server: server loop will exhaust data inputs before sleeping

By exhausting data on input, the performance will be more
consistent + the code more clearly distinguishes between
polling and processing. A test showed gdb packet load
performance go from ~1550kByte/s to 1650kBytes/s + being
more stable.

Signed-off-by: Øyvind Harboe <oyvind.harboe@zylin.com>
__archive__
Øyvind Harboe 2009-12-15 15:43:38 +01:00
parent 4a2f4e3433
commit 74ce435d97
1 changed files with 31 additions and 17 deletions

View File

@ -292,9 +292,10 @@ int server_loop(struct command_context *command_context)
{
struct service *service;
bool poll = true;
/* used in select() */
fd_set read_fds;
struct timeval tv;
int fd_max;
/* used in accept() */
@ -305,10 +306,6 @@ int server_loop(struct command_context *command_context)
LOG_ERROR("couldn't set SIGPIPE to SIG_IGN");
#endif
/* do regular tasks after at most 10ms */
tv.tv_sec = 0;
tv.tv_usec = 10000;
while (!shutdown_openocd)
{
/* monitor sockets for activity */
@ -351,12 +348,24 @@ int server_loop(struct command_context *command_context)
#endif
#endif
openocd_sleep_prelude();
kept_alive();
/* Only while we're sleeping we'll let others run */
retval = socket_select(fd_max + 1, &read_fds, NULL, NULL, &tv);
openocd_sleep_postlude();
struct timeval tv;
tv.tv_sec = 0;
if (poll)
{
/* we're just polling this iteration, this is faster on embedded
* hosts */
tv.tv_usec = 0;
retval = socket_select(fd_max + 1, &read_fds, NULL, NULL, &tv);
} else
{
/* Every 100ms */
tv.tv_usec = 100000;
/* Only while we're sleeping we'll let others run */
openocd_sleep_prelude();
kept_alive();
retval = socket_select(fd_max + 1, &read_fds, NULL, NULL, &tv);
openocd_sleep_postlude();
}
if (retval == -1)
{
@ -385,15 +394,20 @@ int server_loop(struct command_context *command_context)
#endif
}
target_call_timer_callbacks();
process_jim_events(command_context);
if (retval == 0)
{
/* do regular tasks after at most 100ms */
tv.tv_sec = 0;
tv.tv_usec = 10000;
/* We only execute these callbacks when there was nothing to do or we timed out */
target_call_timer_callbacks();
process_jim_events(command_context);
FD_ZERO(&read_fds); /* eCos leaves read_fds unchanged in this case! */
/* We timed out/there was nothing to do, timeout rather than poll next time */
poll = false;
} else
{
/* There was something to do, next time we'll just poll */
poll = true;
}
for (service = services; service; service = service->next)