void* scan_nodes(void* threadid) {
int i, j, k;
long tid;
unsigned int min_nbr_height;
Edge nbr_edge, min_height_edge;
long d;
long local_e;
long local_c;
int max_flow;
int isInQ;
int height_before_lift;
int node_push_times = 0, node_lift_times = 0;
int op_times = 0;
tid = (long)threadid;
Node* cur_node;
ThreadEnv* thisThread;
thisThread = threadEnv + tid;
#if 0
if (tid == 0)
global_relabel(tid);
#endif
pthread_barrier_wait(&start_barrier);
while (!flow_done()) {
cur_node = thisThread->Q[0];
while (cur_node != NoNode) {
#ifdef GR
if (op_times > gr_threshold) {
op_times = 0;
if (pthread_mutex_trylock(&gr_mutex) == 0) {
global_relabel(tid);
pthread_mutex_unlock(&gr_mutex);
}
}
#endif
while (cur_node->excess > 0) {
#ifdef DEBUG
fprintf(stderr,
"%d h=%d, e=%ld\n",
cur_node - g_node,
cur_node->height,
cur_node->excess);
fflush(stderr);
#endif
min_nbr_height = UINT_MAX;
for (nbr_edge = cur_node->adj_list; nbr_edge < (cur_node + 1)->adj_list;
nbr_edge++) {
if (nbr_edge->capacity > 0 &&
(nbr_edge->endpoint)->height < min_nbr_height) {
min_nbr_height = (nbr_edge->endpoint)->height;
min_height_edge = nbr_edge;
}
}
#ifdef DEBUG
fprintf(stderr, "work on %d\n", cur_node - g_node);
fflush(stderr);
#endif
if (cur_node->height > min_nbr_height) {
local_e = cur_node->excess;
local_c = min_height_edge->capacity;
d = MIN(local_e, local_c);
if (min_height_edge->endpoint->wave == cur_node->wave &&
cur_node->height > min_height_edge->endpoint->height) {
node_push_times++;
op_times++;
atomic_add(d, &(min_height_edge->mateedge->capacity));
atomic_sub(d, &(min_height_edge->capacity));
atomic_add(d, &((min_height_edge->endpoint)->excess));
atomic_sub(d, &(cur_node->excess));
#if defined(PUSH) || defined(DEBUG)
fprintf(stderr,
"[%ld] %ld(%ld) -> %ld -> %ld(%ld) \n",
tid,
cur_node - g_node,
cur_node->excess,
d,
min_height_edge->endpoint - g_node,
(min_height_edge->endpoint)->excess);
fflush(stderr);
#endif
// add min_nbr to local queue
isInQ = cmpxchg(&(min_height_edge->endpoint->inQ), 0, tid + 1);
if (isInQ == 0)
enQ(thisThread, min_height_edge->endpoint);
}
} else {
// if we cannot push to any nodes, then we must be able to lift
node_lift_times++;
op_times++;
pthread_mutex_lock(&(node_mutex[cur_node - g_node]));
if (cur_node->height < min_nbr_height + 1)
cur_node->height = min_nbr_height + 1;
pthread_mutex_unlock(&(node_mutex[cur_node - g_node]));
#if defined(LIFT) || defined(DEBUG)
fprintf(stderr,
"%ld ^ %d, ref %ld(%d)\n",
cur_node - g_node,
cur_node->height,
min_height_edge->endpoint - g_node,
min_height_edge->endpoint->height);
fflush(stderr);
#endif
}
} // while( g_node[i].excess > 0 )
set0(&(cur_node->inQ));
if (cur_node->excess > 0) {
isInQ = cmpxchg(&(cur_node->inQ), 0, tid + 1);
if (isInQ == 0) {
reenQ(thisThread, cur_node);
} else {
deQ(thisThread);
}
} else {
deQ(thisThread);
}
#ifdef HELP
if (thisThread->request < MAX_THRD)
send_work(tid);
#endif
cur_node = thisThread->Q[thisThread->head];
} // while (i != -1)
#ifdef HELP
// Q is empty, find something to do;
request_work(tid);
#else
break;
#endif
} // while(!flow_done())
atomic_add(node_push_times, &(totalPushes));
atomic_add(node_lift_times, &(totalLifts));
} // scan_node
/*===========================================================================*
* get_work *
*===========================================================================*/
static void get_work()
{
/* Normally wait for new input. However, if 'reviving' is
* nonzero, a suspended process must be awakened.
