0 == status) ? 0 : ERRNO;
}
int get_send_sock_buff_size(int sockfd, int *send_buffsize)
{
return get_sock_buff_size(sockfd, send_buffsize, SO_SNDBUF);
}
int get_recv_sock_buff_size(int sockfd, int *recv_buffsize)
{
return get_sock_buff_size(sockfd, recv_buffsize, SO_RCVBUF);
}
static int set_sock_buff_size(int sockfd, int buflen, int which_buf)
{
int status;
#ifndef sun
size_t optlen;
#else
int optlen;
#endif
optlen = SIZEOF(buflen);
status = setsockopt(sockfd, SOL_SOCKET, which_buf, (void *)&buflen, (GTM_SOCKLEN_TYPE)optlen);
return (0 == status) ? 0 : ERRNO;
}
int set_send_sock_buff_size(int sockfd, int buflen)
{
return set_sock_buff_size(sockfd, buflen, SO_SNDBUF);
}
int set_recv_sock_buff_size(int sockfd, int buflen)
{
return set_sock_buff_size(sockfd, buflen, SO_RCVBUF);
}
void repl_log_conn_info(int sock_fd, FILE *log_fp)
{
struct sockaddr_in local, remote;
GTM_SOCKLEN_TYPE len;
int save_errno;
size_t errlen;
char *errptr, local_ip[16], remote_ip[16];
in_port_t local_port, remote_port;
error_def(ERR_GETSOCKNAMERR);
error_def(ERR_TEXT);
len = SIZEOF(local);
if (0 == getsockname(sock_fd, (struct sockaddr *)&local, (GTM_SOCKLEN_TYPE *)&len))
{
local_port = ntohs(local.sin_port);
strcpy(local_ip, inet_ntoa(local.sin_addr));
} else
{
save_errno = errno;
errptr = (char *)STRERROR(save_errno);
errlen = strlen(errptr);
gtm_putmsg(VARLSTCNT(9) ERR_GETSOCKNAMERR, 3, save_errno, errlen, errptr, ERR_TEXT, 2,
LEN_AND_LIT("LOCAL"));
local_port = (unsigned short)-1;
strcpy(local_ip, "*UNKNOWN*");
}
len = SIZEOF(remote);
if (0 == getpeername(sock_fd, (struct sockaddr *)&remote, (GTM_SOCKLEN_TYPE *)&len))
{
remote_port = ntohs(remote.sin_port);
strcpy(remote_ip, inet_ntoa(remote.sin_addr));
} else
{
save_errno = errno;
errptr = (char *)STRERROR(save_errno);
errlen = strlen(errptr);
gtm_putmsg(VARLSTCNT(9) ERR_GETSOCKNAMERR, 3, save_errno, errlen, errptr, ERR_TEXT, 2,
LEN_AND_LIT("REMOTE"));
remote_port = (unsigned short)-1;
strcpy(remote_ip, "*UNKNOWN*");
}
repl_log(log_fp, TRUE, TRUE, "Connection information:: Local: %s:%d Remote: %s:%d\n", local_ip, local_port,
remote_ip, remote_port);
return;
}