*/
int r, found_one, proc_p;
register struct fproc *rp;
while (reviving != 0) {
found_one = FALSE;
/* Find a suspended process. */
for (rp = &fproc[0]; rp < &fproc[NR_PROCS]; rp++)
if (rp->fp_pid != PID_FREE && (rp->fp_flags & FP_REVIVED)) {
found_one = TRUE; /* Found a suspended process */
if (unblock(rp))
return; /* So main loop can process job */
send_work();
}
if (!found_one) /* Consistency error */
panic("VFS: get_work couldn't revive anyone");
}
for(;;) {
/* Normal case. No one to revive. Get a useful request. */
if ((r = sef_receive(receive_from, &m_in)) != OK) {
panic("VFS: sef_receive error: %d", r);
}
proc_p = _ENDPOINT_P(m_in.m_source);
if (proc_p < 0) fp = NULL;
else fp = &fproc[proc_p];
if (m_in.m_type == EDEADSRCDST) return; /* Failed 'sendrec' */
/* Negative who_p is never used to access the fproc array. Negative
* numbers (kernel tasks) are treated in a special way.
*/
if (who_p >= (int)(sizeof(fproc) / sizeof(struct fproc)))
panic("receive process out of range: %d", who_p);
if (who_p >= 0 && fproc[who_p].fp_endpoint == NONE) {
printf("VFS: ignoring request from %d: NONE endpoint %d (%d)\n",
m_in.m_source, who_p, m_in.m_type);
continue;
}
/* Internal consistency check; our mental image of process numbers and
* endpoints must match with how the rest of the system thinks of them.
*/
if (who_p >= 0 && fproc[who_p].fp_endpoint != who_e) {
if (fproc[who_p].fp_endpoint == NONE)
printf("slot unknown even\n");
printf("VFS: receive endpoint inconsistent (source %d, who_p "
"%d, stored ep %d, who_e %d).\n", m_in.m_source, who_p,
fproc[who_p].fp_endpoint, who_e);
panic("VFS: inconsistent endpoint ");
}
return;
}
}
void process_request(char* code_sign_key) {
PLATFORM* platform;
int retval;
double last_rpc_time, x;
struct tm *rpc_time_tm;
bool ok_to_send_work = !config.dont_send_jobs;
bool have_no_work = false;
char buf[256];
HOST initial_host;
unsigned int i;
time_t t;
memset(&g_reply->wreq, 0, sizeof(g_reply->wreq));
// if client has sticky files we don't need any more, tell it
//
do_file_delete_regex();
// if different major version of BOINC, just send a message
//
if (wrong_core_client_version()
|| unacceptable_os()
|| unacceptable_cpu()
) {
ok_to_send_work = false;
}
// if no jobs reported and none to send, return without accessing DB
//
if (!ok_to_send_work && !g_request->results.size()) {
return;
}
warn_user_if_core_client_upgrade_scheduled();
if (requesting_work()) {
if (config.locality_scheduling || config.locality_scheduler_fraction || config.enable_assignment) {
have_no_work = false;
} else {
lock_sema();
have_no_work = ssp->no_work(g_pid);
if (have_no_work) {
g_wreq->no_jobs_available = true;
}
unlock_sema();
}
}
// If:
// - there's no work,
// - a config flag is set,
// - client isn't returning results,
// - this isn't an initial RPC,
// - client is requesting work
// then return without accessing the DB.
// This is an efficiency hack for when servers are overloaded
//
if (
have_no_work
&& config.nowork_skip
&& requesting_work()
&& (g_request->results.size() == 0)
&& (g_request->hostid != 0)
) {
g_reply->insert_message("No work available", "low");
g_reply->set_delay(DELAY_NO_WORK_SKIP);
if (!config.msg_to_host && !config.enable_vda) {
log_messages.printf(MSG_NORMAL, "No work - skipping DB access\n");
return;
}
}
// FROM HERE ON DON'T RETURN; "goto leave" instead
// (because ssp->no_work() may have tagged an entry in the work array
// with our process ID)
retval = open_database();
if (retval) {
send_error_message("Server can't open database", 3600);
g_reply->project_is_down = true;
goto leave;
}
retval = authenticate_user();
if (retval) goto leave;
if (g_reply->user.id == 0) {
log_messages.printf(MSG_CRITICAL, "No user ID!\n");
}
initial_host = g_reply->host;
g_reply->host.rpc_seqno = g_request->rpc_seqno;
g_reply->nucleus_only = false;
log_request();
// is host blacklisted?
//
if (g_reply->host._max_results_day == -1) {
send_error_message("Not accepting requests from this host", 86400);
goto leave;
}
if (strlen(config.sched_lockfile_dir)) {
int pid_with_lock = lock_sched();
if (pid_with_lock > 0) {
log_messages.printf(MSG_CRITICAL,
"Another scheduler instance [PID=%d] is running for this host\n",
pid_with_lock
);
} else if (pid_with_lock) {
log_messages.printf(MSG_CRITICAL,
"Error acquiring lock for [HOST#%d]\n", g_reply->host.id
);
}
if (pid_with_lock) {
send_error_message(
"Another scheduler instance is running for this host", 60
);
goto leave;
}
}
// in deciding whether it's a new day,
// add a random factor (based on host ID)
// to smooth out network traffic over the day
//
retval = rand();
srand(g_reply->host.id);
x = drand()*86400;
srand(retval);
last_rpc_time = g_reply->host.rpc_time;
t = (time_t)(g_reply->host.rpc_time + x);
rpc_time_tm = localtime(&t);
g_request->last_rpc_dayofyear = rpc_time_tm->tm_yday;
t = time(0);
g_reply->host.rpc_time = t;
t += (time_t)x;
rpc_time_tm = localtime(&t);
g_request->current_rpc_dayofyear = rpc_time_tm->tm_yday;
retval = modify_host_struct(g_reply->host);
// write time stats to disk if present
//
if (g_request->have_time_stats_log) {
write_time_stats_log();
}
// look up the client's platform(s) in the DB
//
platform = ssp->lookup_platform(g_request->platform.name);
if (platform) g_request->platforms.list.push_back(platform);
// if primary platform is anonymous, ignore alternate platforms
//
if (strcmp(g_request->platform.name, "anonymous")) {
for (i=0; i<g_request->alt_platforms.size(); i++) {
platform = ssp->lookup_platform(g_request->alt_platforms[i].name);
if (platform) g_request->platforms.list.push_back(platform);
}
}
if (g_request->platforms.list.size() == 0) {
sprintf(buf, "%s %s",
_("This project doesn't support computers of type"),
g_request->platform.name
);
g_reply->insert_message(buf, "notice");
log_messages.printf(MSG_CRITICAL,
"[HOST#%d] platform '%s' not found\n",
g_reply->host.id, g_request->platform.name
);
g_reply->set_delay(DELAY_PLATFORM_UNSUPPORTED);
goto leave;
}
handle_global_prefs();
read_host_app_versions();
update_n_jobs_today();
handle_results();
handle_file_xfer_results();
if (config.enable_vda) {
handle_vda();
}
// Do this before resending lost jobs
//
if (bad_install_type()) {
ok_to_send_work = false;
}
if (!requesting_work()) {
ok_to_send_work = false;
}
send_work_setup();
if (g_request->have_other_results_list) {
if (ok_to_send_work
&& (config.resend_lost_results || g_wreq->resend_lost_results)
&& !g_request->results_truncated
) {
if (resend_lost_work()) {
ok_to_send_work = false;
}
}
if (config.send_result_abort) {
send_result_abort();
}
}
if (requesting_work()) {
if (!send_code_sign_key(code_sign_key)) {
ok_to_send_work = false;
}
if (have_no_work) {
if (config.debug_send) {
log_messages.printf(MSG_NORMAL,
"[send] No jobs in shmem cache\n"
);
}
}
// if last RPC was within config.min_sendwork_interval, don't send work
//
if (!have_no_work && ok_to_send_work) {
if (config.min_sendwork_interval) {
double diff = dtime() - last_rpc_time;
if (diff < config.min_sendwork_interval) {
ok_to_send_work = false;
log_messages.printf(MSG_NORMAL,
"Not sending work - last request too recent: %f\n", diff
);
sprintf(buf,
"Not sending work - last request too recent: %d sec", (int)diff
);
g_reply->insert_message(buf, "low");
// the 1.01 is in case client's clock
// is slightly faster than ours
//
g_reply->set_delay(1.01*config.min_sendwork_interval);
}
}
if (ok_to_send_work) {
send_work();
}
}
if (g_wreq->no_jobs_available) {
g_reply->insert_message("Project has no tasks available", "low");
}
}
handle_msgs_from_host();
if (config.msg_to_host) {
handle_msgs_to_host();
}
update_host_record(initial_host, g_reply->host, g_reply->user);
write_host_app_versions();
leave:
if (!have_no_work) {
ssp->restore_work(g_pid);
}
}
/*===========================================================================*
* main *
*===========================================================================*/
int main(void)
{
/* This is the main program of the file system. The main loop consists of
* three major activities: getting new work, processing the work, and sending
* the reply. This loop never terminates as long as the file system runs.
*/
int transid;
struct job *job;
/* SEF local startup. */
sef_local_startup();
printf("Started VFS: %d worker thread(s)\n", NR_WTHREADS);
/* This is the main loop that gets work, processes it, and sends replies. */
while (TRUE) {
yield_all(); /* let other threads run */
self = NULL;
job = NULL;
send_work();
get_work();
transid = TRNS_GET_ID(m_in.m_type);
if (IS_VFS_FS_TRANSID(transid)) {
job = worker_getjob( (thread_t) transid - VFS_TRANSID);
if (job == NULL) {
printf("VFS: spurious message %d from endpoint %d\n",
m_in.m_type, m_in.m_source);
continue;
}
m_in.m_type = TRNS_DEL_ID(m_in.m_type);
}
if (job != NULL) {
do_fs_reply(job);
continue;
} else if (who_e == PM_PROC_NR) { /* Calls from PM */
/* Special control messages from PM */
sys_worker_start(do_pm);
continue;
} else if (is_notify(call_nr)) {
/* A task notify()ed us */
sys_worker_start(do_control_msgs);
continue;
} else if (who_p < 0) { /* i.e., message comes from a task */
/* We're going to ignore this message. Tasks should
* send notify()s only.
*/
printf("VFS: ignoring message from %d (%d)\n", who_e, call_nr);
continue;
}
/* At this point we either have results from an asynchronous device
* or a new system call. In both cases a new worker thread has to be
* started and there might not be one available from the pool. This is
* not a problem (requests/replies are simply queued), except when
* they're from an FS endpoint, because these can cause a deadlock.
* handle_work() takes care of the details. */
if (IS_DEV_RS(call_nr)) {
/* We've got results for a device request */
handle_work(do_async_dev_result);
continue;
} else {
/* Normal syscall. */
handle_work(do_work);
}
}
return(OK); /* shouldn't come here */
}
/*===========================================================================*
* main *
*===========================================================================*/
int main(void)
{
/* This is the main program of the file system. The main loop consists of
* three major activities: getting new work, processing the work, and sending
* the reply. This loop never terminates as long as the file system runs.
*/
int transid;
struct worker_thread *wp;
/* SEF local startup. */
sef_local_startup();
printf("Started VFS: %d worker thread(s)\n", NR_WTHREADS);
/* This is the main loop that gets work, processes it, and sends replies. */
while (TRUE) {
worker_yield(); /* let other threads run */
send_work();
/* The get_work() function returns TRUE if we have a new message to
* process. It returns FALSE if it spawned other thread activities.
*/
if (!get_work())
continue;
transid = TRNS_GET_ID(m_in.m_type);
if (IS_VFS_FS_TRANSID(transid)) {
wp = worker_get((thread_t) transid - VFS_TRANSID);
if (wp == NULL || wp->w_fp == NULL) {
printf("VFS: spurious message %d from endpoint %d\n",
m_in.m_type, m_in.m_source);
continue;
}
m_in.m_type = TRNS_DEL_ID(m_in.m_type);
do_reply(wp);
continue;
} else if (who_e == PM_PROC_NR) { /* Calls from PM */
/* Special control messages from PM */
service_pm();
continue;
} else if (is_notify(call_nr)) {
/* A task ipc_notify()ed us */
switch (who_e) {
case DS_PROC_NR:
/* Start a thread to handle DS events, if no thread
* is pending or active for it already. DS is not
* supposed to issue calls to VFS or be the subject of
* postponed PM requests, so this should be no problem.
*/
if (worker_can_start(fp))
handle_work(ds_event);
break;
case KERNEL:
mthread_stacktraces();
break;
case CLOCK:
/* Timer expired. Used only for select(). Check it. */
expire_timers(m_in.m_notify.timestamp);
break;
default:
printf("VFS: ignoring notification from %d\n", who_e);
}
continue;
} else if (who_p < 0) { /* i.e., message comes from a task */
/* We're going to ignore this message. Tasks should
* send ipc_notify()s only.
*/
printf("VFS: ignoring message from %d (%d)\n", who_e, call_nr);
continue;
}
if (IS_BDEV_RS(call_nr)) {
/* We've got results for a block device request. */
bdev_reply();
} else if (IS_CDEV_RS(call_nr)) {
/* We've got results for a character device request. */
cdev_reply();
} else {
/* Normal syscall. This spawns a new thread. */
handle_work(do_work);
}
}
return(OK); /* shouldn't come here */
}
